Communication method and apparatus

Abstract

The application provides a communication method and device. In the method, a terminal side can obtain a data set for model training, and then can perform model training in a corresponding training manner according to the data set and a performance index, to obtain a first model meeting requirements, thereby facilitating improvement of feedback performance of CSI. The performance index corresponds to an inference task of the first model, thereby facilitating monitoring of credibility of an inference result corresponding to the inference task of the first model based on the performance index in the process of model training of the terminal side, and thereby facilitating obtaining of the first model meeting requirements.

Description

Technical Field

[0001] This application relates to the field of wireless communication, and more particularly to a communication method and apparatus. Background Technology

[0002] In wireless communication systems, artificial intelligence (AI) models can be used to compress and reconstruct (or restore) channel state information (CSI). The sender of the CSI report (such as a terminal device) can use the AI ​​model to compress the CSI and send the resulting CSI report. The receiver of the CSI report (such as a network device) can use the AI ​​model to reconstruct the CSI based on the received CSI report.

[0003] To support the integration and development of dual-end AI models (such as models deployed on both the sender and receiver sides of a CSI report), a dataset is needed for model training. One possible implementation is that the sender provides the dataset to the receiver for model training. However, the receiver's model training may not match the dataset provided by the sender, resulting in a trained model that is unsuitable or even unusable. Summary of the Invention

[0004] This application provides a communication method and apparatus that facilitates the training of models that meet the requirements and improves the feedback performance of CSI.

[0005] Firstly, a communication method is provided, which can be applied to the terminal side, for example, it can be executed by a terminal device; or it can be executed by a component deployed in the terminal device, such as circuits or chips inside the terminal device (such as modem chips, also known as baseband chips, or system-on-chip (SoC) chips or system-in-package (SIP) chips containing modem cores, etc.); or it can be executed by a device deployed outside the terminal device (such as the host of an over-the-top (OTT) system or a cloud server) or a component in the device (such as chips, processors, or circuits inside the device, etc.); or it can be implemented by a logic module or software capable of realizing all or part of the functions of the terminal device, etc. Alternatively, the method can be executed by a first device, which can be a terminal device; it can also be a component within the terminal device, such as internal circuitry or a chip (e.g., a modem chip), or a SoC chip or SIP chip containing a modem core; it can also be a device outside the terminal device (e.g., the host of an OTT system or a cloud server) or a component within a device (e.g., internal chips, processors, or circuitry); or it can be a logic module or software capable of implementing some or all of the functionalities of the terminal side, etc. This application does not limit this. For ease of understanding and explanation, the method provided in the first aspect will be described below using the terminal side as an example.

[0006] For example, the method includes: acquiring a dataset for model training to obtain a first model; the first model is obtained by training the model based on the dataset and performance metrics, the performance metrics corresponding to the inference task of the first model, the inference task of the first model being to perform a first processing on the first data to obtain second data, the first processing including compression, or the first processing including compression and quantization; the first data being used to characterize first channel information; the second data being used to determine third data based on the second model, the third data being used to characterize second channel information, the second channel information including at least one of the following: part or all of the first channel information, or third channel information, the third channel information being different from the first channel information, the inference task of the second model including processing the second data to obtain the third data.

[0007] The first model can be obtained by training the model on the terminal device or by training it on other devices besides the terminal device; this application does not limit this.

[0008] The dataset can be sent from the network side to the terminal side. The terminal side can obtain the dataset in the following ways: the terminal device can receive the dataset from the network device on the network side; or, the host or cloud server of the OTT system on the terminal side can receive the dataset from the intelligent network element on the network side; or, the terminal device can obtain the dataset received from the intelligent network element from the host or cloud server of the OTT system on the terminal side; or, the host or cloud server of the OTT system on the terminal side can obtain the dataset received from the network device from the terminal device.

[0009] Based on the above scheme, after the terminal obtains the dataset, it can train the model using appropriate training methods according to the dataset and performance metrics to obtain a model that meets the requirements, thereby improving the feedback performance of CSI. Specifically, the performance metrics correspond to the inference task of the first model, which allows the terminal to monitor the reliability of the inference results corresponding to the inference task of the first model during model training, thus facilitating the acquisition of a model that meets the requirements.

[0010] For example, the processing of the second data by the second model may include decompression, or decompression and dequantization. For instance, if the first processing includes compression, then the processing of the second data by the second model may include decompression. If the first processing includes compression and quantization, then the processing of the second data by the second model may include decompression and dequantization.

[0011] For example, the second channel information includes the third channel information, as well as part or all of the first channel information.

[0012] In conjunction with the first aspect, in some implementations of the first aspect, before model training, the method further includes: receiving first information indicating performance metrics.

[0013] This first information can be sent from the network side to the terminal side. Based on this first information, the terminal side can determine the performance metrics for training the first model.

[0014] In conjunction with the first aspect, in some implementations of the first aspect, the dataset includes one or more of the following data: fourth data, fifth data, sixth data, or seventh data.

[0015] The fourth data is used to characterize the fourth channel information.

[0016] The fifth data is obtained by performing the first processing on the fourth data. The fifth data is used to determine the eighth data, or the fifth data and the seventh data are used to determine the eighth data.

[0017] The sixth data is the true value corresponding to the eighth data.

[0018] The eighth data is used to characterize the fifth channel information, and the seventh data is used to characterize the sixth channel information. The fifth channel information includes at least one of the following: part or all of the fourth channel information, part or all of the sixth channel information, or the seventh channel information, wherein the seventh channel information is different from the fourth channel information and different from the sixth channel information.

[0019] Based on the above technical solutions, it is beneficial to design datasets that include different data to adapt to the capabilities of different terminal sides and / or model training requirements.

[0020] In one possible implementation, the dataset includes fourth and fifth data. The fourth and fifth data are used for training a first model, where the input to the first model includes the fourth data, and the ground truth value corresponding to the output of the first model includes the fifth data; or, the fourth and fifth data are used for training a third model, and the third model is used for training the fourth model to obtain the first model, where the fourth model includes both the first and third models; in the training of the third model, the input to the third model includes the fifth data, and the ground truth value corresponding to the output of the third model includes the fourth data; the training of the third model for the fourth model includes: in the training of the fourth model, data #1 is used as the input to the first model and as the ground truth value corresponding to the output of the third model, and the output of the first model is used as the input to the third model.

[0021] Optionally, data #1 includes the fourth data and / or other data besides the fourth data. For example, data #1 includes data acquired by the terminal itself, such as data obtained through measurement or simulation. The channel information represented by data #1 corresponds to the same channel as the fourth channel information.

[0022] Based on the above technical solutions, the terminal side can flexibly choose the appropriate model training method according to its own needs and / or model training requirements.

[0023] Optionally, before model training, the method further includes: receiving second information, which indicates that fourth and fifth data are used for model training of a first model, or indicates that fourth and fifth data are used for model training of a third model.

[0024] Based on the above technical solution, the terminal side can determine the model training method based on the second information, and then determine the corresponding performance indicators.

[0025] For example, the fourth and fifth data are used for training the first model. Specifically, the terminal can use the fourth data as input to the first model and the fifth data as a label to train the first model.

[0026] Based on the above technical solution, the terminal can directly train the first model based on the acquired dataset. The training steps for this model are fewer and the training process is simpler, which helps the terminal to obtain the first model more quickly through training.

[0027] Optionally, the performance metric corresponds to the inference task of the first model, including: the performance metric is the performance requirement that should be met when performing the inference task of the first model.

[0028] Optionally, the performance metrics include one or more of the following ranges that should be satisfied: a first mean square error (MSE); a first normalized mean square error (NMSE); a first mean absolute error (MAE); or, a weighted sum or a weighted average of multiples of the first MSE, the first NMSE, or the first MAE; the first MSE is the MSE between the information obtained by the first model through the first processing of the fourth data and the fifth data; the first NMSE is the NMSE between the information obtained by the first model through the first processing of the fourth data and the fifth data; the first MAE is the MAE between the information obtained by the first model through the first processing of the fourth data and the fifth data.

[0029] The difference between the output and the label of the first model is characterized by a weighted sum or weighted average of the first NMSE, the first MSE, the first MAE, or two or more of the above. The terminal side can then optimize the parameters of the first model based on this difference to reduce the discrepancy, thereby obtaining a more accurate first model for model inference.

[0030] For example, the fourth and fifth data are used for training the third model, and the third model is used for training the fourth model to obtain the first model, wherein the fourth model includes the first model. The fourth model also includes the third model, that is, the third model has the functionality of both the first and third models.

[0031] For example, the fourth model can be used to compress the fourth data to obtain the fifth data (i.e., the function of the first model); the fourth model can also be used to decompress the fifth data to obtain the reconstructed CSI (i.e., the function of the third model).

[0032] Based on the above technical solution, the terminal side can combine the third model and the first model to perform end-to-end model training. This training process can monitor model performance from the entire process of CSI compression and reconstruction, which is beneficial to improving the performance of CSI feedback through the first model.

[0033] Optionally, the performance metric corresponds to the inference task of the first model, including: the performance metric is the performance requirement that the output of the fourth model should meet.

[0034] Optionally, the performance metrics include the following ranges that should be satisfied: second MSE; second NMSE; second MAE; third MSE; third NMSE; third MAE; third generalized cosine similarity (GCS); third square generalized cosine similarity (SGCS); or, a weighted sum or weighted average of multiples of the following: second MSE, second NMSE, second MAE, third MSE, third NMSE, third MAE, third GCS, or third SGCS; the second MSE is the MSE between the reconstructed CSI obtained by the second processing of the fifth data using the third model and the fourth data; the second NMSE is the NMSE between the reconstructed CSI obtained by the second processing of the fifth data using the third model and the fourth data; the second MAE is the MSE between the reconstructed CSI obtained by the second processing of the fifth data using the third model and the fourth data. MAE; the third MSE is the MSE between the reconstructed CSI and data #1 obtained by processing data #1 using the fourth model; the third NMSE is the NMSE between the reconstructed CSI and data #1 obtained by processing data #1 using the fourth model; the third MAE is the MAE between the reconstructed CSI and data #1 obtained by processing data #1 using the fourth model; the third GCS is the GCS between the reconstructed CSI and data #1 obtained by processing data #1 using the fourth model; the third SGCS is the SGCS between the reconstructed CSI and data #1 obtained by processing data #1 using the fourth model.

[0035] The second process includes decompression, or decompression and dequantization.

[0036] The difference between the output of the third model and the input of the first model is characterized by a weighted sum or weighted average of the third NMSE, third MSE, third MAE, third GCS, third SGCS, or two or more of the above. The terminal side can optimize the parameters of the first model based on this difference to reduce the difference between the two, thereby obtaining a more accurate first model for model inference.

[0037] In one possible implementation, the dataset includes fourth, fifth, and sixth data sets. The fifth and sixth data sets are used for training the third model, and the third model is used for training the fourth model to obtain the first model, which includes the first model.

[0038] The fourth model also includes the third model, which has the functions of both the first and third models.

[0039] For example, the fourth model can be used to compress the fourth data to obtain the fifth data (i.e., the function of the first model); the fourth model can also be used to process the fifth data to obtain the reconstructed CSI (i.e., the function of the third model).

[0040] Based on the above technical solution, the terminal side can combine the third model and the first model to perform end-to-end model training. This training process can monitor model performance from the entire process of CSI compression and reconstruction, which is beneficial to improving the performance of CSI feedback through the first model.

[0041] In the training of the third model, the input of the third model includes the fifth data, and the true value corresponding to the output of the third model includes the sixth data.

[0042] The training of the third model for the fourth model includes the following: In the training of the fourth model, the ninth data is used as the input of the first model, the output of the first model is used as the input of the third model, and the true value corresponding to the output of the third model includes the tenth data.

[0043] Optionally, the ninth and tenth data points are both the fourth data points, or the ninth and tenth data points are both the sixth data points, or the ninth data point is the fourth data point and the tenth data point is the sixth data point.

[0044] Optionally, the ninth data includes data other than the fourth and sixth data, such as data acquired by the terminal itself, such as data acquired through measurement or simulation.

[0045] Optionally, the tenth data includes data other than the fourth and sixth data, such as data acquired by the terminal itself, such as data acquired through measurement or simulation.

[0046] Optionally, the performance metric corresponds to the inference task of the first model, including: the performance metric is the performance requirement that the output of the fourth model should meet.

[0047] Optionally, the performance metrics include the following ranges that should be met: Fourth MSE; Fourth NMSE; Fourth MAE; Fifth MSE; Fifth NMSE; Fifth MAE; Fifth GCS; Fifth SGCS; or, a weighted sum or weighted average of multiples of the following: Fourth MSE, Fourth NMSE, Fourth MAE, Fifth MSE, Fifth NMSE, Fifth MAE, Fifth GCS, or Fifth SGCS; Fourth MSE is the MSE between the reconstructed CSI obtained by processing the fifth data using the third model and the sixth data; Fourth NMSE is the NMSE between the reconstructed CSI obtained by processing the fifth data using the third model and the sixth data; Fourth MAE is the MSE between the reconstructed CSI obtained by processing the fifth data using the third model and the sixth data; Fourth MAE is the MSE between the reconstructed CSI obtained by processing the fifth data using the third model and the sixth data. The MAE between the reconstructed CSI and the sixth data obtained by processing the fifth data using the fourth model; the MSE between the reconstructed CSI and the tenth data obtained by processing the ninth data using the fourth model; the NMSE between the reconstructed CSI and the tenth data obtained by processing the ninth data using the fourth model; the MAE between the reconstructed CSI and the tenth data obtained by processing the ninth data using the fourth model; the GCS between the reconstructed CSI and the tenth data obtained by processing the ninth data using the fourth model; and the SGCS between the reconstructed CSI and the tenth data obtained by processing the ninth data using the fourth model.

[0048] The difference between the output of the third model and the input of the first model is characterized by a weighted sum or weighted average of the fifth NMSE, fifth MSE, fifth MAE, fifth GCS, fifth SGCS, or two or more of the above. The terminal can then optimize the parameters of the first model based on this difference to reduce the discrepancy, thereby obtaining a more accurate first model for model inference.

[0049] In one possible implementation, the dataset includes fifth, sixth, and seventh data points. These data points are used for training a third model, which in turn is used to train a fourth model to obtain a first model, which in turn includes the first model.

[0050] The fourth model also includes the third model, which has the functions of both the first and third models.

[0051] For example, the fourth model can be used to compress the sixth data (i.e., the function of the first model); the fourth model can also be used to process the compressed data of the sixth data to obtain the reconstructed CSI (i.e., the function of the third model).

[0052] Based on the above technical solution, the terminal side can combine the third model and the first model to perform end-to-end model training. This training process can monitor model performance from the entire process of CSI compression and reconstruction, which is beneficial to improving the performance of CSI feedback through the first model.

[0053] In the training of the third model, the input to the third model includes the fifth and seventh data points, and the output of the third model corresponds to the sixth data point.

[0054] The third model is used for training the fourth model, which includes: In the training of the fourth model, data #2 is used as the input of the first model and as the true value corresponding to the output of the third model. The output of the first model is used as the input of the third model, and the true value corresponding to the output of the third model includes data #3.

[0055] Optionally, data #2 includes the sixth data and / or data other than the sixth data. For example, data #2 includes data acquired by the terminal itself, such as data acquired through measurement or simulation. The channel information represented by data #2 corresponds to the same channel as the channel information represented by the sixth data.

[0056] Optionally, the performance metric corresponds to the inference task of the first model, including: the performance metric is the performance requirement that the output of the fourth model should meet.

[0057] Optionally, the performance metrics include the following ranges that should be met: Sixth MSE; Sixth NMSE; Sixth MAE; Seventh MSE; Seventh NMSE; Seventh MAE; Seventh GCS; Seventh SGCS; or, a weighted sum or weighted average of multiples of the following: Sixth MSE, Sixth NMSE, Sixth MAE, Seventh MSE, Seventh NMSE, Seventh MAE, Seventh GCS, or Seventh SGCS; Sixth MSE is the MSE between the reconstructed CSI obtained by processing the fifth and seventh data using the third model and the sixth data; Sixth NMSE is the NMSE between the reconstructed CSI obtained by processing the fifth and seventh data using the third model and the sixth data; Sixth MAE is the MSE between the reconstructed CSI obtained by processing the fifth and seventh data using the third model and the sixth data; Sixth MAE is the MSE between the reconstructed CSI obtained by processing the fifth and seventh data using the third model and the sixth data; The MAE between the reconstructed CSI and the sixth data obtained by processing the fifth and seventh data using the three models; the MSE between the reconstructed CSI and data #2 obtained by processing data #2 using the fourth model; the NMSE between the reconstructed CSI and data #2 obtained by processing data #2 using the fourth model; the MAE between the reconstructed CSI and data #2 obtained by processing data #2 using the fourth model; the GCS between the reconstructed CSI and data #2 obtained by processing data #2 using the fourth model; and the SGCS between the reconstructed CSI and data #2 obtained by processing data #2 using the fourth model.

[0058] The difference between the output of the third model and the input of the first model is characterized by a weighted sum or weighted average of the seventh NMSE, seventh MSE, seventh MAE, seventh GCS, seventh SGCS, or two or more of the above. The terminal side can optimize the parameters of the first model based on this difference to reduce the difference between the two, thereby obtaining a more accurate first model for model inference.

[0059] Optionally, the range that the above-mentioned MSE should satisfy includes the range that the average and / or distribution of the MSE should satisfy.

[0060] Optionally, the range that the above-mentioned NMSE should satisfy includes the range that the average and / or distribution of the NMSE should satisfy.

[0061] Optionally, the range that the above-mentioned MAE should satisfy includes the range that the average and / or distribution of the MAE should satisfy.

[0062] Optionally, the range that GCS should satisfy includes the range that the average and / or distribution of GCS should satisfy.

[0063] Optionally, the range that the SGCS should satisfy includes the range that the average and / or distribution of the SGCS should satisfy.

[0064] Optionally, the performance metrics are related to the measurement configuration, which is used to configure the reference signal, and the first channel information is related to the measurement results of the reference signal.

[0065] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: performing the inference task of the first model, provided that the performance metrics can be met.

[0066] If the performance metrics are met, the inference task of the first model is executed, thereby ensuring the performance of the model on the terminal side, which is beneficial to business needs.

[0067] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes: sending a third message when the performance metric cannot be met, the third message being used for one or more of the following: indicating that the performance metric cannot be met; requesting a change of the performance metric; requesting a change of the dataset; or requesting the shutdown of model training.

[0068] In this way, the network can respond when performance metrics cannot be met, such as shutting down model training, changing performance metrics, or changing the dataset. Since the network's response can be made after understanding the situation on the terminal side, the distributed datasets and performance metrics are more compatible with the capabilities of the terminal side, thereby helping to improve the performance of the dual-end model.

[0069] Secondly, a communication method is provided, which can be applied to the network side. For example, it can be executed by a network device; or by a component deployed within the network device, such as internal circuits or chips (e.g., modem chips, also known as baseband chips, or SoC chips or SIP chips containing modem cores); or by a device outside the network device (e.g., intelligent network elements on the network side) or a component within the device (e.g., internal chips, processors, or circuits); or by a logic module or software capable of implementing all or part of the functions of the network device, etc. Alternatively, the method can be executed by a second device, which can be a network device; or a component within the network device, such as internal circuits or chips (e.g., modem chips), or SoC chips or SIP chips containing modem cores; or a device outside the network device (e.g., intelligent network elements) or a component within the device (e.g., internal chips, processors, or circuits); or a logic module or software capable of implementing some or all of the functions of the network side, etc. This application does not limit this.

[0070] For example, the method includes: sending a dataset for model training to obtain a first model; the performance metrics of the first model correspond to the inference task of the first model, the first model is obtained by model training based on the dataset and the performance metrics, the inference task of the first model is to perform a first processing on the first data to obtain second data, the first processing includes compression, or the first processing includes compression and quantization; the first data is used to characterize first channel information; the second data is used to determine third data based on the second model, the third data is used to characterize second channel information, the second channel information includes at least one of the following: part or all of the first channel information, or third channel information, the third channel information is different from the first channel information, and the inference task of the second model includes processing the second data to obtain the third data.

[0071] For example, the second model is deployed on the network side.

[0072] In conjunction with the second aspect, in some implementations of the second aspect, the method further includes: sending first information, the first information being used to indicate performance metrics.

[0073] In conjunction with the second aspect, some implementations of the second aspect further include obtaining a dataset. For example, the network side obtains a trained fifth model and a second model through model training, and then obtains the dataset by running the fifth model and the second model. The inference task of the fifth model is to compress the input CSI or to perform compression and quantization. For example, the inference task of the fifth model is the same as the inference task of the first model.

[0074] In conjunction with the second aspect, in some implementations of the second aspect, the second channel information includes the third channel information, as well as part or all of the first channel information.

[0075] In conjunction with the second aspect, in some implementations of the second aspect, the dataset includes one or more of the following data: fourth data, fifth data, sixth data, or seventh data.

[0076] The fourth data is used to characterize the fourth channel information.

[0077] The fifth data is obtained by performing the first processing on the fourth data. The fifth data is used to determine the eighth data, or the fifth data and the seventh data are used to determine the eighth data.

[0078] The sixth data is the true value corresponding to the eighth data.

[0079] The eighth data is used to characterize the fifth channel information, and the seventh data is used to characterize the sixth channel information. The fifth channel information includes at least one of the following: part or all of the fourth channel information, part or all of the sixth channel information, or the seventh channel information, wherein the seventh channel information is different from the fourth channel information and different from the sixth channel information.

[0080] In conjunction with the second aspect, in some implementations of the second aspect, the dataset includes fourth data and fifth data. The fourth and fifth data are used for model training of the first model, where the input to the first model includes the fourth data, and the ground truth value corresponding to the output of the first model includes the fifth data; or, the fourth and fifth data are used for model training of the third model, where the third model is used for model training against the fourth model to obtain the first model, and the fourth model includes both the first and third models; in the model training of the third model, the input to the third model includes the fifth data, and the ground truth value corresponding to the output of the third model includes the fourth data; the model training of the third model for the fourth model includes: in the model training of the fourth model, data #1 is used as the input to the first model and as the ground truth value corresponding to the output of the third model, and the output of the first model is used as the input to the third model.

[0081] Optionally, data #1 includes the fourth data and / or other data besides the fourth data. For example, data #1 includes data acquired by the terminal itself, such as data acquired through measurement or simulation.

[0082] In conjunction with the second aspect, in some implementations of the second aspect, the method further includes: sending second information, which is used to instruct the fourth and fifth data to be used for model training of the first model, or to instruct the fourth and fifth data to be used for model training of the third model.

[0083] In conjunction with the second aspect, in some implementations of the second aspect, the fourth data and the fifth data are used for model training of the first model, and the performance metrics include one or more of the following ranges that should be satisfied: first MSE; first NMSE; first MAE; or, a weighted sum of multiples of the first MSE, first NMSE, or first MAE, or a weighted average; the first MSE is the MSE between the information obtained by the first model through the first processing of the fourth data and the fifth data; the first NMSE is the NMSE between the information obtained by the first model through the first processing of the fourth data and the fifth data; the first MAE is the MAE between the information obtained by the first model through the first processing of the fourth data and the fifth data.

[0084] In conjunction with the second aspect, in some implementations of the second aspect, the fourth and fifth data are used for model training of the third model, and the performance metrics include one or more of the following ranges that should be satisfied: second MSE; second NMSE; second MAE; third MSE; third NMSE; third MAE; third GCS; third SGCS; or, a weighted sum or weighted average of multiples of the second MSE, second NMSE, second MAE, third MSE, third NMSE, third MAE, third GCS, or third SGCS; the second MSE is the MSE between the reconstructed CSI obtained by the third model through the second processing of the fifth data and the fourth data; the second NMSE is the MSE between the reconstructed CSI obtained by the third model through the second processing of the fifth data and the fourth data. The second MAE is the MAE between the reconstructed CSI and the fourth data obtained by processing the fifth data using the third model; the third MSE is the MSE between the reconstructed CSI and the data #1 obtained by processing the data #1 using the fourth model; the third NMSE is the NMSE between the reconstructed CSI and the data #1 obtained by processing the data #1 using the fourth model; the third MAE is the MAE between the reconstructed CSI and the data #1 obtained by processing the data #1 using the fourth model; the third GCS is the GCS between the reconstructed CSI and the data #1 obtained by processing the data #1 using the fourth model; the third SGCS is the SGCS between the reconstructed CSI and the data #1 obtained by processing the data #1 using the fourth model.

[0085] The second process includes decompression, or decompression and dequantization.

[0086] In conjunction with the second aspect, in some implementations of the second aspect, the dataset includes fourth, fifth, and sixth data. The fifth and sixth data are used for training the third model, and the third model is used for training the fourth model to obtain the first model, which includes both the first and third models.

[0087] In the training of the third model, the input of the third model includes the fifth data, and the true value corresponding to the output of the third model includes the sixth data.

[0088] The training of the third model for the fourth model includes the following: In the training of the fourth model, the ninth data is used as the input of the first model, the output of the first model is used as the input of the third model, and the true value corresponding to the output of the third model includes the tenth data.

[0089] Optionally, the ninth and tenth data points are both the fourth data points, or the ninth and tenth data points are both the sixth data points, or the ninth data point is the fourth data point and the tenth data point is the sixth data point.

[0090] Optionally, the ninth data includes data other than the fourth and sixth data, such as data acquired by the terminal itself, such as data acquired through measurement or simulation.

[0091] Optionally, the tenth data includes data other than the fourth and sixth data, such as data acquired by the terminal itself, such as data acquired through measurement or simulation.

[0092] Optionally, the performance metrics include the following ranges that should be met: Fourth MSE; Fourth NMSE; Fourth MAE; Fifth MSE; Fifth NMSE; Fifth MAE; Fifth GCS; Fifth SGCS; or, a weighted sum or weighted average of multiples of the following: Fourth MSE, Fourth NMSE, Fourth MAE, Fifth MSE, Fifth NMSE, Fifth MAE, Fifth GCS, or Fifth SGCS; Fourth MSE is the MSE between the reconstructed CSI obtained by processing the fifth data using the third model and the sixth data; Fourth NMSE is the NMSE between the reconstructed CSI obtained by processing the fifth data using the third model and the sixth data; Fourth MAE is the MSE between the reconstructed CSI obtained by processing the fifth data using the third model and the sixth data; Fourth MAE is the MSE between the reconstructed CSI obtained by processing the fifth data using the third model and the sixth data. The MAE between the reconstructed CSI and the sixth data obtained by processing the fifth data using the fourth model; the MSE between the reconstructed CSI and the tenth data obtained by processing the ninth data using the fourth model; the NMSE between the reconstructed CSI and the tenth data obtained by processing the ninth data using the fourth model; the MAE between the reconstructed CSI and the tenth data obtained by processing the ninth data using the fourth model; the GCS between the reconstructed CSI and the tenth data obtained by processing the ninth data using the fourth model; and the SGCS between the reconstructed CSI and the tenth data obtained by processing the ninth data using the fourth model.

[0093] In conjunction with the second aspect, in some implementations of the second aspect, the dataset includes the fifth, sixth, and seventh data sets. The fifth, sixth, and seventh data sets are used for training the third model, and the third model is used for training the fourth model to obtain the first model, which includes both the first and third models.

[0094] In the training of the third model, the input to the third model includes the fifth and seventh data points, and the output of the third model corresponds to the sixth data point.

[0095] The third model is used for training the fourth model, which includes: In the training of the fourth model, data #2 is used as the input of the first model and as the true value corresponding to the output of the third model. The output of the first model is used as the input of the third model, and the true value corresponding to the output of the third model includes data #3.

[0096] Optionally, data #2 includes the sixth data and / or data other than the sixth data. For example, data #2 includes data acquired by the terminal itself, such as data acquired through measurement or simulation.

[0097] Optionally, the performance metrics include the following ranges that should be met: Sixth MSE; Sixth NMSE; Sixth MAE; Seventh MSE; Seventh NMSE; Seventh MAE; Seventh GCS; Seventh SGCS; or, a weighted sum or weighted average of multiples of the following: Sixth MSE, Sixth NMSE, Sixth MAE, Seventh MSE, Seventh NMSE, Seventh MAE, Seventh GCS, or Seventh SGCS; Sixth MSE is the MSE between the reconstructed CSI obtained by processing the fifth and seventh data using the third model and the sixth data; Sixth NMSE is the NMSE between the reconstructed CSI obtained by processing the fifth and seventh data using the third model and the sixth data; Sixth MAE is the MSE between the reconstructed CSI obtained by processing the fifth and seventh data using the third model and the sixth data; Sixth MAE is the MSE between the reconstructed CSI obtained by processing the fifth and seventh data using the third model and the sixth data; The MAE between the reconstructed CSI and the sixth data obtained by processing the fifth and seventh data using the three models; the MSE between the reconstructed CSI and data #2 obtained by processing data #2 using the fourth model; the NMSE between the reconstructed CSI and data #2 obtained by processing data #2 using the fourth model; the MAE between the reconstructed CSI and data #2 obtained by processing data #2 using the fourth model; the GCS between the reconstructed CSI and data #2 obtained by processing data #2 using the fourth model; and the SGCS between the reconstructed CSI and data #2 obtained by processing data #2 using the fourth model.

[0098] In conjunction with the second aspect, in some implementations of the second aspect, the method further includes: receiving third information, which is used for one or more of the following: indicating that a performance metric cannot be met; requesting a change of performance metric; requesting a change of dataset; or requesting to shut down model training.

[0099] It should be understood that the methods provided in the second aspect correspond to those in the first aspect. The descriptions and technical effects of the various implementation methods in the second aspect can be found in the relevant descriptions in the first aspect, and will not be repeated here.

[0100] Thirdly, an apparatus is provided. This apparatus may include functional modules corresponding to each of the methods / operations / steps / actions described in any possible implementation of the first aspect, or may include functional modules corresponding to each of the methods / operations / steps / actions described in any of the second aspects. The module may be a hardware circuit, software, or a combination of hardware circuitry and software implementation.

[0101] In one design, the device may include a processing module and a communication module. The communication module is used to perform the sending and receiving actions performed by the terminal side in the method described in the first aspect above, while the processing module is used to perform processing-related actions performed by the terminal side in the method described in the first aspect above.

[0102] In one design, the device can be a terminal device, or a device, module, circuit, or chip configured in the terminal device, or a device that can be used in conjunction with the terminal device, such as an OTT host or cloud server.

[0103] In one design, the device may include a processing module and a communication module. The communication module is used to perform the sending and receiving actions performed by the network side in the method described in the second aspect above, while the processing module is used to perform processing-related actions performed by the network side in the method described in the second aspect above.

[0104] In one design, the device can be a network device, or a device, module, circuit, or chip configured in the network device, or a device that can be used in conjunction with the network device, such as an intelligent network element with a radio access network (RAN) intelligent controller (RIC) deployed thereon.

[0105] Fourthly, an apparatus is provided, comprising a processor and a storage medium storing instructions that, when executed by the processor, cause a method as described in the first aspect or any possible implementation thereof to be implemented, or cause a method as described in the second aspect or any possible implementation thereof to be implemented.

[0106] Fifthly, an apparatus is provided, comprising a processing circuit for processing data and / or information such that a method as in the first aspect or any possible implementation thereof is implemented, or a method as in the second aspect or any possible implementation thereof is implemented.

[0107] The processing circuit may include one or more processors, or all or part of the circuitry in one or more processors used for control or processing functions.

[0108] Optionally, the apparatus may further include a memory for storing programs or instructions, and the processor for running the programs or instructions to implement the methods as described in the first aspect or any possible implementation thereof, or to implement the methods as described in the second aspect or any possible implementation thereof.

[0109] Optionally, the device may also include the transceiver circuit, or an input / output interface.

[0110] In a sixth aspect, a chip is provided, including processing circuitry for running a program or instructions to cause the method as described in the first aspect or any possible implementation thereof to be implemented, or to cause the method as described in the second aspect or any possible implementation thereof to be implemented.

[0111] Optionally, the chip may further include a memory for storing programs or instructions.

[0112] Optionally, the chip may also include transceiver circuitry, or input / output interfaces.

[0113] A seventh aspect provides a computer-readable storage medium comprising instructions that, when executed by a processor, cause the method as described in the first aspect or any possible implementation thereof to be implemented, or cause the method as described in the second aspect or any possible implementation thereof to be implemented.

[0114] Eighthly, a computer program product is provided, the computer program product comprising computer program code or instructions, which, when executed, cause the method as described in the first aspect and any possible implementation thereof to be implemented, or cause the method as described in the second aspect and any possible implementation thereof to be implemented.

[0115] Ninth aspect, a communication system is provided, the communication system including means for performing the first aspect and any possible implementation thereof, or including means for performing the second aspect and any possible implementation thereof.

[0116] It should be understood that the third to ninth aspects of this application correspond to the technical solutions of the first to second aspects of this application, and the beneficial effects obtained by each aspect and the corresponding feasible implementation are similar, and will not be repeated here. Attached Figure Description

[0117] Figure 1 This is a schematic diagram of a communication system applicable to the communication method in the embodiments of this application;

[0118] Figure 2 This is a schematic diagram of another communication system applicable to the communication method of the embodiments of this application;

[0119] Figure 3 This is a schematic diagram of a possible application framework in a communication system;

[0120] Figure 4 This is a schematic diagram of another possible application framework in a communication system;

[0121] Figure 5 This is a schematic diagram of CSI feedback using an auto-encoder (AE) model provided in an embodiment of this application;

[0122] Figure 6 An example of a neuron structure is shown;

[0123] Figure 7 This is a schematic diagram of a deep neural network (DNN);

[0124] Figure 8 This is a schematic diagram illustrating the data set connection between the network side and the terminal side provided in an embodiment of this application;

[0125] Figure 9 This is a schematic diagram of model interfacing between the network side and the terminal side provided in an embodiment of this application;

[0126] Figure 10 This is a schematic flowchart of the communication method provided in the embodiments of this application;

[0127] Figures 11 to 15 This is a flowchart of the model training provided in the embodiments of this application;

[0128] Figure 16 This is another illustrative flowchart of the communication method provided in the embodiments of this application;

[0129] Figure 17 and Figure 18 This is a schematic block diagram of the communication device provided in the embodiments of this application;

[0130] Figure 19 This is a schematic block diagram of the AI ​​processor provided in the embodiments of this application;

[0131] Figure 20 This is a schematic diagram illustrating the application of a dual-ended model encoder and decoder in a channel information processing scenario, as provided in an embodiment of this application. Detailed Implementation

[0132] The technical solution provided in this application will now be described in conjunction with the accompanying drawings.

[0133] To facilitate understanding of the embodiments of this application, the following points will be explained first:

[0134] First, in this application, the terminal side can also be referred to as the user equipment (UE) side, terminal-side equipment, etc., including: terminal equipment (or user equipment, terminal, etc.), components deployed in the terminal equipment (such as circuits or chips inside the terminal equipment), equipment deployed outside the terminal equipment (such as the host or cloud server of an OTT system, hereinafter referred to as the OTT system server), or components deployed in equipment outside the terminal equipment (such as circuits or chips inside the equipment). The network side (NW side) can also be referred to as network-side equipment, including: network equipment communicating with the terminal equipment, components deployed in the network equipment (such as circuits or chips inside the network equipment with near real-time radio access network (RAN) intelligent control functions), equipment deployed outside the network equipment (such as intelligent network elements, for example, intelligent network elements with near real-time RAN intelligent control functions), or components deployed in the intelligent network element (such as circuits or chips inside the intelligent network element). Among them, network equipment can include: access network equipment, core network equipment, or operation administration and maintenance (OAM).

[0135] Second, in this application, the indication includes direct indication (also known as explicit indication) and indirect indication (also known as implicit indication). Directly indicating information A means including information A; indirectly indicating information A can mean indicating information A through the correspondence between information A and information B and by directly indicating information B; or by indicating information A through a preset rule that can be used to determine A based on B and by directly indicating information B. The correspondence between information A and information B, and the preset rule, can be predefined, pre-stored, pre-burned, or pre-configured.

[0136] Third, in this application, "at least one" means one or more, and "more than one" means two or more. "And / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A alone, A and B simultaneously, or B alone, where A and B can be singular or plural. The character " / " generally indicates an "or" relationship between the preceding and following related objects, but it does not exclude the possibility of indicating an "and" relationship; the specific meaning can be understood in context. "At least one of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b, or c can mean: a, b, c; a and b; a and c; b and c; or a and b and c. Here, a, b, and c can be single or multiple.

[0137] Fourth, in this application, the use of prefixes such as "first" and "second" is merely for the purpose of distinguishing and describing different things belonging to the same category of names, and does not constrain the order, size, or quantity of things. For example, "first model" and "second model" are simply different models, and do not limit the number, size, or priority of the models; similarly, "first information" and "second information" are simply different indicative information, and do not limit the quantity, chronological order, size, or priority of the information.

[0138] Fifth, in this application, "send" and "receive" indicate the direction of signal transmission. For example, "send information to a terminal device" can be understood as the destination of the information being the terminal device, which may include direct transmission via the air interface or indirect transmission by other units or modules via the air interface. "Receive information from a network device" can be understood as the source of the information being the network device, which may include direct reception from the network device via the air interface or indirect reception from the network device by other units or modules via the air interface. "Send" can also be understood as the "output" of the chip interface, and "receive" can also be understood as the "input" of the chip interface. In other words, sending and receiving can occur between devices, such as between a terminal device and a computing node, or within a device, such as between components, modules, chips, software modules, or hardware modules within the device via a bus, wiring, or interface.

[0139] Sixth, in the embodiments of this application, "when," "if," and "if" all refer to the device making corresponding processing under certain objective circumstances, and are not limited to a time, nor do they require the device to make a judgment action when it is implemented, nor do they mean that there are other limitations.

[0140] Seventh, in this application, the words "example," "exemplarily," "for example," or "such as" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "example," "exemplarily," "for example," or "such as" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the words "example," "exemplarily," "for example," or "such as" is intended to present the relevant concepts in a specific manner.

[0141] Eighth, for ease of distinction and explanation, this paper introduces the terms encoder and decoder, quantizer and dequantizer. These names are given only to distinguish different functions and do not limit the structure of the device. For example, if the terminal side (such as a terminal device or OTT system server) has the inference function of the encoder, it can be said that the terminal side contains the encoder, which can be understood as a functional module of the terminal side; if the terminal side has the inference function of the decoder, it can be said that the terminal side contains the decoder, which can be understood as a functional module of the terminal side. If the encoder has the quantization function, it can be said that the encoder contains the quantizer. If the decoder has the dequantization function, it can be said that the decoder contains the dequantizer. In specific implementations, the encoder and decoder can be implemented through a neural network model. More specifically, the functions of the encoder and decoder, quantizer and dequantizer can be implemented separately in hardware, or in software, or in a combination of hardware and software; this application does not limit this.

[0142] The technical solutions provided in this application can be applied to various communication systems, such as 5th generation (5G) or new radio (NR) systems, frequency division duplex (FDD) systems, time division duplex (TDD) systems, wireless local area network (WLAN) systems, satellite communication systems, future communication systems, or integrated systems of multiple systems. The technical solutions provided in this application can also be applied to device-to-device (D2D) communication, vehicle-to-everything (V2X) communication, machine-to-machine (M2M) communication, machine-type communication (MTC), and Internet of Things (IoT) communication systems or other communication systems.

[0143] In a communication system, one network element can send signals to or receive signals from another network element. These signals can include information, signaling, or data. The term "network element" can also be replaced by an entity, network entity, device, communication equipment, communication module, node, communication node, etc. This disclosure uses a network element as an example. For instance, a communication system can include at least one terminal device and at least one network device. The network device can send downlink signals to the terminal device, and / or the terminal device can send uplink signals to the network device. It is understood that the terminal device in this disclosure can be replaced by a first network element, and the network device can be replaced by a second network element, both performing the corresponding methods described in this disclosure.

[0144] Figure 1 This is a schematic diagram of a communication system applicable to the communication method in the embodiments of this application. For example... Figure 1 As shown, the communication system 100A may include at least one access network device, such as Figure 1 The access network device 110 shown; the communication system 100A may also include at least one terminal device, such as Figure 1 Terminal devices 120 and 130 are shown. Access network device 110 can communicate with terminal devices (such as terminal devices 120 and 130) via a wireless link. Communication devices in this communication system, for example, between access network device 110 and terminal device 120, can communicate via multi-antenna technology.

[0145] In wireless communication networks, such as mobile communication networks, the services supported by the networks are becoming increasingly diverse, thus requiring increasingly diverse demands. For example, networks need to support ultra-high speeds, ultra-low latency, and / or massive connectivity. This characteristic makes network planning, network configuration, and / or resource scheduling increasingly complex. Furthermore, as network functions become more powerful, such as supporting higher spectrum, higher-order multiple-input multiple-output (MIMO) technology, beamforming (BF), and beam management, network energy efficiency has become a hot research topic. These new demands, scenarios, and characteristics bring unprecedented challenges to network planning, operation, and efficient operation. To meet these challenges, artificial intelligence (AI) technology can be introduced into wireless communication networks to achieve network intelligence. To support AI technology in wireless networks, AI nodes may also be introduced. AI nodes can be AI network elements or AI modules. Figure 2 This is a schematic diagram of another communication system applicable to the communication method in the embodiments of this application. Compared to Figure 1 Regarding the communication system 100A shown, Figure 2 The communication system 100B shown also includes an AI network element 140. The AI ​​network element 140 is used to perform AI-related operations, such as building datasets or AI models. The AI ​​network element can also be simply referred to as an intelligent network element. In this disclosure, the AI ​​model can be simply referred to as a model.

[0146] In one possible implementation, access network device 110 can send data related to the training of the AI ​​model to AI network element 140, whereby AI network element 140 constructs a dataset and trains the AI ​​model. For example, the data related to the training of the AI ​​model may include data reported by terminal devices. AI network element 140 can send the results of operations related to the AI ​​model to access network device 110, and then forward them to terminal devices via access network device 110. For example, the results of operations related to the AI ​​model may include at least one of the following: a trained AI model, model evaluation results, or test results, etc. Exemplarily, a portion of the trained AI model may be deployed on access network device 110, and another portion on terminal devices 120 and / or 130. Alternatively, the trained AI model may be deployed on access network device 110. Or, the trained AI model may be deployed on terminal devices 120 and / or 130.

[0147] It should be understood that Figure 2This explanation only uses the direct connection between AI network element 140 and access network device 110 as an example. In other scenarios, AI network element 140 can also be connected to terminal devices. Alternatively, AI network element 140 can be connected to both access network device 110 and terminal devices simultaneously. Alternatively, AI network element 140 can also be connected to access network device 110 through a third-party network element. This application embodiment does not limit the connection relationship between AI network element and other network elements. For example, AI network element 140 can also be configured as a module in access network device and / or terminal device, for example, configured in... Figure 1 In the access network device 110 or terminal device shown.

[0148] It should be noted that, Figure 1 and Figure 2 This is a simplified illustration for ease of understanding only. For example, the communication system may also include other devices, such as wireless relay devices and / or wireless backhaul devices. Figure 1 and Figure 2 The figures are not shown. In practical applications, this communication system may include multiple access network devices or multiple terminal devices. This application does not limit the number of access network devices and terminal devices included in the communication system.

[0149] In the embodiments of this application, the terminal device may also be referred to as UE, access terminal, user unit, user station, mobile station, mobile station, remote station, remote terminal, mobile device, user terminal, terminal, wireless communication device, user agent, or user equipment.

[0150] Terminal devices can be devices that provide voice / data, such as handheld devices with wireless connectivity, in-vehicle devices, etc. Currently, examples of terminals include: mobile phones, tablets, laptops, PDAs, mobile internet devices (MIDs), wearable devices, virtual reality (VR) devices, augmented reality (AR) devices, wireless terminals in industrial control, wireless terminals in self-driving, wireless terminals in remote medical surgery, wireless terminals in smart grids, wireless terminals in transportation safety, wireless terminals in smart cities, wireless terminals in smart homes, cellular phones, cordless phones, session initiation protocol (SIP) phones, wireless local loop (WLL) stations, personal digital assistants (PDAs), handheld devices with wireless communication capabilities, computing devices or other processing devices connected to a wireless modem, wearable devices, terminal devices in 5G networks, or future public land mobile communication networks. Terminal devices in a network (PLMN), etc., are not limited to this in the embodiments of this application.

[0151] By way of example and not limitation, in this embodiment, the terminal device can also be a wearable device. Wearable devices, also known as wearable smart devices, are a general term for devices that utilize wearable technology to intelligently design and develop everyday wearables, such as glasses, gloves, watches, clothing, and shoes. Wearable devices are portable devices that are worn directly on the body or integrated into the user's clothing or accessories. Wearable devices are not merely hardware devices, but also achieve powerful functions through software support, data interaction, and cloud interaction. Broadly speaking, wearable smart devices include those that are feature-rich, large in size, and can achieve complete or partial functions without relying on a smartphone, such as smartwatches or smart glasses, as well as those that focus on a specific type of application function and require the use of other devices such as smartphones, such as various smart bracelets and smart jewelry for vital sign monitoring.

[0152] In this embodiment, the device for implementing the functions of the terminal device can be the terminal device itself, or it can be any device capable of supporting the terminal device in implementing those functions, such as a chip system. This device can be installed in or used in conjunction with the terminal device. In this embodiment, the chip system can be composed of chips or may include chips and other discrete components. This embodiment only uses the terminal device as an example to illustrate the device for implementing the functions of the terminal device, and does not constitute a limitation on the solution of this embodiment.

[0153] The network device in this application embodiment can be a device for communicating with a terminal device. This network device may include access network equipment, core network equipment, or other equipment in the communication system. The access network equipment may be, for example, a base station. In this application embodiment, the access network equipment may refer to a RAN node (or device) that connects the terminal device to the wireless network. A base station can broadly encompass, or be replaced by, various names such as: NodeB, evolved NodeB (eNB), next-generation NodeB (gNB), relay station, access point, transmit / receive point (TRP), transmitting point (TP), master station, auxiliary station, motor slide retainer (MSR) node, home base station, network controller, access node, wireless node, access point (AP), transmission node, transceiver node, baseband unit (BBU), remote radio unit (RRU), active antenna unit (AAU), remote radio head (RRH), central unit (CU), distributed unit (DU), radio unit (RU), positioning node, etc. A base station can be a macro base station, micro base station, relay node, donor node, or similar entities, or combinations thereof. A base station can also refer to a communication module, modem, or chip installed within the aforementioned equipment or apparatus. A base station can also be a mobile switching center, a device that performs base station functions in D2D, V2X, and M2M communications, or a device that performs base station functions in future communication systems. A base station can support networks using the same or different access technologies. Optionally, a RAN node can also be a server, wearable device, vehicle, or in-vehicle equipment. For example, the access network equipment in V2X technology can be a roadside unit (RSU). The embodiments of this application do not limit the specific technology or equipment form used in the access network equipment.

[0154] Base stations can be fixed or mobile. For example, a helicopter or drone can be configured to act as a mobile base station, and one or more cells can move depending on the location of the mobile base station. In other examples, a helicopter or drone can be configured as a device to communicate with another base station.

[0155] In some deployments, the access network equipment mentioned in the embodiments of this application may be a device including a CU, or a DU, or a device including both CU and DU, or a device with a control plane CU node (central unit-control plane (CU-CP)) and a user plane CU node (central unit-user plane (CU-UP)) and a DU node. For example, the access network equipment may include gNB-CU-CP, gNB-CU-UP, and gNB-DU.

[0156] In some deployments, multiple RAN nodes collaborate to assist terminals in achieving wireless access, with different RAN nodes each implementing some of the base station's functions. For example, RAN nodes can be CUs, DUs, CU-CPs, CU-UPs, or RUs. CUs and DUs can be configured separately or included in the same network element, such as a BBU. RUs can be included in radio frequency equipment or radio frequency units, such as RRUs, AAUs, or RRHs.

[0157] RAN nodes can support one or more types of fronthaul interfaces, each corresponding to a DU and RU with different functions. If the fronthaul interface between the DU and RU is a Common Public Radio Interface (CPRI), the DU is configured to implement one or more baseband functions, and the RU is configured to implement one or more radio frequency functions. If the fronthaul interface between the DU and RU is another type of interface, relative to CPRI, it moves some downlink and / or uplink baseband functions—for example, for downlink, precoding, digital beamforming, or one or more of inverse fast Fourier transform (IFFT) / adding a cyclic prefix (CP)—from the DU to the RU; and for uplink, digital beamforming, or one or more of fast Fourier transform (FFT) / removing CP—from the DU to the RU. In one possible implementation, this interface can be an enhanced common public radio interface (eCPRI). Under the eCPRI architecture, the splitting methods between DU and RU are different, corresponding to different types (category, Cat) of eCPRI, such as eCPRI Cat A, B, C, D, E, F.

[0158] Taking eCPRI Cat A as an example, for downlink transmission, layer mapping is used as the dividing line. The DU is configured to implement one or more functions preceding layer mapping (i.e., coding, rate matching, scrambling, modulation, and layer mapping itself), while other functions following layer mapping (e.g., resource element (RE) mapping, digital beamforming, or one or more of IFFT / CP addition) are implemented in the RU. For uplink transmission, de-RE mapping is used as the dividing line. The DU is configured to implement one or more functions preceding de-mapping (i.e., decoding, rate matching de-matching, descrambling, demodulation, inverse discrete Fourier transform (IDFT), channel equalization, and de-RE mapping itself), while other functions following de-mapping (e.g., digital BF or FFT / CP removal) are implemented in the RU. It is understood that descriptions of the functions of the DU and RU corresponding to various types of eCPRI can be found in the eCPRI protocol and will not be elaborated upon here.

[0159] In one possible design, the processing unit in the BBU used to implement baseband functions is called the baseband high (BBH) unit, and the processing unit in the RRU / AAU / RRH used to implement baseband functions is called the baseband low (BBL) unit.

[0160] In different systems, CU (or CU-CP and CU-UP), DU, or RU may have different names, but those skilled in the art will understand their meaning. For example, in an open RAN (ORAN) architecture, CU can also be called open CU (open-CU, O-CU), DU can also be called open DU (open-DU, O-DU), CU-CP can also be called open CU-CP (open-CU-CP) O-CU-CP, CU-UP can also be called open CU-UP (open-CU-UP, O-CU-UP), and RU can also be called open RU (open-RU, O-RU). Any of the units among CU (or CU-CP, CU-UP), DU, and RU in this application can be implemented through software modules, hardware modules, or a combination of software modules and hardware modules.

[0161] In this embodiment, the apparatus for implementing the functions of a network device can be a network device itself; it can also be an apparatus capable of supporting the network device in implementing those functions, such as a chip system, hardware circuit, software module, or a hardware circuit plus a software module. This apparatus can be installed in the network device or used in conjunction with the network device. In this embodiment, the example of a network device being used to implement the functions of a network device is provided only and does not constitute a limitation on the solutions described in this embodiment.

[0162] Network devices and / or terminal devices can be deployed on land, including indoors or outdoors, handheld or vehicle-mounted; they can also be deployed on water; and they can also be deployed in the air on airplanes, balloons, and satellites. This application does not limit the scenario in which the network devices and terminal devices are located. Furthermore, terminal devices and network devices can be hardware devices, or software functions running on dedicated hardware or general-purpose hardware, such as virtualization functions instantiated on a platform (e.g., a cloud platform), or entities that include dedicated or general-purpose hardware devices and software functions. This application does not limit the specific form of the terminal devices and network devices.

[0163] Optionally, the AI ​​node can be deployed in one or more of the following locations within the communication system: access network equipment, terminal equipment, or core network elements. Alternatively, the AI ​​node can also be deployed independently, for example, in a location other than any of the aforementioned devices, such as in the host or cloud server of an over-the-top (OTT) system. The AI ​​node can communicate with other devices in the communication system, which can be, for example, one or more of the following: access network equipment, terminal equipment, or core network elements.

[0164] It is understood that this application does not limit the number of AI nodes. For example, when there are multiple AI nodes, they can be divided based on function, such as different AI nodes being responsible for different functions.

[0165] It can also be understood that AI nodes can be independent devices, or they can be integrated into the same device to achieve different functions. Alternatively, they can be network elements in hardware devices, software functions running on dedicated hardware, or virtualization functions instantiated on a platform (e.g., a cloud platform). This application does not limit the specific form of the aforementioned AI nodes.

[0166] Figure 3 This is a schematic diagram of a possible application framework in a communication system. For example... Figure 3As shown, network elements in a communication system are connected via interfaces (such as next-generation (NG) interfaces or Xn interfaces) or air interfaces. The NG interface is the interface between the radio access network and the 5G core network. The Xn interface is the interface between access network devices, and the air interface is the interface between access network devices and terminal devices. These network element nodes, such as core network devices, RAN nodes, terminal devices, or one or more devices in the OAM, are equipped with one or more AI modules (for clarity, ...). Figure 3 (Only one is shown in the image). The access network node can be a single RAN node or can include multiple RAN nodes, such as RU, CU, and DU. One or more of the RU, CU, or DU can also be equipped with one or more AI modules. Optionally, the CU can also be split into CU-CP and CU-UP. One or more AI modules are provided in CU-CP and / or CU-UP.

[0167] The AI ​​module is used to implement corresponding AI functions. AI modules deployed in different network elements can be the same or different. Depending on the parameter configuration, the AI ​​module can implement different functions. The AI ​​module model can be configured based on one or more of the following parameters: structural parameters (e.g., at least one of the following: number of neural network layers, neural network width, inter-layer connections, neuron weights, neuron activation function, or bias in the activation function), input parameters (e.g., type and / or dimension of input parameters), or output parameters (e.g., type and / or dimension of output parameters). The bias in the activation function can also be referred to as the neural network bias.

[0168] An AI module can have one or more models. A model can infer an output, which includes one or more parameters. The learning, training, or inference processes of different models can be deployed on different nodes or devices, or they can be deployed on the same node or device.

[0169] The network device can be a network device equipped with one or more AI modules. The network device may include... Figure 3 This refers to one or more devices within the core network, RAN, or OAM. For example, the AI ​​module could be... Figure 4The RAN intelligent controller (RIC) shown can be a near-real-time (near-RT) RIC or a non-real-time (non-RT) RIC. For example, a near-real-time RIC is located in a RAN node (e.g., in a CU or DU), while a non-real-time RIC is located in an OAM, a cloud server, a core network device, or other access network devices. The RIC can obtain subsets from multiple terminal devices from RAN nodes (e.g., CU, CU-CP, CU-UP, DU, and / or RU), reassemble them into a dataset, and train based on the dataset. Exemplarily, near-real-time and non-real-time RICs can also be set up as separate network elements, and access network devices can be either near-real-time or non-real-time RICs.

[0170] Figure 4 This is a schematic diagram of another possible application framework in a communication system. Figure 4 The communication system shown includes access network nodes (CU, DU, and RU shown in the figure) and terminals, as well as a RIC. For example, the RIC could be... Figure 3 The AI ​​module shown can be used to implement AI-related functions. The RIC includes near real-time RIC and non-real-time RIC. Non-real-time RIC primarily processes non-real-time information, such as data that is not sensitive to latency, with latency on the order of seconds. Real-time RIC primarily processes near real-time information, such as data that is relatively sensitive to latency, with latency on the order of tens of milliseconds.

[0171] The near real-time RIC is used for model training and inference. For example, it is used to train an AI model and then use that AI model for inference. The near real-time RIC can obtain network-side and / or terminal-side information from RAN nodes (e.g., CU, CU-CP, CU-UP, DU, and / or RU) and / or terminals. This information can be used as training data or inference data. Optionally, the near real-time RIC can deliver the inference results to the RAN nodes and / or terminals. Optionally, inference results can be exchanged between CU and DU, and / or between DU and RU. For example, the near real-time RIC delivers the inference results to the DU, and the DU sends them to the RU.

[0172] The non-real-time RIC is also used for model training and inference. For example, it can be used to train an AI model and then use that model for inference. The non-real-time RIC can obtain network-side and / or terminal-side information from RAN nodes (e.g., one or more of CU, CU-CP, CU-UP, DU, or RU) and / or terminals. This information can be used as training data or inference data, and the inference results can be delivered to the RAN nodes and / or terminals. Optionally, inference results can be exchanged between CU and DU, and / or between DU and RU; for example, the non-real-time RIC delivers the inference results to the DU, which then forwards them to the RU.

[0173] The near real-time RIC and non-real-time RIC can also be set up as separate network elements. Optionally, the near real-time RIC and non-real-time RIC can also be part of other devices. For example, the near real-time RIC can be set in the RAN node (e.g., in CU, DU), while the non-real-time RIC can be set in the OAM, cloud server, core network device, or other network device.

[0174] With the development of wireless communication technology and the increasing number of supported services, higher demands are being placed on communication systems in terms of system capacity, communication latency, and other indicators. Among these advancements, massive MIMO (Multi-User MIMO) systems, by configuring large-scale antenna arrays at the transceiver end, can achieve spatial diversity gain, thereby significantly increasing system capacity. For example, access network equipment can simultaneously transmit data to multiple terminal devices using the same time-frequency resources, i.e., multi-user MIMO (MU-MIMO); or, access network equipment can simultaneously transmit multiple data streams to the same terminal device, i.e., single-user MIMO (SU-MIMO). The data between these multiple terminal devices or the multiple data streams within the same terminal device are spatially multiplexed, thus becoming a key direction in the evolution of communication systems.

[0175] Access network equipment needs to obtain the CSI of the downlink channel to determine the resources, modulation and coding scheme (MCS), precoding and other configurations of the downlink data channel of the scheduling terminal equipment.

[0176] Taking precoding as an example, in massive MIMO, access network equipment needs to precode downlink data using a precoding matrix. By using precoding techniques, access network equipment can achieve spatial multiplexing between terminal devices or data streams. This means that data between different terminal devices or between different data streams of the same terminal device is spatially isolated, thereby reducing interference between different terminal devices or different data streams and improving the received signal-to-interference-plus-noise ratio (SINR) of the terminal devices. To calculate the precoding matrix, the access network equipment needs to obtain the CSI of the downlink channel and determine the precoding matrix based on the CSI.

[0177] In TDD systems, due to the reciprocity of uplink and downlink channels, access network devices can obtain the uplink CSI by measuring the uplink reference signal and then infer a relatively accurate downlink CSI, for example, using the uplink CSI as the downlink CSI. However, in FDD systems, uplink and downlink reciprocity cannot be guaranteed. The downlink CSI is obtained by the terminal device measuring the downlink reference signal, such as the channel state information reference signal (CSI-RS) or the synchronizing signal block (SSB). Therefore, the terminal device needs to generate a CSI report according to the protocol predefined method or the access network device configuration, and then feed the generated CSI report back to the access network device so that it can obtain the downlink CSI.

[0178] In FDD systems, a crucial part of CSI feedback is the precoding matrix indicator (PMI), which uses 0-1 bits in the CSI to quantize the channel matrix or precoding matrix. PMI design (also known as codebook design) is a fundamental issue in mobile communication systems. Traditional codebook design methods predefine (or agree upon) a series of precoding matrices and their corresponding numbers in the protocol; these precoding matrices are called codewords. The channel matrix or precoding matrix can be approximated using predefined codewords or linear combinations of multiple predefined codewords. Therefore, the terminal device can feed back the corresponding codeword numbers and one or more weighting coefficients to the access network device via the PMI, which is used by the access network device to reconstruct the channel matrix or precoding matrix.

[0179] As the antenna array size of MIMO systems continues to increase, the number of supported antenna ports also increases, leading to a growth in the dimensionality of the corresponding channel matrix and precoding matrix. To enable terminal devices to estimate (or measure) the downlink channel, the overhead of the access network equipment transmitting reference signals increases. Simultaneously, the error of approximating large-scale channel matrices and precoding matrices with a finite number of predefined codewords increases. One method to improve channel reconstruction accuracy is to increase the number of codewords in the codebook, but this simultaneously increases the overhead of CSI feedback (including the corresponding codeword number and one or more weighting coefficients), thereby reducing the available resources for data transmission and causing system capacity loss. Therefore, it is necessary to investigate how to more effectively compress and represent channel information without increasing the overhead of transmitting reference signals and CSI feedback, and how to more effectively reconstruct the channel based on feedback information. Correlation exists between different elements in the downlink channel matrix between access network equipment and terminal equipment; furthermore, correlation exists between downlink channel matrices in different time slots. For example, the correlation between different elements in the channel matrix implies the existence of a basis (which can be represented by matrices U1 and U2). Projecting the channel matrix H onto this basis yields a sparse equivalent channel, i.e., H' = U1H. H U2 is a sparse matrix, where the superscript H denotes the conjugate transpose operation. Theoretically, the channel matrix H can be reconstructed simply by estimating the non-zero elements in H' using the reference signal and feeding them back. Therefore, the overhead of the reference signal transmission and CSI feedback has room for compression. However, traditional CSI feedback schemes, such as the codebook-based feedback method mentioned above, do not fully utilize the channel compression space, and the channel compression process may cause significant information loss. Machine learning methods (such as deep learning (DL)) have stronger nonlinear feature extraction capabilities, thus enabling more effective extraction of correlations between channel matrices. Consequently, compared to traditional schemes, they can more effectively compress and represent channel information, and more effectively reconstruct channel information based on the feedback information.

[0180] In some implementations, neural network models can be used for compressed feedback of channel information. After the terminal measures the reference signal to obtain channel information, it inputs this information into its neural network model to obtain compressed channel information. The terminal then feeds back the compressed channel information to the network device. The network device inputs this compressed information into its neural network model to reconstruct the channel information measured by the terminal. This leverages the nonlinear feature extraction capabilities of neural networks to improve the accuracy of channel measurements.

[0181] As an example, such as Figure 20The diagram illustrates the application of a dual-ended model encoder and decoder in a channel information processing scenario. The terminal model is the encoder, used to compress CSI-RS measurement data. The terminal sends the compressed CSI-RS measurement data to the network device. Upon receiving the compressed CSI-RS measurement data, the network device uses its decoder to reconstruct the CSI-RS measurement data, obtaining the final CSI-RS measurement data. This allows for large-port, full-band channel measurement between the terminal and network device with limited CSI-RS measurement overhead, based on a dual-ended model. During CSI-RS measurement, multiple resource sets can be used to achieve multi-port channel measurement. Currently, one resource set supports a maximum of 32 antenna ports for channel measurement; adding multiple resource sets can support channel measurement with larger arrays. In related technologies, to support high-bandwidth channel measurement, the number of resource blocks (RBs) that the terminal needs to measure can be indicated. The density indicates the number of channels measured within one RB, and the value can be 0.5, 1, or 3.

[0182] In the embodiments of this application, combined with Figure 20 In the CSI-RS measurement scenario shown, the terminal can measure only a portion of the CSI-RS and compress that portion of the CSI-RS measurement data; correspondingly, the network device recovers the compressed portion of the CSI-RS measurement data to obtain the full CSI-RS measurement data. Alternatively, the terminal can measure only a portion of the CSI-RS and compress that portion of the CSI-RS measurement data; however, the network device not only recovers the compressed portion of the CSI-RS measurement data to obtain the full CSI-RS measurement data, but it can also predict the unmeasured CSI-RS measurement data to obtain the full CSI-RS measurement data. Alternatively, the terminal can measure only a portion of the CSI-RS and compress that portion of the CSI-RS measurement data; after the terminal sends the compressed portion of the CSI-RS measurement data multiple times, the network device predicts the full CSI-RS measurement data based on these multiple transmissions of the compressed portion of the CSI-RS measurement data.

[0183] To facilitate understanding of the embodiments of this application, the terms involved in this application will be briefly explained below.

[0184] 1. Channel State Information (CSI): Channel state information, also known as channel state information or channel environment information, is a type of information that reflects channel characteristics and channel quality.

[0185] Channel information measurement refers to the process by which the receiver deciphers the channel information based on a reference signal transmitted by the transmitter; that is, it involves estimating the channel information using channel estimation methods. For example, the reference signal may include one or more of the following: channel state information reference signal (CSI-RS), synchronizing signal / physical broadcast channel block (SSB), sounding reference signal (SRS), or demodulation reference signal (DMRS). One or more of CSI-RS, SSB, and DMRS can be used to measure downlink channel information. SRS and / or DMRS can be used to measure uplink channel information.

[0186] Channel information can be determined based on channel measurements of a reference signal. Alternatively, channel information can be derived from channel measurements of a reference signal.

[0187] In this embodiment, the meaning of CSI is broader than that in traditional schemes. It is not limited to precoding matrix indicator (PMI) or CSI-RS resource indicator (CRI). It can also be one or more of the following: channel response information (such as channel response matrix, frequency domain channel response information, time domain channel response information), weight information corresponding to the channel response, precoding matrix information corresponding to the channel response, channel matrix, channel feature matrix, precoding matrix, reference signal receiving power (RSRP), signal to interference plus noise ratio (SINR), identity (ID) of the optimal beam, or ID of the top K beams. The signal to interference plus noise ratio can also be called the signal-to-interference-plus-noise ratio. The optimal beam refers to the beam that maximizes the received or transmitted energy. For example, if the receiver uses different receiving beams to receive signals, the optimal beam may include the beam with the largest RSRP among the signals received from multiple different receiving beams. For example, if the transmitter uses different transmit beams to send signals, the optimal beam can include the beam with the highest RSRP (Receiving Power Ratio) of the signal when it arrives at the receiver. Similarly, the optimal top K beams refer to the top K beams that maximize the received or transmitted energy.

[0188] The following are some of the CSI-related terms used in this application:

[0189] Target CSI: Also known as the full information of CSI or ground-truth CSI. For ease of distinction and explanation, target CSI will be represented by H4 in the following text.

[0190] CSI feedback information: also known as CSI feedback information, channel measurement result feedback information, channel information feedback information, compressed information, compressed channel information, compressed CSI information, compressed channel information, or compressed CSI, etc. For ease of distinction and explanation, CSI feedback information will be represented by C2 in the following text.

[0191] Input CSI: Also known as measurement CSI, raw CSI, uncompressed CSI information, or raw CSI information. Input CSI can be understood as the CSI input to the AI ​​model (such as the first model mentioned below), meaning the input CSI can be the input parameters of the AI ​​model. For ease of distinction and explanation, input CSI will be represented by H1 in the following text.

[0192] It should be noted that in the embodiments of this application, during the training phase, CSI feedback information can refer to the information obtained after compressing the input CSI, which can be simply referred to as compressed CSI; or it can refer to the information obtained by compressing and quantizing the input CSI, which can be simply referred to as quantized CSI. This application does not limit this. During the model deployment phase, CSI feedback information can be information sent from the terminal side to the network side. To save feedback overhead, CSI feedback information can refer to the information obtained by compressing and quantizing the measured CSI (also called input CSI), i.e., quantized CSI.

[0193] It should be understood that quantizing compressed CSI yields quantized CSI; dequantizing quantized CSI yields compressed CSI. Compressed CSI obtained by dequantizing quantized CSI is compressed CSI with quantization loss.

[0194] Reconstructed CSI: Also known as recovered CSI, reconstructed CSI, reconstructed channel information, decompressed CSI, decompressed channel information, or output CSI, etc. Reconstructed CSI can be understood as the CSI output by the AI ​​model; that is, the reconstructed CSI can be the output parameter of the AI ​​model (such as the second model mentioned below). For example, if the input to the AI ​​model is compressed CSI, then the reconstructed CSI output by the AI ​​model can be understood as the CSI obtained by decompressing the compressed CSI. In this case, the purpose of decompressing the compressed CSI is to obtain a CSI that is infinitely close to the uncompressed CSI; therefore, the reconstructed CSI can also be called recovered CSI. As another example, if the input to the AI ​​model is quantized CSI obtained by quantizing compressed CI, then the reconstructed CSI output by the AI ​​model can be understood as the CSI obtained by decompressing and dequantizing the quantized CSI. In this case, the AI ​​model decompresses and dequantizes the quantized CSI to obtain a CSI that is infinitely close to the uncompressed CSI; therefore, the reconstructed CSI can also be called the recovered CSI. For ease of distinction and explanation, the reconstructed CSI will be referred to below using H3 or... express.

[0195] 2. AI Model: A function model that maps an input of a certain dimension to an output of a certain dimension. Its parameters can be obtained through machine learning (ML). For example, f(x) = ax 2 +b is a quadratic function model, which can be viewed as an AI model. a and b correspond to the parameters of the model and can be obtained through machine learning training.

[0196] 3. AE Model: This can generally refer to a network structure consisting of two AI models, such as an encoder and a decoder. Each model can be an AI model. AE models are also called bilateral models, two-end models, collaborative models, etc. The encoder and decoder of an AE are usually trained together and can be used in a coordinated manner.

[0197] In this application, the feedback of CSI can be implemented based on the AI ​​model of AE. Figure 5 This is a schematic diagram of CSI feedback using an AE model provided in an embodiment of this application.

[0198] As shown in the figure, the terminal side compresses the target CSI using an encoder, while the network side reconstructs the CSI using a decoder. For example, the terminal side can use the target CSI as input to the encoder, which compresses the target CSI to obtain the compressed CSI. The terminal side can then quantize the compressed CSI to obtain the quantized CSI. The terminal side can then send the quantized CSI as CSI feedback information to the network side, for example, through a CSI report.

[0199] The network side can first dequantize the CSI feedback information to obtain compressed CSI with quantization loss. The network side can then use this compressed CSI as input to the decoder, which can decompress the CSI to obtain the reconstructed CSI.

[0200] The quantizer used to perform quantization can be predefined, such as protocol predefined, or it can be indicated by the network side; this application does not limit this.

[0201] In another implementation, the encoder on the terminal side can also compress and quantize the target CSI, outputting CSI feedback information. The decoder on the network side can also dequantize and decompress the CSI feedback information to obtain the reconstructed CSI. In this case, Figure 5 The encoder output is CSI feedback information, and the decoder input is also CSI feedback information.

[0202] It should be noted that the encoder on the terminal side can be deployed inside the terminal device or in other devices outside the terminal device, such as the aforementioned OTT server; the decoder on the network side can be deployed inside the network device or in other devices outside the network device, such as the aforementioned intelligent network element.

[0203] It should be understood that although the figure shows an encoder and a decoder, this is only a model division from a functional perspective. The encoder can also be called the first model, and the decoder can also be called the second model, etc. This application does not limit this. In addition, this application does not limit the number of models included in the AE model.

[0204] It should also be understood that the AE model is only one possible model for achieving the above functions and should not constitute any limitation on this application. The AE model can also be replaced by other AI models that can achieve the same or similar functions.

[0205] 4. Model training: By selecting an appropriate function (such as a loss function), the model parameters are trained using optimization algorithms to minimize the difference between the model's predicted values ​​and the ground truth (or, the true value, the target value, and the label).

[0206] For example, model training methods include, but are not limited to, supervised learning, self-supervised learning, and knowledge distillation. These methods are briefly explained below.

[0207] Supervised learning, also known as supervised instruction, involves using machine learning algorithms to learn the mapping relationship between sample values ​​and labels based on collected sample values ​​and labels. This learned mapping relationship is then expressed using a machine learning model. The process of training the machine learning model is the process of learning this mapping relationship. For example, in signal detection, the noisy received signal is the sample, and the corresponding real constellation point is the label. Machine learning aims to learn the mapping relationship between samples and labels through training, that is, to enable the machine learning model to learn a signal detector. During training, the model parameters are optimized by calculating the error between the model's predicted values ​​and the real labels. Once the mapping relationship is learned, it can be used to predict the sample label of each new sample. The mapping relationship learned in supervised learning can include linear and nonlinear mappings. Based on the type of label, the learning task can be divided into classification tasks and regression tasks.

[0208] Self-supervised learning is a type of unsupervised learning. Unsupervised learning relies on collected sample values ​​to allow algorithms to discover inherent patterns within the samples themselves. Self-supervised learning uses the samples themselves as supervisory signals; that is, the model learns the mapping relationship from sample to sample. During training, model parameters are optimized by calculating the error between the model's predicted values ​​and the actual samples. Self-supervised learning can be used in signal compression and decompression recovery applications; common algorithms include autoencoders and generative adversarial networks.

[0209] Knowledge distillation: Generally, large models are often single complex networks or collections of networks, possessing excellent performance and generalization ability, while small models, due to their smaller network size, have limited expressive power. Therefore, the knowledge learned by the large model can be used to guide the training of the small model; this process is called knowledge distillation. Knowledge distillation can enable small models to achieve performance comparable to large models, but with fewer parameters and shorter inference latency, thus achieving model compression and acceleration. Furthermore, directly training small models with massive amounts of data often does not yield good performance, while training large models with massive amounts of data and then using the large model to perform knowledge distillation on the small model can achieve better continuation results. In addition, knowledge distillation can also be used to integrate and transfer datasets from different domains.

[0210] Knowledge distillation employs a teacher-student model, where a teacher model assists in training a student model. The teacher model is a complex, large model, while the student model is a simple, small model. Because the teacher model has strong learning capabilities, it can transfer the knowledge it learns to the relatively weaker student model, thereby enhancing the student model's generalization ability. The complex, cumbersome, but effective teacher model remains offline, simply acting as a mentor; the flexible and lightweight student model is the one actually deployed for prediction tasks.

[0211] The model training methods listed above are merely examples, and this application does not limit the methods used for model training.

[0212] Model training in this application embodiment may include initial training (or original training) or retraining of the model. Model training enables model enhancement. Model enhancement can refer to improving model functionality through model structure optimization and / or model training, such as adding other functions or neural networks to the existing model functionality, so that the enhanced model can be applied to more scenarios or richer requirements.

[0213] 5. Performance metric function: also simply referred to as a function. The performance metric function in this application is used to describe the performance of the model, specifically to describe the difference between the model's predicted values ​​and the true values. Model training can be terminated when the value of the performance metric function reaches the range required by the performance index.

[0214] The performance metric function can be a loss function, also known as an objective function or cost function, used to measure the difference between the predicted and actual values. A higher output value (loss) indicates a greater difference, and model training can be a process of minimizing this difference. Loss functions involved in the embodiments of this application include, for example, L1Loss (also known as mean absolute error (MAE), MSE, and NMSE).

[0215] The performance metric function can also be used to measure the similarity between predicted and actual values. In this application, examples of performance metric functions used to measure the similarity between predicted and actual values ​​include GCS and SGCS.

[0216] The calculation methods for each performance metric function can be found in the formulas below:

[0217]

[0218] in, These are predicted values, including M values: H represents the true values, including M values: h0, h1, ..., h M-1 .

[0219] In another implementation, the above GCS and SGCS can also be replaced by the following loss functions: the difference between GCS and 1, and the difference between SGCS and 1. In this case, the model training process is the process of reducing this difference.

[0220] In the embodiments of this application, the performance metric function and the performance index are corresponding. For example, the performance metric function is NMSE, and the corresponding performance index is the upper bound of NMSE; another example is that the performance metric function is SGCS, and the corresponding performance index is the lower bound of SGCS, or the upper bound of the difference between SGCS and 1, and so on, and will not be listed further.

[0221] Furthermore, this application does not impose any restrictions on the specific values ​​of the performance metrics (such as upper and / or lower bounds) corresponding to each performance metric function.

[0222] 6. Model Files and Model Parameters: Model files and / or model parameters can be used to determine the model. Therefore, the model in this application can refer to the model itself, or it can refer to the model files and / or model parameters used to determine the model.

[0223] The model file can be used to indicate the model structure, which includes, but is not limited to, feedforward neural networks (FNNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs). The model file can have a fixed format, such as a standard predefined format or a format pre-negotiated by both ends of the connection. Model parameters can refer to parameters in the neural network model, such as, but not limited to, the number of layers in the neural network, the type and weights of neurons in each layer, etc. This application does not limit the method of distributing model parameters.

[0224] Take DNN as an example. The idea behind DNN comes from the neuronal structure of the brain. Each neuron can perform a weighted summation operation on its inputs and then use the result of the weighted summation operation to generate the output through a nonlinear function. Figure 6 This is an example of a neuron structure. Figure 6 The input to the neuron shown is x = [x0x1…x] N-1 The weights corresponding to the inputs are w = [w0w1…w] N-1The bias of the weighted summation is b. The nonlinear function f() can take many forms; for example, the nonlinear function f() can be the maximum value function max{0, x}. Then the effect of a neuron's execution is... Where N is a positive integer; n is a positive integer greater than or equal to 0 and less than or equal to (N-1).

[0225] A typical DNN has multiple neural network layers, including an input layer, one or more hidden layers, and an output layer. Generally, the first layer is the input layer, the last layer is the output layer, and the layers in between are hidden layers. Each layer contains multiple neurons. Layers are fully connected; that is, any neuron in the i-th layer is connected to any neuron in the (i+1)-th layer. The input layer processes the received values ​​(i.e., the DNN's input) through neurons and then passes them to the hidden layers. Similarly, the hidden layers pass the computation results to the final output layer, producing the DNN's output. Figure 7 An example of a DNN is shown. Figure 7 The DNN model shown has three neural network layers: an input layer, a hidden layer, and an output layer.

[0226] It should be understood that the above text, in combination with... Figure 6 and Figure 7 The examples provided are for illustrative purposes only and should not be construed as limiting the scope of this application. This application does not limit the structure and parameters used in the AI ​​model.

[0227] One of the model structure or model parameters can be predefined, while the other can be sent by the sender (e.g., the network side). Alternatively, both the model structure and model parameters can be sent by the sender (e.g., the network side). This application does not impose any restrictions on this.

[0228] 7. Two-way connection: This refers to the connection between the sender (e.g., the network side) and the receiver (e.g., the terminal side). It can be a connection between datasets or between models. A dataset is used in machine learning for model training, validation, and testing; the quantity and quality of the dataset affect the effectiveness of machine learning. In this embodiment, the dataset can be used for model training.

[0229] The following section uses CSI compressed inference tasks as an example to explain the connection between the dataset and the model.

[0230] Dataset integration mainly refers to the sender sending a dataset to the receiver for model training. Figure 8 This is a schematic diagram illustrating the data set connection between the network side and the terminal side.

[0231] like Figure 8As shown, the network can first acquire the dataset through joint training. For example, the network can use a CSI compression model to compress the target CSI, and then use a CSI reconstruction model to decompress the target CSI. That is, the target CSI is the input to the CSI compression model, and the output of the CSI compression model can be the compressed CSI. The compressed CSI can be the input to the CSI reconstruction model, and the CSI reconstruction model can output the reconstructed CSI.

[0232] Optionally, the compressed CSI can be further quantized to obtain quantized CSI. During the model training phase, both the compressed CSI and the quantized CSI can be used as CSI feedback information.

[0233] For example, the dataset sent from the network side to the terminal side may include one or more of the following: target CSI, CSI feedback information, or reconstructed CSI. Optionally, the dataset may also include a quantization method. In other words, the quantization method is also indicated by the network side. Optionally, the quantization method is predefined, such as protocol predefined. The quantization method is used to determine how to quantize the compressed CSI and / or how to dequantize the quantized CSI. The quantization method can also be replaced by a dequantization method, a quantizer, or a dequantizer.

[0234] The network side can distribute this dataset to the terminal side. The terminal side can then train a model based on the received dataset to obtain a CSI compressed model for the terminal side.

[0235] Model integration mainly refers to the sender sending model files and / or model parameters to the receiver for the receiver's model deployment, or for the receiver's model development (e.g., model training) and model deployment. Figure 9 This is a schematic diagram illustrating the model interoperability between the network side and the terminal side.

[0236] like Figure 9 As shown, the network side can first obtain the CSI compression model through joint training. The specific process of network-side joint training can be found in the above text. Figure 8 The explanation will not be repeated here. The network side can distribute the CSI compressed model or CSI reconstructed model obtained from joint training to the terminal side. For example, the network side can send the model file and / or model parameters of the CSI compressed model to the terminal side, or send the model file and / or model parameters of the CSI reconstructed model to the terminal side. The terminal side can then train the model based on the received model file and / or model parameters.

[0237] It should be understood that the above text, in combination with... Figure 8 and Figure 9 The CSI compression model shown can be an encoder, and the CSI reconstruction model can be a decoder.

[0238] As mentioned earlier, dual-end AI models are typically co-trained and can be used interchangeably. To support the integration and development / deployment of dual-end AI models, datasets or models need to be provided to one or both of the models. For example, the network side can provide a dataset to the terminal side for model development and / or deployment; conversely, the network side can provide a model to the terminal side for model development and / or deployment. (Based on the above...) Figure 8 and Figure 9 As can be seen from the description, the data contained in the dataset is not fixed; datasets may be of various types, and different types of datasets contain different data. Therefore, how the terminal side can train the model based on the received dataset has become a technical problem that urgently needs to be solved.

[0239] In view of this, this application provides a method to associate the dataset used for training an asymmetric dual-end model with performance metrics, and to associate the performance metrics with inference tasks. This allows the terminal to train the model using appropriate training methods based on the received dataset and the corresponding performance metrics, thereby obtaining a model that meets the requirements and improving the feedback performance of CSI.

[0240] The method provided in this application will now be described in detail with reference to the accompanying drawings.

[0241] To facilitate understanding, before introducing the method provided in this application, the channel information involved in the following embodiments will be explained first.

[0242] Channel information refers to channel-related information between the first device side and the second device side (e.g., the terminal side and the network side), which is information about the path and / or the measured channel obtained by the device through channel measurement.

[0243] For a more detailed description of the channel information, please refer to the description of CSI in the terminology section above. That is, the channel information in the embodiments of this application can be replaced with CSI, and will not be repeated here.

[0244] A portion of channel information A can be referred to as a subset of channel information. For example, if channel information A corresponds to channel A, then the channel corresponding to a portion of channel information A is a subset of channel A. For instance, channel A corresponding to channel information A might be a channel between the first and second device sides, corresponding to resource A, while the channel corresponding to a portion of channel information A might be a channel between the first and second device sides, corresponding to resource B. Here, resource B is a subset of resource A. For example, resource A includes resource B and resource C, which is different from resource B.

[0245] Channel information A is different from channel information B. This can be understood as channel A corresponding to channel information A being different from channel B corresponding to channel information B. For example, channel A is the channel between the first device side and the second device side, corresponding to resource A, while channel B is the channel between the first device side and the second device side, corresponding to resource D. Resource A is different from resource D.

[0246] Channel information A includes channel information C, which can be understood as channel information A including channel information C itself, or as channel information A including channel information C'. The similarity between channel information C' and channel information C is greater than threshold #A, or the difference between channel information C' and channel information C is less than threshold #B. The similarity or difference between channel information C' and channel information C can be characterized by one or more of the following: MSE, NMSE, MAE, SGCS, or GCS.

[0247] The aforementioned resources may include one or more of the following: time-domain resources, frequency-domain resources, or spatial-domain resources.

[0248] A time-domain resource may include one or more time-domain units. A time-domain unit can be any of the following: a slot, a subframe, a frame, an orthogonal frequency division multiplexing (OFDM) symbol, a second (s), or a millisecond (ms). Taking time-domain resources A and B as examples, if time-domain resources A and B are different, then time-domain resources A includes different time-domain units than time-domain resources B. For example, time-domain resources A includes slot #a and slot #b. If time-domain resources A and B are the same, then time-domain resources A and B include the same time-domain units. For example, both time-domain resources A and B include slot #a. If time-domain resources A include time-domain resources B, then time-domain resources A includes at least the time-domain units belonging to time-domain resources B. For example, if time-domain resources B include slot #a, then time-domain resources A includes at least slot #a.

[0249] Frequency domain resources can include one or more frequency domain elements. A frequency domain element can be any of the following: subcarrier, component carrier (CC), resource block (RB), subchannel, resource pool, bandwidth, bandwidthpart (BWP), channel, physical resource block (PRB), resource blockgroup (RBG), PRB bundling, or an interlaced RB, etc. Taking frequency domain resources A and B as examples, if frequency domain resources A and B are different, then the frequency domain elements included in frequency domain resources A are different from those included in frequency domain resources B. For example, frequency domain resources A includes subband #a, and frequency domain resources B include subband #b. If frequency domain resources A and B are the same, then frequency domain resources A and B include the same frequency domain elements. For example, both frequency domain resources A and B include subband #a. If frequency domain resource A includes frequency domain resource B, then frequency domain resource A must include at least the frequency domain elements belonging to frequency domain resource B. For example, if frequency domain resource B includes subband #a, then frequency domain resource A must include at least subband #a.

[0250] A spatial resource is associated with one or more of the following: a sending port, a receiving port, a sending stream (a stream can be replaced with a layer), or a receiving stream. For example, if spatial resource A is different from spatial resource B, then spatial resource A differs from spatial resource B in at least one of the following: a sending port, a receiving port, a sending stream, or a receiving stream. For example, if spatial resource A includes spatial resource B, then the sending port / receiving port corresponding to spatial resource A includes the sending port / receiving port corresponding to spatial resource B, and / or, the sending stream / receiving stream corresponding to spatial resource A includes the sending stream / receiving stream corresponding to spatial resource B. If spatial resource A includes spatial resource B, then each of the above-mentioned items corresponding to spatial resource A includes at least the corresponding item corresponding to spatial resource B. For example, if spatial resource B corresponds to sending port A and sending stream A, then the sending port corresponding to spatial resource A includes at least sending port A, and the sending stream corresponding to spatial resource A includes at least sending stream A.

[0251] If a resource includes multiple resources in the time domain, frequency domain, or spatial domain, then resource A1 is different from resource A2 if at least one of the multiple resources included in resource A1 and resource A2 is different; if resource A1 is the same as resource A2, then every single one of the multiple resources included in resource A1 and resource A2 is the same; if resource A1 includes resource A2, then every single one of the multiple resources included in resource A1 includes at least the corresponding single one included in resource A2. For example, resource A1 includes time domain resource A1 and frequency domain resource A2, and resource A2 includes time domain resource A2 and frequency domain resource A2. Resource A1 being different from resource A2 means that time domain resource A1 and time domain resource A2 are different, and / or, frequency domain resource A1 and frequency domain resource A2 are different. Resource A1 being the same as resource A2 means that time domain resource A1 and time domain resource A2 are the same, and frequency domain resource A1 and frequency domain resource A2 are the same. Resource A1 includes resource A2, meaning that time-domain resource A1 includes at least time-domain resource A2, and frequency-domain resource A1 includes at least frequency-domain resource A2.

[0252] It should also be noted that in the embodiments below, the different data numbers are merely for the purpose of distinguishing different types of data, and do not constrain the size or quantity of the data. For example, "first data" and "second data" in the embodiments below are simply data of different types or purposes, and do not limit the quantity, size relationship, or priority relationship of the data. Exemplarily, first data may include one or more data. Exemplarily, second data may include one or more data. Exemplarily, a single data may also be referred to as a sample, channel measurement result, or channel information, etc.

[0253] It should also be noted that in the embodiments below, "model" can be replaced by names such as encoder, decoder, compressor, decompressor, compression quantizer, decompression and dequantizer, quantizer, or dequantizer, etc., and this application does not limit this. For example, if the inference task corresponding to the model (such as the first model) is to compress data, then the model can be replaced by encoder, compressor, compression quantizer, or quantizer, etc. As another example, if the inference task corresponding to the model (such as the second model) is to decompress data, then the model can be replaced by decoder, decompressor, decompression and dequantizer, or dequantizer, etc.

[0254] Figure 10 This is a schematic flowchart of the communication method 1000 provided in the embodiments of this application. Figure 10 The method shown illustrates the flow of method 1000 from the perspective of interaction between the terminal side and the network side. More detailed explanations regarding the terminal side and the network side can be found above and will not be repeated here.

[0255] Figure 10The method 1000 shown may include steps 1010 to 1020. Optionally, the method may also include one or more of steps 1030a to 1070. The various steps in method 1000 are described in detail below.

[0256] In step 1010, the terminal side acquires a dataset, which is used for model training.

[0257] In the embodiments of this application, the device that performs model training on the terminal side can be referred to as a training device. The training device can be a terminal device, or it can be a device other than a terminal device, such as the aforementioned OTT system server. This application does not limit this.

[0258] In one example, the training device is a terminal device, and step 1010 may specifically include: the terminal device receiving a dataset from a network device, such as receiving it over the air interface; or, the terminal device obtaining a dataset from an OTT system server, which is received by the OTT system from a smart network element on the network side, such as receiving it through a wired network.

[0259] In another example, the training device is an OTT system server, and step 1010 may specifically include: the OTT system server receiving a dataset from a smart network element, such as receiving it through a wired network; or, the OTT system server obtaining a dataset from a terminal device, which is received by the terminal device from the network side, such as a network device, for example, receiving it through an air interface.

[0260] In summary, whether the data set is obtained by the terminal device or the OTT system server, it is received from the network side. Therefore, one possible implementation of step 1010 is that the terminal receives the data set from the network side, and correspondingly, the network side sends the data set to the terminal.

[0261] Sending the dataset can refer to transmitting the dataset via a message carrier. The network side can send the dataset to the terminal side through one or more messages, and this application does not limit this. The figure is only an example, showing the process of sending the dataset from the network side to the terminal side through one step, and should not constitute any limitation on this application.

[0262] Optionally, before sending the dataset from the network side to the terminal side, method 1000 includes step 1010a: the network side obtains the dataset. For example, the network side obtains the dataset by running a fifth model and a second model. The inference task of the fifth model is to compress the input CSI or to compress and quantize it. For example, the inference task of the fifth model is the same as the inference task of the first model described below. The inference task of the second model is described below. The method by which the network device jointly trains the fifth model and the second model can be found in the description in Example 1 below.

[0263] For example, if the device performing model training on the network side is a network device, then in step 1010a, the network device obtains the dataset based on the fifth model and the second model, for example, based on the trained fifth model and the second model. Optionally, the network device can also send the dataset to devices outside the network device, such as intelligent network elements.

[0264] For example, if the device performing model training on the network side is a device other than the network device, such as an intelligent network element, then in step 1010a, the intelligent network element obtains the dataset through model training. Optionally, the intelligent network element can also send the dataset to the network device.

[0265] In step 1020, the terminal side trains the model based on the dataset and performance metrics to obtain the first model.

[0266] Here, the first model is the model that the terminal side hopes to obtain through model training. The inference task of this first model can be to perform a first process on the measured CSI to obtain CSI feedback information. The measured CSI can be called the input CSI. If the measured CSI is denoted as the first data and the CSI feedback information is denoted as the second data, then the inference task of the first model can be said to be to perform a first process on the first data to obtain the second data. The first processing includes compression, or the first processing includes compression and quantization.

[0267] The first data is used to characterize the first channel information. The second data is used to determine the third data based on the second model.

[0268] The third data is used to characterize the second channel information. The second channel information includes at least one of the following: a portion of the first channel information, all of the first channel information, or third channel information. The third channel information is different from the first channel information. For example, the second channel information includes the third channel information, and also includes a portion or all of the first channel information. The third data may also be replaced with names such as reconstructed CSI or predicted CSI.

[0269] The inference task of the second model includes processing the second data to obtain the third data. For example, the inference task of the second model may be to decompress the second data to obtain the third data. Alternatively, the inference task of the second model may be to decompress and dequantize the second data to obtain the third data. As an example, the second model is deployed on the network side.

[0270] It should be understood that the second model is a model that matches the first model; that is, the combination of the first and second models is a two-ended model. Based on the description above, it can be seen that the input parameters of the first model (such as the first data representing the first channel information) and the output parameters of the second model (such as the third data representing the second channel information) may be different. Therefore, the combination of the first and second models can be called an asymmetric two-ended model.

[0271] Examples of first data and third data for different types of channel information are as follows.

[0272] For example, if the first channel information is a channel matrix, then the first data is the first channel information itself. Correspondingly, the third data is a channel matrix, and the second channel information is a channel matrix. The first channel information can be a channel matrix between the terminal and the network, corresponding to resource X. Similarly, the second channel information can be a channel matrix between the terminal and the network, corresponding to resource Y. Resource Y is different from resource X. For example, resource Y includes resource X. Or, resource Y and resource X do not overlap at all.

[0273] The channel matrix described above can also be replaced with RSRP, SINR, etc.

[0274] For example, if the first channel information is a channel feature matrix, then the first data can be a channel matrix recovered from the channel feature matrix. Correspondingly, the third data is a channel matrix, and the second channel information can be either a channel matrix or a channel feature matrix. Specifically, the first channel information can be the channel feature matrix of the channel matrix corresponding to resource X between the terminal and the network. The second channel information can be the channel matrix or the channel feature matrix of the channel matrix corresponding to resource Y between the terminal and the network.

[0275] The aforementioned channel feature matrix can be replaced by PMI, CRI, precoding matrix, or channel response information, etc.

[0276] In this embodiment of the application, the performance metrics correspond to the inference task of the first model.

[0277] Since performance metrics may vary depending on the dataset, and the model training process may also differ, the terminal may determine whether the performance metric can be met during the training of the first model, or it may determine whether the performance metric can be met during the training of other models, which could be the models used to train the first model. In other words, the terminal may have already started training the first model before determining whether the performance metric can be met, or it may not have started training the first model yet.

[0278] Therefore, the performance metric corresponding to the inference task of the first model may include: the performance metric being the performance requirement that should be met when performing the inference task of the first model, and / or, the performance metric being the performance requirement that should be met when performing the inference task through the fourth model. The fourth model includes the first model. For example, the inference task of the first model is inference that compresses the input CSI. Correspondingly, the inference task of the fourth model includes: inference that compresses the input CSI to obtain CSI feedback information, and inference that decompresses the CSI feedback information.

[0279] For example, the performance metric can specifically be the range of performance requirements that the values ​​of one or more loss functions should meet, such as the accuracy range. This loss function can include, but is not limited to, GCS, SGCS, MSE, NMSE, MAE, etc. MAE can also be called L1 loss (L1 Loss).

[0280] For example, the value of MSE may include the average MSE and / or the distribution of MSE. Accordingly, the range of performance requirements that the value of MSE should meet may include: the range of performance requirements that the average MSE should meet, and / or the range of performance requirements that the distribution of MSE should meet.

[0281] The MAE value can include the average MAE and / or the distribution of MAE. Accordingly, the range of performance requirements that the MAE value should meet can include: the range of performance requirements that the average MAE should meet, and / or the range of performance requirements that the distribution of MAE should meet.

[0282] The value of GCS can include the average GCS and / or the distribution of GCS. Accordingly, the range of performance requirements that the value of GCS should meet can include: the range of performance requirements that the average GCS should meet, and / or the range of performance requirements that the distribution of GCS should meet.

[0283] The value of SGCS can include either the average SGCS or the distribution of SGCS. Correspondingly, it refers to the range of performance requirements that the SGCS value should meet, including the range of performance requirements that the average SGCS should meet. Alternatively, it refers to the range of performance requirements that the SGCS value should meet, including the range of performance requirements that the distribution of SGCS should meet.

[0284] The NMSE value can include either the average NMSE or the distribution of NMSE in the SGCS. Correspondingly, the range of performance requirements that the NMSE value should meet includes the range of performance requirements that the average NMSE should meet. Alternatively, the range of performance requirements that the NMSE value should meet includes the range of performance requirements that the distribution of NMSE should meet.

[0285] Taking SGCS as an example, the average SGCS refers to the average of the SGCS values ​​between the model's K outputs and the labels. The distribution of SGCS can be the SGCS value corresponding to the L% points on the cumulative distribution function (CDF) distribution curve determined by the SGCS values ​​between the model's K outputs and the labels. Here, the model's K outputs refer to the model's outputs during training, and K is a positive integer. 0 < L < 100, for example, L = 5, 10, 20, or 30.

[0286] For example, the CDF determined by the SGCS between the K outputs of the model and the label is defined as: F(z) = P(S≤z). Here, T represents the SGCS between the model's output and the label. F(z) represents the probability that the SGCS between the model's output and the label is less than z. Taking L=20 as an example, the 20% point on the CDF distribution curve determined by the SGCS between the K outputs of the model and the label refers to a value s that satisfies the following condition. 0.2 :F(s 0.2 )=P(S≤s 0.2 = 0.2. In other words, it means that there is a 20% probability that the SGCS between the model's output and the label does not exceed s. 0.2 .

[0287] The average value or distribution of functions such as MSE, NMSE, MAE, or GCS can be found in the description above and will not be repeated here.

[0288] The loss function can be predefined, such as protocol predefined, or it can be indicated by the network side; this application does not limit this.

[0289] For example, taking performance metric a and performance metric b as examples, performance metric a and performance metric b are different, including the following situations.

[0290] Scenario 1: Performance index 'a' refers to the range of performance requirements that the values ​​of functions in function set 'a' should satisfy, and performance index 'b' refers to the range of performance requirements that the values ​​of functions in function set 'b' should satisfy. Furthermore, function set 'a' and function set 'b' are not entirely identical. Function set 'a' or function set 'b' includes one or more of the aforementioned GCS, SGCS, MSE, NMSE, or MAE. For example, function set 'a' includes GCS and SGCS. Function set 'b' includes MSE, NMSE, and MAE.

[0291] In scenario 2, both performance index 'a' and performance index 'b' refer to the range of performance requirements that the values ​​of functions in function set 'a' should satisfy. However, the range of performance requirements that function 'a' should satisfy, as indicated by performance index 'a', differs from the range of performance requirements that function 'a' should satisfy, as indicated by performance index 'b'. Function 'a' belongs to function set 'a'.

[0292] The following describes how performance metrics are obtained on the terminal side.

[0293] Alternatively, the performance metric is indicated by the network side. The network side can indicate this performance metric to the terminal side via information (the first information shown below).

[0294] Accordingly, prior to step 1020, method 1000 further includes 1030a: the terminal side receives first information, which is used to indicate the performance indicator. Accordingly, the network side sends the first information.

[0295] Similar to step 1010, in one example, the device performing model training on the terminal side is a terminal device. Step 1030a may include: the terminal device receiving first information from the network device, or the terminal device receiving first information from the OTT server. The information indicating the performance metric (e.g., the first information, or other information) may be received by the OTT system server from the intelligent network element. In another example, the device performing model training on the terminal side is an OTT system server. Step 1030a may include: the OTT system server receiving first information from the intelligent network element, or the OTT system server receiving first information from the terminal device. The information indicating the performance metric (e.g., the first information, or other information) may be received by the terminal device from the network side, such as the network device.

[0296] For example, as described above, the performance metric may be a range of performance requirements that the values ​​of one or more loss functions should satisfy. Therefore, the first information indicates the performance metric, including: the first information indicates at least one loss function, and / or indicates a range of performance requirements that the values ​​of the at least one loss function should satisfy.

[0297] For example, the first information may include an identifier, number, or index of at least one loss function to indicate at least one loss function. Where the first information indicates at least one loss function, it may also indicate the range of performance requirements that the value of the loss function should satisfy; alternatively, the range of performance requirements that the value of the loss function should satisfy may be predefined.

[0298] For example, the first information may include an identifier, number, or index corresponding to the range of performance requirements that the value of at least one loss function should satisfy, to indicate the range of performance requirements that the value of at least one loss function should satisfy. In cases where the first information indicates the range of performance requirements that the value of at least one loss function should satisfy, the first information may also indicate the loss function, or the loss function may be predefined.

[0299] For example, the network side determines the performance metric indicated by the first information sent to the terminal side based on the correspondence between different datasets and different performance metrics. Table 1 below shows an example of the correspondence between different datasets and different performance metrics. As shown in Table 1, if the terminal side obtains dataset 1 in step 1010, then in step 1030a, the first information sent by the network side indicates performance metric 1 or performance metric 2. If the terminal side obtains dataset 2 in step 1010, then in step 1030a, the first information sent by the network side indicates performance metric 3. If the terminal side obtains dataset 3 in step 1010, then in step 1030a, the first information sent by the network side indicates performance metric 4.

[0300] Table 1

[0301] Dataset types Model training methods Performance indicators Dataset 1 Training Method 1 Performance Indicator 1 Dataset 1 Training Method 2 Performance index 2 Dataset 2 Training Method 3 Performance index 3 Dataset 3 Training Method 4 Performance index 4

[0302] Optionally, performance metrics are related to measurement configurations; in other words, different measurement configurations corresponding to the same dataset may correspond to different performance metrics. Accordingly, the network side can determine the performance metrics indicated by the first information sent to the terminal side based on the correspondence between different datasets, different measurement configurations, and different performance metrics. Measurement configuration refers to the reference signal measurement configuration, i.e., the measurement configuration is used to configure the reference signal. The measurement results of the reference signal configured by the measurement configuration are related to the first channel information. Table 2 below shows an example of the correspondence between dataset 1, different measurement configurations, and different performance metrics. For example, the measurement configuration may include one or more of the following: antenna port configuration, subband configuration, or payload configuration. Antenna port configuration indicates the antenna port on the network side that transmits the reference signal and / or the antenna port on the terminal side that receives the reference signal. Subband configuration indicates the subband on the network side that transmits the reference signal and / or the subband on the terminal side that receives the reference signal. Payload configuration indicates the number of bits of CSI feedback overhead.

[0303] As shown in Table 2, in step 1010, the terminal side obtains dataset 1, and the network side configures the measurement configuration for the terminal side as antenna port configuration 1 and subband configuration 1. Then, in step 1030a, the first information sent by the network side indicates performance indicator 1a. In step 1010, the terminal side obtains dataset 1, and the network side configures the measurement configuration for the terminal side as antenna port configuration 1 and subband configuration 2. Then, in step 1030a, the first information sent by the network side indicates performance indicator 1b. In step 1010, the terminal side obtains dataset 1, and the network side configures the measurement configuration for the terminal side as antenna port configuration 3 and subband configuration 2. Then, in step 1030a, the first information sent by the network side indicates performance indicator 1c. In step 1010, the terminal side obtains dataset 1, and the network side configures the measurement configuration for the terminal side as antenna port configuration 1, subband configuration 1, and load configuration 1. Then, in step 1030a, the first information sent by the network side indicates performance indicator 2a. In step 1010, the terminal side obtains dataset 1, and the network side configures the measurement configuration for the terminal side as antenna port configuration 1 and sub-band configuration 3. Then, in step 1030a, the first information sent by the network side indicates performance index 2b.

[0304] Table 2

[0305] Dataset types Measurement configuration Model training methods Performance indicators Dataset 1 Antenna port configuration 1, sub-band configuration 1 Training Method 1 Performance Indicator 1a Dataset 1 Antenna port configuration 1, sub-band configuration 2 Training Method 1 Performance Indicator 1b Dataset 1 Antenna port configuration 2, subband configuration 2 Training Method 1 Performance metrics 1c Dataset 1 Antenna port configuration 1, subband configuration 1, load configuration 1 Training Method 2 Performance index 2a Dataset 1 Antenna port configuration 1, subband configuration 3 Training Method 2 Performance Indicator 2b

[0306] In one possible example, antenna port configuration 1 in Table 2 indicates that the number of antenna ports for transmitting reference signals on the network side is 256, and the number of antenna ports for receiving reference signals on the terminal side is 32. That is, antenna port configuration 1 can be represented as 256T32R. In this case, performance index 1a, performance index 1c, or performance index 2 may include the following range for SGCS: SGCS ≥ 0.6. Antenna port configuration 2 in Table 2 indicates that the number of antenna ports for transmitting reference signals on the network side is 256, and the number of antenna ports for receiving reference signals on the terminal side is 4. That is, antenna port configuration 2 can be represented as 256T4R. In this case, performance index 1c may include the following range for SGCS: SGCS ≥ 0.8.

[0307] Optionally, different training methods for different models correspond to different performance metrics. Therefore, the network side can indicate the training method of the model to the terminal side through information (such as the second information shown below), thereby indicating the performance metrics to the terminal side.

[0308] Accordingly, prior to step 1020, method 1000 further includes 1030b: the terminal side receives second information, which is used to indicate the training method of the model training corresponding to the dataset. Accordingly, the network side sends the second information.

[0309] Similar to step 1010, in one example, the device performing model training on the terminal side is a terminal device. Step 1030b may include: the terminal device receiving second information from the network device, or the terminal device receiving second information from the OTT server. Information indicating the training method of model training (e.g., the second information itself, or other information) may be received by the OTT system server from the intelligent network element. In another example, the device performing model training on the terminal side is an OTT system server. Step 1030b may include: the OTT system server receiving second information from the intelligent network element, or the OTT system server receiving second information from the terminal device. Information indicating the training method of model training (e.g., the second information itself, or other information) may be received by the terminal device from the network side, such as the network device.

[0310] For example, the second information may include an identifier, number, or index of the training method to indicate the training method.

[0311] For example, as shown in Table 1 above, if the second information sent by the network side indicates training method 1, it is equivalent to the second information indicating performance index 1; or, if the second information sent by the network side indicates training method 2, it is equivalent to the second information indicating performance index 2; or, if the second information sent by the network side indicates training method 3, it is equivalent to the second information indicating performance index 3; or, if the second information sent by the network side indicates training method 4, it is equivalent to the second information indicating performance index 4.

[0312] It should be understood that method 1000 may include both steps 1030a and 1030b described above, or method 1000 may include any one of steps 1030a and 1030b described above.

[0313] Optionally, performance metrics can be predefined, such as protocol predefined metrics. For example, a protocol can predefine the dataset used for model training, along with the corresponding loss function and performance metrics. Or, for another example, a protocol can predefine the correspondence between various combinations of loss functions and performance metrics and the dataset.

[0314] Accordingly, before step 1020, method 1000 also includes 1030c, whereby the terminal side obtains the performance index from its own end.

[0315] Similar to step 1010, in one example, the device performing model training on the terminal side is a terminal device, and step 1030c may include: the terminal device obtaining the performance metric locally (corresponding to the case where the performance metric is stored on the terminal device), or the terminal device obtaining the performance metric from the OTT system server (corresponding to the case where the performance metric is stored on the OTT system server). In another example, the device performing model training on the terminal side is an OTT system server, and step 1030c may include: the OTT system server obtaining the performance metric locally (corresponding to the case where the performance metric is stored on the OTT system server), or the OTT system server obtaining the performance metric from the terminal device (corresponding to the case where the performance metric is stored on the terminal device).

[0316] For example, the correspondence between the predefined datasets and performance metrics in the protocol can be shown in Table 1 above.

[0317] The datasets available on the terminal side may be of various types, and the model training process will differ depending on the type of dataset. Since the following text will provide a detailed explanation of the model training process on the terminal side based on different types of datasets, it will not be elaborated upon here.

[0318] It should be noted that the model training process and performance metrics may vary depending on the dataset. Therefore, during model training, the terminal may determine whether the performance metric can be met while training the first model, or it may determine whether the performance metric can be met while training other models, which could be the models used to train the first model. In other words, the terminal may have already started training the first model before determining whether the performance metric can be met, or it may not have started training the first model yet.

[0319] During model training on the terminal side, this performance metric may or may not be met. The following sections will explain these two scenarios separately.

[0320] Optionally, the method further includes step 1040, whereby the terminal side performs the inference task of the first model if the performance indicators are met.

[0321] If the performance metrics are met, the terminal can execute the inference task of the first model. For example, the terminal can obtain the first model through model training. This first model can be deployed to a real-world scenario to perform the inference task if its performance in executing the inference task meets the performance metrics.

[0322] Furthermore, the method also includes step 1050: the terminal side sends fourth information to the network side, the fourth information being used to indicate that the performance indicators have been met. Accordingly, the network side receives the fourth information from the terminal side.

[0323] The network side can determine that the model training on the terminal side can meet the performance indicators based on the fourth information, and can continue to monitor the first model based on the performance indicators during the process of the model being deployed to the device that performs model inference, such as the terminal device, and then the device that performs model inference performs model inference.

[0324] Furthermore, the method also includes step 1060, in which the network side monitors the first model.

[0325] One possible approach is that the network side can continue to monitor the first model based on performance metrics.

[0326] For example, the performance metric is performance metric 2 as exemplified below, and the network side can continue to monitor the first model based on performance metric 2.

[0327] Another possible implementation is that the network side monitors the first model based on other metrics related to performance indicators.

[0328] For example, the performance metric is performance metric 2-1 as illustrated below. The network side can monitor the first model based on performance metric 2-1. Since performance metric 2-1 and performance metric 2-2 correspond to training 2-1 and training 2-2 respectively, and the first model has been deployed in a real-world scenario, the network side can monitor the first model based on performance metric 2-2.

[0329] The specific implementation of network-side monitoring of the first model is determined by the network side, and this application does not impose any restrictions on it.

[0330] Optionally, during the process of monitoring the first model based on performance metrics on the network side, if the performance metrics cannot be met, the network device may send a sixth message to the terminal side. The sixth message is used for one or more of the following: indicating to shut down model inference, indicating to perform model update or model retraining, indicating the performance metrics used for model update or model retraining, or indicating the dataset used for model update or model retraining.

[0331] Optionally, the method further includes step 1070, whereby the terminal sends third information to the network side if the performance indicators cannot be met. Accordingly, the network side receives the third information.

[0332] If the performance metric cannot be met, the terminal can report to the network side via third-party information.

[0333] For example, the third information is used for one or more of the following:

[0334] a. The indicated performance indicators cannot be met;

[0335] b. Request a change to the performance metrics;

[0336] c. Request a change of dataset; or

[0337] d. Request to shut down model training.

[0338] For example, if the terminal's computing power is limited, making it difficult to meet a particular performance metric, a third-party message can be used to request a change to the performance metric, such as a request to replace it with a less demanding one. The terminal can also simultaneously indicate that the aforementioned performance metric cannot be met via the third-party message.

[0339] For example, since the terminal can train the model based on the dataset sent by the network, it can request a change of dataset via third-party information, such as requesting a different type of dataset. The terminal can also simultaneously indicate via third-party information that the aforementioned performance metrics cannot be met.

[0340] For example, due to limited computing power on the terminal side, this performance metric may be difficult to meet. Therefore, the terminal side can request to disable model training via third-party information. The terminal side can also simultaneously indicate via third-party information that the aforementioned performance metric cannot be met. Disabling model training can be understood as the terminal side no longer training the model based on the dataset provided by the network side. In other words, the terminal side has not deployed the first model. The network side also no longer configures the terminal side for inference tasks using the first model. It should be understood that the network side's actions after disabling model training are at its own discretion, and this application does not impose any restrictions on this.

[0341] For example, the terminal can also indicate that the aforementioned performance metrics cannot be met via third-party information, without indicating other information (such as any of the items b to d listed above). The network's response after receiving the third-party information can be determined by the network itself, or it can be predefined by the protocol. For example, the protocol predefines shutting down model training when the performance metric is not met, or changing the performance metric when the performance metric is not met, and so on, to be listed further.

[0342] Similar to the steps described above, in one example, the device performing model training on the terminal side is a terminal device. The terminal device sends third information, which can be either sent to a network device or sent to an OTT system server. This information, through the OTT system server, is then transmitted to the intelligent network element, providing information that can be used to achieve one or more of the above-mentioned steps a to d (e.g., third information or other information). In another example, the device performing model training on the terminal side is an OTT system server. The terminal device sends third information, which can be either sent to a smart network element or sent to the terminal device. This information, through the terminal device, is then transmitted to the network device, providing information that can be used to achieve one or more of the above-mentioned steps a to d (e.g., third information or other information).

[0343] It should be understood that steps 1040-1060 and step 1070 are steps that the terminal side executes respectively when the performance indicators are met or not. The terminal side can choose to execute one of them according to the actual situation, and does not necessarily have to execute all of them.

[0344] Based on the above technical solution, the terminal can obtain a dataset for training the asymmetric two-end model and perform model training based on the performance metrics corresponding to the dataset, which also correspond to the inference task of the first model. In other words, after obtaining the dataset for training the asymmetric two-end model, the terminal can determine how to train the model and the performance metrics that should be met, and thus train the model accordingly. In this way, the terminal can obtain a model that meets the requirements through model training, thereby improving the feedback performance of CSI.

[0345] The training process will be explained in detail below with reference to the accompanying diagrams.

[0346] For ease of distinction and explanation, the dataset sent by the network side will be referred to as the first dataset in the following text, and the letters representing the various parameters of the first dataset will be identified by the subscript "NW". The letters representing the various parameters obtained by the terminal side through other means for training will be identified by the subscript "UE".

[0347] The first dataset includes one or more of the following: fourth data (which can be represented as H1) NW The fifth data (which can be represented as C2), the sixth data (which can be represented as H4) NW ) or the seventh data (which can be represented as H2) NW ).

[0348] The fourth data is used to characterize the fourth channel information. The fourth data can also be referred to as the first input CSI. The relationship between the fourth data and the fourth channel information can be found in the description of the relationship between the first data and the first channel information above.

[0349] The fifth data is obtained by performing a first process on the fourth data. This first process includes compression, or, alternatively, compression and quantization. The fifth data is used to determine the eighth data. The fifth data can also be referred to as the first CSI feedback information.

[0350] The sixth data point is the true value corresponding to the eighth data point. The sixth data point can also be called the target CSI or the true CSI.

[0351] The seventh data is used to characterize the sixth channel information. The seventh data, along with the fifth data, is used to determine the eighth data. The seventh data can be determined by the network side through measurements of uplink reference signals (such as sounding reference signals, SRS). The seventh data can also be referred to as uplink reference signal measurements, such as SRS measurements. The relationship between the seventh data and the sixth channel information can be found in the description of the relationship between the first data and the first channel information above.

[0352] The eighth data is used to characterize the fifth channel information. The fifth channel information includes at least one of the following: part or all of the fourth channel information, part or all of the sixth channel information, or the seventh channel information. The seventh channel information differs from both the fourth and sixth channel information. The eighth data can also be referred to as reconstructed CSI. The relationship between the eighth data and the fifth channel information can be found in the description of the relationship between the third data and the second channel information above.

[0353] The data used for training on the terminal side includes the second input CSI (which can be represented as H). UEThe second input CSI can be obtained through measurement by the terminal device, simulation by the terminal side, or predefined by the protocol, etc., without limitation. The CSI obtained by the terminal side itself can include a CSI with the same format as the first input CSI (or a CSI corresponding to the same channel as the first input CSI), or a CSI with the same format as the target CSI (or a CSI corresponding to the same channel as the target CSI). For example, the terminal side obtains a CSI with the same format as the first input CSI by measuring a first reference signal from the network side. Another example is that the terminal side obtains a CSI with the same format as the target CSI by measuring a second reference signal from the network side. Yet another example is that the terminal side obtains a CSI with the same format as the first input CSI or a CSI with the same format as the target CSI through simulation.

[0354] In addition, for ease of distinction, the models involved in the training process will be briefly explained below.

[0355] First Model: The model that the terminal needs to obtain through model training. This first model can be used to perform inference tasks; more specifically, it can be used to perform inference on the compression (or compression and quantization) of the input CSI (such as denoted as the first data), or in other words, to compress (or compress and quantize) the input CSI. As an example, the first model can be an encoder.

[0356] The second model is a network-side model that can work in conjunction with the first model. The second model can be used to process CSI feedback information (e.g., decompress, or decompress and dequantize) it. Alternatively, the second model can be used to process information obtained by dequantizing the CSI feedback information (e.g., decompress). As an example, the second model could be a decoder.

[0357] The third model: The third model is a model obtained by training the terminal model and can work in conjunction with the first model. The third model can be used to process CSI feedback information (such as decompressing, or decompressing and dequantizing) or to process information obtained by dequantizing CSI feedback information (such as decompressing). As an example, the third model can be a decoder.

[0358] The fourth model includes the first and third models, and possesses the functions of both. For example, the fourth model can be used to compress the input CSI to obtain compressed CSI (i.e., the function of the first model); the fourth model can also be used to decompress the compressed CSI to obtain reconstructed CSI (i.e., the function of the third model). As another example, the fourth model can be used to compress and quantize the input CSI to obtain quantized CSI (i.e., the function of the first model); the fourth model can also be used to dequantize and decompress the quantized CSI to obtain reconstructed CSI (i.e., the function of the third model). As an example, the fourth model is an AE model that includes an encoder and a decoder.

[0359] Example 1: The network side sends the first dataset, which includes the fourth data (which can be represented as H1). NW The first dataset contains two data points: the first input CSI and the fifth data point (which can be represented as C2). The terminal side trains the model based on the first dataset. The fifth data point can be either compressed CSI (i.e., unquantized CSI) or quantized CSI, without limitation. For ease of understanding, the fourth data point will be referred to as the first input CSI, and the fifth data point as the first CSI feedback information.

[0360] For example, before the network side sends the first dataset, it can first perform joint training of encoder 1 and decoder 1 (an example of the second model) to obtain the first dataset.

[0361] For example, the process of joint training on the network side is as follows: Figure 11 As shown. The network side will input the first input CSI (which can be represented as H1). NW The first input CSI is used as the input to encoder 1. Encoder 1 can compress the first input CSI to obtain first CSI feedback information (which can be represented as C2), and this first CSI feedback information can be used as the input to decoder 1. Decoder 1 can decompress the first CSI feedback information to obtain the reconstructed CSI (which can be represented as H3). NW Furthermore, the network side can target the CSI (which can be represented by H4). NW Using the labels, encoder 1 and decoder 1 are trained. By training encoder 1 and decoder 1, the reconstructed CSI output by decoder 1 is made to approach the target CSI.

[0362] Optionally, encoder 1 can compress and quantize the first input CSI to obtain first CSI feedback information. Correspondingly, decoder 1 can decompress and dequantize the first CSI feedback information to obtain reconstructed CSI.

[0363] Optionally, encoder 1 can compress the first input CSI to obtain compressed CSI, and the compressed CSI is quantized to obtain first CSI feedback information. Correspondingly, decoder 1 can decompress the dequantized information of the first CSI feedback information to obtain reconstructed CSI. The dequantized information of the first CSI feedback information corresponds to the compressed CSI.

[0364] Optionally, such as Figure 11 As shown, the input to decoder 1 can also include SRS measurements (which can be represented as H2). NW In other words, decoder 1 is used to process the second CSI feedback information and the second SRS measurement.

[0365] For example, further descriptions of the first input CSI, the first CSI feedback information, the SRS measurement, and the reconstructed CSI can be found in the descriptions of the fourth, fifth, seventh, and eighth data above.

[0366] For example, after the network side completes the joint training of encoder 1 and decoder 1, it sends the first dataset containing the first input CSI and the first CSI feedback information to the terminal side.

[0367] Based on the first dataset, the terminal side can use the following training methods to train the model.

[0368] Training Method 1: The terminal side trains the first model (such as encoder 2) based on the first dataset, or in other words, encoder 2 is trained separately. In other words, in training method 1, the first dataset (or replaced by the first input CSI and the first CSI feedback information) is used for model training of the first model.

[0369] In the training of the first model, the first input CSI is used as the input of the first model, and the first CSI feedback information is used as the true value corresponding to the output of the first model.

[0370] In one possible design (hereinafter referred to as Design 1), the encoder 2 to be trained on the terminal side can be used for compression and quantization, or in other words, encoder 2 has the function of compression and quantization. The first CSI feedback information in the first dataset can be quantized CSI. In this case, the inference task of the first model can also be said to be used to compress and quantize the first input CSI.

[0371] In another possible design (hereinafter referred to as Design 2), the encoder 2 to be trained on the terminal side can be used for compression but not for quantization; in other words, encoder 2 has compression functionality but not quantization functionality. The first CSI feedback information in the first dataset can be either quantized CSI or compressed CSI.

[0372] It should be understood that whether an encoder has quantization functionality (or whether it includes a quantizer) can be predefined, such as by protocol predefinition, or indicated by the network side, or determined by the terminal side itself. This application does not limit this.

[0373] The following sections will explain the two possible designs mentioned above.

[0374] In Design 1, the encoder can be used for compression and quantization, and the first CSI feedback information in the first dataset can be the quantized CSI.

[0375] For the process of model training on the terminal side, please refer to [link / reference]. Figure 12 Specifically, as follows: The first input CSI (i.e., H1) from the first dataset... NW The first CSI feedback information (i.e., C2) from the first dataset is used as the input to encoder 2, and the first CSI feedback information (i.e., C2) is used as the label to train encoder 2. Encoder 2 can be used to compress and quantize the first input CSI, and output the CSI feedback information (i.e., C2). It should be understood that in Design 1, the CSI feedback information output by encoder 2 is quantized CSI. Through training encoder 2, the CSI feedback information output by encoder 2 (i.e., The training approaches the label (i.e., C2). For example, this training can be achieved through supervised learning targeting C2.

[0376] It should be understood that at the output of encoder 2 When the performance metrics for label C2 are met, training of encoder 2 on the terminal side can be stopped, thus obtaining the trained encoder 2. The terminal side can then deploy the trained encoder 2 to a specific device to perform model inference.

[0377] In Design 2, the encoder can be used for compression but not for quantization. The first CSI feedback information in the first dataset can be either compressed CSI or quantized CSI. The different first CSI feedback information leads to different model training procedures.

[0378] If the first CSI feedback information is compressed CSI, then the process of model training on the terminal side is combined with the above. Figure 12 The process is similar, except that encoder 2 is used to compress the first input CSI, so the CSI feedback information output by encoder 2 is the compressed CSI. Other processes can be found in the above text. Figure 12 The relevant descriptions will not be repeated here.

[0379] If the first CSI feedback information is a quantized CSI, the process of model training on the terminal side can include the following steps.

[0380] Training 1-0: Dequantize the first CSI feedback information in the first dataset to obtain a compressed CSI (denoted as C2') that includes quantization loss.

[0381] Training 1-1: Take the first input CSI (i.e., H1) from the first dataset. NW Using the first input CSI (i.e., C2'), which is the compressed CSI containing quantization loss obtained from training 1-0, as the input to encoder 2, encoder 2 is trained using the label. Encoder 2 can then be used to train encoder 2 on the first input CSI (i.e., H1). NW The encoder 2 is compressed, and the compressed CSI is output. Through training the encoder 2, the compressed CSI output by the encoder 2 (e.g., denoted as...) is improved. The CSI and the label (i.e., C2') tend to be close. For example, this training can be achieved by supervised learning of the compressed CSI (i.e., C2') with quantization loss, which is obtained by dequantizing the first CSI feedback information in the first dataset, and therefore can also be called supervised learning of the first CSI feedback information.

[0382] During the training processes 1-0 and 1-1 described above, the dequantization operation can be implemented by modules other than the encoder and decoder. For example, dequantization can be implemented by a dequantizer. The dequantizer can be predefined, such as protocol predefined, or it can be indicated by the network side; this application does not limit this. Optionally, the first dataset also includes a dequantizer. Since dequantization and quantization are relative, another possible implementation of the predefined dequantizer is a predefined quantizer, and another possible implementation of the network-indicated quantizer and dequantizer is a network-indicated quantizer.

[0383] For example, corresponding to training method one, during model training, the set of functions used to monitor model performance may include one or more of the following functions: NMSE between the output of encoder 2 (i.e., an example of the first model) and the label, MSE between the output of encoder 2 and the label, MAE between the output of encoder 2 and the label, GCS between the output of encoder 2 and the label, SGCS between the output of encoder 2 and the label, or a weighted sum or weighted average of two or more of the above. The weighting coefficients for each item may be predefined or indicated by the network side; this application does not limit this.

[0384] The performance metric corresponding to the aforementioned function set can be the range that the values ​​of each function in the function set should satisfy. That is, the performance metric can include the range that the values ​​of one or more of the following functions should satisfy: NMSE between the output of encoder 2 and the tag, MSE between the output of encoder 2 and the tag, MAE between the output of encoder 2 and the tag, GCS between the output of encoder 2 and the tag, SGCS between the output of encoder 2 and the tag, or a weighted sum or weighted average of two or more of the above. The weighting coefficients for each item can be predefined or indicated by the network side; this application does not limit this.

[0385] For a more detailed description of the range that the values ​​of NMSE, MSE, MAE, GCS, or SGCS should satisfy, please refer to the description of step 1020 above. For example, the NMSE / SGCS between the output of encoder 2 and the tag may include the average value and / or distribution of the NMSE / SGCS between the output of encoder 2 and the tag.

[0386] As can be seen from the above, the performance of this model can be monitored by monitoring the difference between the output of encoder 2 and the label.

[0387] As seen in Design 1 and Design 2 above, the output of encoder 2 is not necessarily the same, and the labels used for training are also different. In Design 1, the label is the first CSI feedback information C2, where C2 represents the quantized CSI. The CSI feedback information output by encoder 2 is different. This represents quantized CSI; in Design 2, if the first CSI feedback information is compressed CSI, then the tag is the first CSI feedback information C2, where C2 represents compressed CSI, and the CSI feedback information output by encoder 2. The compressed CSI is indicated. The set of functions used to monitor model performance (referred to as function set 1 for ease of distinction and explanation) may include one or more of the following functions.

[0388] The output of encoder 2 (i.e., The MSE between the label (i.e., C2) and the tag (i.e., C2) (an example of the first MSE), i.e.

[0389] The output of encoder 2 (i.e., The NMSE (example of the first NSME) between the label (i.e., C2) and the tag (i.e., C2).

[0390] The output of encoder 2 (i.e., The MAE between the label (i.e., C2) and the tag (i.e., C2) (an example of the first MAE), i.e.

[0391] The output of encoder 2 (i.e., The GCS between the label (i.e., C2) and the tag (i.e., C2)

[0392] The output of encoder 2 (i.e., The SGCS between the label (i.e., C2) and the tag (i.e., C2) is... or

[0393] The output of encoder 2 (i.e., The weighted sum or weighted average of two or more of the following: NMSE, MSE, MAE, GCS, or SGCS, between the label (i.e., C2) and the tag (i.e., C2). The weighting coefficients of each item may be predefined or indicated by the network side, and this application does not limit them.

[0394] The output of encoder 2 is the first input CSI (i.e., H1) processed by encoder 2. NW The CSI feedback information obtained by performing the first processing (i.e., compression, or compression and quantization) is as follows: The label is the first CSI feedback information (i.e., C2).

[0395] The performance metric corresponding to the above function set 1 can be the range that the values ​​of each function in function set 1 should satisfy. That is, the performance metric can include the range that the values ​​of one or more of the following functions should satisfy: Or a weighted sum or weighted average of two or more of the above terms. For ease of distinction and explanation, the performance index corresponding to this function set 1 will be referred to as performance index 1 below.

[0396] In Design 2, the first CSI feedback information is a quantized CSI, and the label is a compressed CSI (i.e., C2') obtained by dequantizing the first CSI feedback information and including quantization loss. The output of the second model... This is a compressed CSI; the set of functions used to monitor model performance can be found in the above text. The descriptions of performance metrics and indicators in C2 are only a part of it. It was replaced with C2 has been replaced with C2', which will not be elaborated further.

[0397] As can be seen, the above performance metrics are directly or indirectly related to the inference task of the first model. These metrics can be used to monitor decoder-only training, end-to-end training, and both decoder-only and end-to-end training. Therefore, it can be said that the above performance metrics correspond to the inference task of the first model.

[0398] We can also see that among the performance metric functions included in function set 1, and The loss function, in this example, is the desired output after encoder 2 processing. It tends to approach the input C2, therefore when the performance metric function includes Alternatively, when using a weighted sum or weighted average of two or more of the aforementioned terms, the performance metric used for model training can be the upper bound of the values ​​of each performance metric function. In other words, if the value of a performance metric function is less than its corresponding upper bound, the performance metric is satisfied; if the value of a performance metric function is greater than its corresponding upper bound, the performance metric is not satisfied.

[0399] Alternatively, their weighted sum or weighted average can be used to describe the degree of similarity between them. In this example, since we want the output of encoder 2 to be... It tends to approach the input C2, therefore when the performance metric function includes Alternatively, when using a weighted sum or weighted average, the performance metric used for training the model can be the lower bound of the values ​​of each performance metric function. In other words, if the value of a performance metric function is greater than its corresponding lower bound, the performance metric is satisfied; if the value of a performance metric function is less than its corresponding lower bound, the performance metric is not satisfied.

[0400] For example, the performance metric function is Performance indicators are The upper bound is then The value is less than the upper bound, thus satisfying the performance index; The value is greater than the upper bound, and the performance index is not met.

[0401] For example, the performance metric function is and The performance indicators are as follows: upper bound and The lower bound. Then The value is less than this The upper bound, and The value is greater than this The lower bound of satisfies the performance indicators; The value is greater than this The upper bound, and / or, The value is less than this The lower bound does not meet the performance specifications.

[0402] For example, the performance metric function is and The weighted sum or weighted average is given by the metric, which is the lower bound of the weighted sum or weighted average. and If the weighted sum or weighted average is greater than the lower bound, the performance index is satisfied. and If the weighted sum or weighted average is less than the lower bound, the performance index is not met.

[0403] It should be noted that whether the value of the performance metric function equals the upper or lower bound and is determined to meet or not meet the performance metric can be predefined, such as protocol predefined or pre-negotiated between the network side and the terminal side. This application does not limit this. Furthermore, the upper and lower bounds exemplified above are provided for ease of understanding of the performance metric; this application does not limit the specific form of the performance metric, nor does it specify the specific values ​​of the upper and lower bounds. For example, the performance metric can be determined based on business requirements, or it can be determined in conjunction with the computing power of the terminal side, etc., without limitation.

[0404] It should also be noted that if the above GCS and SGCS are replaced with the difference between GCS and 1, and the difference between SGCS and 1, respectively, the performance index can also be the upper bound of the corresponding performance measurement function. The substitution of GCS and SGCS and the corresponding changes in the performance index here are applicable to the performance indices of GCS and SGCS discussed later; for the sake of brevity, they will not be repeated here.

[0405] Training Method 2: The terminal first trains a third model (e.g., decoder 2) based on the first dataset, and then trains a fourth model (e.g., a combination of encoder 2 and decoder 2) based on the third model. The fourth model includes the first model (e.g., encoder 2) and the third model. In other words, in training method 2, the first dataset (or replaced by the first input CSI and the first CSI feedback information) is used to train the third model, and the third model is used to train the fourth model to obtain the first model.

[0406] In the training of the third model, the first CSI feedback information is used as the input of the third model, and the first input CSI is used as the true value corresponding to the output of the third model.

[0407] In the training of the fourth model, data #1 is used as the input to the first model (or replaced with the input to the fourth model) and the ground truth value corresponding to the output of the third model (or replaced with the ground truth value corresponding to the output of the fourth model). The output of the first model is used as the input to the first model. For example, data #1 includes the first input CSI and / or data other than the first input CSI, such as the CSI acquired by the terminal itself. The channel information represented by data #1 corresponds to the same channel as the fourth channel information. Hereinafter, data #1 will be referred to as the second input CSI.

[0408] Referring to the description in Training Method 1 above, there are two possible designs for the encoder on the terminal side. Design 1: The encoder 2 to be trained on the terminal side can be used for both compression and quantization. Design 2: The encoder 2 to be trained on the terminal side can be used for compression but not for quantization.

[0409] The following sections will explain these two possible designs.

[0410] In Design 1, the encoder can be used for compression and quantization, and correspondingly, the decoder can be used for dequantization and decompression. The first CSI feedback information in the first dataset can be the quantized CSI.

[0411] For the process of model training on the terminal side, please refer to [link / reference]. Figure 13 This may include the following steps.

[0412] Training 2-1: Train decoder 2 separately based on the first dataset. This decoder 2 can be a nominal decoder. On the terminal side, the first CSI feedback information (i.e., C2) from the first dataset is used as the input to decoder 2, and the first input CSI (i.e., H1) from the first dataset is used as the input to decoder 2. NW The first CSI feedback information (i.e., C2) is used as a label to train decoder 2. Decoder 2 can be used to dequantize and decompress the first CSI feedback information (i.e., C2) to output the reconstructed CSI (i.e., C2). By training decoder 2, the output of decoder 2 (i.e., ) and tags (i.e., H1) NW They tend to approach each other. For example, training 2-1 can be done by targeting H1. NW This is achieved through supervised learning.

[0413] It should be understood that at the output of decoder 2 With label H1 NW When the performance metrics are met, the training of the decoder 2 model on the terminal side can be stopped, thus obtaining the trained decoder 2.

[0414] Training 2-2: Train encoder 2 based on decoder 2. Encoder 2 can be a model built by the terminal itself, for example, the structure of which is designed by the terminal. The terminal can fix the parameters of decoder 2 (i.e., use a frozen decoder 2; in other words, do not change the parameters of decoder 2 during the training of encoder 2), and input the second input CSI (i.e., H... UE The second input CSI used in training 2-2 can be the input to encoder 2, used for end-to-end training of encoder 2 and decoder 2. It should be understood that the second input CSI used in training 2-2 can be the first input CSI from the first dataset on the network side (i.e., H). UE It can also be H1 NW It can be CSI obtained by the terminal itself, or it can be CSI obtained by the terminal itself. This application does not limit this.

[0415] For example, the end-to-end training process is as follows: The second input CSI (i.e., H) is... UE The second input CSI is input to encoder 2, which can be used to compress and quantize the CSI, outputting CSI feedback information (i.e., quantized CSI). The output of encoder 2 is used as input to decoder 2, which can be used to dequantize and decompress the CSI feedback information output by encoder 2, outputting reconstructed CSI (i.e., quantized CSI). End-to-end training of encoder 2 and decoder 2 can also be referred to as joint training of encoder 2 and decoder 2. The label for this joint training is the second input CSI (i.e., H). UE Through end-to-end training, the output of decoder 2 (i.e., ) and tags (i.e., H UE They tend to approach each other. For example, training 2-2 can be performed by targeting H. UE This is achieved through self-supervised learning.

[0416] It should be understood that at the output of decoder 2 H with input encoder 2 UE Once the performance metrics are met, training of encoder 2 on the terminal side can be stopped, thus obtaining the trained encoder 2. The terminal side can then deploy the trained encoder 2 to a specific device to perform model inference.

[0417] For example, corresponding to training method two, during model training, the set of functions used to monitor model performance (referred to as function set 2 for ease of distinction and explanation) may include one or more of the following functions: NMSE between the output of decoder 2 (i.e., an example of the third model) and the label, MSE between the output of decoder 2 and the label, MAE between the output of decoder 2 and the label, GCS between the output of decoder 2 and the label, SGCS between the output of decoder 2 and the label, or a weighted sum or weighted average of two or more of the above. The weighting coefficients of each item may be predefined or indicated by the network side; this application does not limit this.

[0418] For a more detailed description of the range that the values ​​of NMSE, MSE, MAE, GCS, or SGCS should satisfy, please refer to the description of step 1020 above. For example, the NMSE / SGCS between the output of decoder 2 and the tag may include the average value and / or distribution of the NMSE / SGCS between the output of decoder 2 and the tag.

[0419] The performance metric corresponding to the aforementioned function set 2 can be the range that the values ​​of each function in function set 2 should satisfy. That is, the performance metric can include the range that the values ​​of one or more of the following functions should satisfy: NMSE between the output of decoder 2 and the label, MSE between the output of decoder 2 and the label, MAE between the output of decoder 2 and the label, GCS between the output of decoder 2 and the label, SGCS between the output of decoder 2 and the label, or a weighted sum or weighted average of two or more of the above. The weighting coefficients for each item can be predefined or indicated by the network side; this application does not limit this.

[0420] As can be seen from Training 2-1 and Training 2-2 above, the output of Decoder 2 is not necessarily the same. Therefore, corresponding to Training 2-1, the set of functions used to monitor model performance (denoted as Function Set 2-1 for ease of distinction and explanation) may include one or more of the following functions.

[0421] The output of decoder 2 (i.e., ) and tags (i.e., H1) NW The NMSE between (example of the second NMSE), i.e.

[0422] The output of decoder 2 (i.e., ) and tags (i.e., H1) NW The MSE between (example of the second MSE), i.e.

[0423] The output of decoder 2 (i.e., ) and tags (i.e., H1) NW MAE between (example of the second MAE), i.e.

[0424] The output of decoder 2 (i.e., ) and tags (i.e., H1) NW GCS between )

[0425] The output of decoder 2 (i.e., ) and tags (i.e., H1) NW SGCS between ) or,

[0426] The output of decoder 2 (i.e., ) and tags (i.e., H1) NW The weighted sum or weighted average of two or more of the following: NMSE, MSE, MAE, GCS, or SGCS. The weighting coefficients for each item can be predefined or indicated by the network side, and this application does not limit them.

[0427] The output of decoder 2 is the reconstructed CSI obtained by performing a second process (i.e., decompression, or decompression and dequantization) on the first CSI feedback information through decoder 2. The label is the first input CSI (i.e., H1). NW ).

[0428] The performance metric corresponding to the function set 2-1 above can be the range that the values ​​of each function in function set 2-1 should satisfy. That is, the performance metric can include the range that the values ​​of one or more of the following functions should satisfy: Or a weighted sum or weighted average of two or more of the above terms. For ease of distinction and explanation, the performance index corresponding to function set 2-1 will be referred to as performance index 2-1 below.

[0429] Corresponding to Training 2-2, the set of functions used to monitor model performance (referred to as Function Set 2-2 for ease of distinction and explanation) may include one or more of the following functions.

[0430] The output of decoder 2 (i.e., ) and tags (i.e., H UE The NMSE between (example of the third NMSE), i.e.

[0431] The output of decoder 2 (i.e., ) and tags (i.e., H UE The MSE between (example of the third MSE), i.e.

[0432] The output of decoder 2 (i.e., ) and tags (i.e., H UE MAE between (example of the third MAE), i.e.

[0433] The output of decoder 2 (i.e., ) and tags (i.e., H UE The GCS between (example of the third GCS), i.e.

[0434] The output of decoder 2 (i.e., ) and tags (i.e., H UE ) between SGCS (example of the third SGCS), i.e. or,

[0435] The output of decoder 2 (i.e., ) and tags (i.e., H UEThe weighted sum or weighted average of two or more of the following: NMSE, MSE, MAE, GCS, or SGCS. The weighting coefficients for each item can be predefined or indicated by the network side, and this application does not limit them.

[0436] The output of decoder 2 is the result of the second input CSI (i.e., H) processed by encoder 2 and decoder 2 (the combination of encoder 2 and decoder 2 is an example of the fourth model). UE The reconstructed CSI obtained by processing ) is, ), labeled as the second input CSI (i.e., H UE ).

[0437] The performance metric corresponding to the function set 2-2 above can be the range that the values ​​of each function in function set 2-2 should satisfy. That is, the performance metric can include the range that the values ​​of one or more of the following functions should satisfy: Or a weighted sum or weighted average of two or more of the above terms. For ease of distinction and explanation, the performance index corresponding to function set 2-2 will be referred to as performance index 2-2 below.

[0438] It should be understood that the functions listed above are merely examples and should not be construed as limiting this application. Those skilled in the art can perform simple mathematical transformations based on the functions listed above to obtain more functions; they can also use other functions to replace the functions listed above, based on the same concept. For example, Alternatively, they can be replaced with: 1 and The absolute value of the difference (i.e., ), 1 and The absolute value of the difference (i.e., For example, Alternatively, they can be replaced with: 1 and The absolute value of the difference (i.e., ), 1 and The absolute value of the difference (i.e., This application includes, but is not limited to, functions such as, etc. The functions mentioned below can also be extended and modified with reference to the descriptions and examples here, and will not be elaborated on further where functions are mentioned below.

[0439] As can be seen, the function sets corresponding to training method two include function set 2-1 and function set 2-2, and the corresponding performance metrics include performance metric 2-1 and performance metric 2-2. Function set 2-1 and performance metric 2-1 correspond to training 2-1, and function set 2-2 and performance metric 2-2 correspond to training 2-2. Therefore, the performance metric corresponding to the first dataset can be {performance metric 2-1}, or {performance metric 2-2}, or {performance metric 2-1, performance metric 2-2}, and the function set corresponding to the first dataset can be {function set 2-1}, or {function set 2-2}, or {function set 2-1, function set 2-2}.

[0440] For example, if the terminal side wants to monitor training 2-1, it can train the model based on function set 2-1 and performance metric 2-1. If the terminal side wants to monitor training 2-2, it can train the model based on function set 2-2 and performance metric 2-2. If the terminal side wants to monitor the entire training process, it can train the model for training 2-1 based on function set 2-1 and performance metric 2-1, and train the model for training 2-2 based on function set 2-2 and performance metric 2-2.

[0441] The situations where performance metrics are met and not met have been explained in detail with multiple examples in Training Method 1 above, which can be found in the relevant description above, and will not be repeated here.

[0442] In Design 2, the encoder can be used for compression, not quantization. Correspondingly, the decoder can be used for decompression, not dequantization. The first CSI feedback information can be either compressed CSI or quantized CSI. The corresponding processes differ depending on the first CSI feedback information.

[0443] If the first CSI feedback information is compressed CSI, the process of model training on the terminal side is the same as described above. Figure 13 The illustrated process is similar, except that decoder 2 is used to decompress the CSI feedback information, so the CSI feedback information input to decoder 2 is compressed CSI; encoder 2 is used to compress the input CSI, so the CSI feedback information output by encoder 2 is compressed CSI. Other processes can be found in the above text. Figure 13 The relevant descriptions will not be repeated here.

[0444] If the first CSI feedback information is a quantized CSI, the process of model training on the terminal side can include the following steps.

[0445] Training 2-0: Dequantize the first CSI feedback information (i.e., C2) in the first dataset to obtain a compressed CSI that includes quantization loss.

[0446] Training 2-1': Similar to the process of Training 2-1 above, the difference is that the input of Decoder 2 in Training 2-1 is the first CSI feedback information (i.e., C2) in the first dataset, while the input of Decoder 2 in Training 2-1' is the compressed CSI containing quantization loss.

[0447] Training 2-2': Train encoder 2 based on decoder 2. For example, the terminal side can fix the parameters of decoder 2 and train the second input CSI (i.e., H...). UE The second input CSI (i.e., H) is used as the input to encoder 2 and the training labels to perform end-to-end training on encoder 2 and decoder 2. End-to-end training can reuse the procedure provided in training 2-2 above; in this case, training 2-2' is the same as training 2-2. Alternatively, the end-to-end training procedure can differ from that provided in training 2-2: the second input CSI (i.e., H) is used as the input to encoder 2 and the training labels to perform end-to-end training on encoder 2 and decoder 2. UE Input encoder 2, encoder 2 can be used to input the second input CSI (i.e., H UE The CSI is compressed to obtain compressed CSI; the compressed CSI is quantized to obtain quantized CSI; the quantized CSI is dequantized to obtain compressed CSI including quantization loss; the compressed CSI including quantization loss is used as input to decoder 2, which can be used to decompress the compressed CSI including quantization loss to obtain reconstructed CSI (i.e., Through end-to-end training, the output of decoder 2 (i.e., ) and tags (i.e., H UE They tend to approach each other; for example, training 2-2' can be achieved by targeting H. UE This is achieved through self-supervised learning.

[0448] Similar to Training 2-2, the second input CSI used in Training 2-2' can be the first input CSI from the first dataset on the network side (i.e., H). UE It can also be H1 NW It can be CSI obtained by the terminal itself, or it can be CSI obtained by the terminal itself. This application does not limit this.

[0449] In the training process provided in Design 2, quantization and dequantization operations can be implemented by modules other than the encoder and decoder. For example, quantization can be implemented by a quantizer, and dequantization can be implemented by a dequantizer. The quantizer and / or dequantizer can be predefined, such as protocol predefined, or indicated by the network side; this application does not limit this. Optionally, the first dataset also includes a quantizer and / or a dequantizer. Since dequantization and quantization are relative, one possible implementation of predefined quantizers and dequantizers is to predefine either a quantizer or a dequantizer, and another possible implementation of network-indicated quantizers and dequantizers is to indicate either a quantizer or a dequantizer. The relevant descriptions of quantization and dequantization involved in the training process described below can be found here, and will not be repeated hereafter.

[0450] For a more detailed description of the set of functions and performance metrics used to monitor model performance for Training 2-1' and Training 2-2', please refer to the description of the set of functions and performance metrics used to monitor model performance for Training 2-1 and Training 2-2 above.

[0451] As mentioned above, when the first dataset includes the first input CSI and the first CSI feedback information, the terminal can train the model using either training method one or training method two. Correspondingly, training method one and training method two correspond to different performance metrics. Therefore, in order for the terminal to determine the performance metrics used for model training (or in other words, to determine the training method used for model training), the network can indicate the performance metrics and / or training method to the terminal device through information (such as first information and / or second information).

[0452] For example, the network side sends first information to the terminal side, and the first information indicates that the performance indicator is the performance indicator 1 corresponding to the above training method 1, or indicates that the performance indicator is the above performance indicator 2-1 and / or performance indicator 2-2.

[0453] For example, the network sends a second message to the terminal device, indicating that the training method is either training method one or training method two. In other words, the second message indicates that the first dataset is used for training the first model (i.e., indicating that the training method is training method one), or the second message indicates that the first dataset is used for training the third model (i.e., indicating that the training method is training method two).

[0454] Example 2: The network side sends the first dataset, which includes the fourth data (which can be represented as H1). NW The fifth data (which can be represented as C2) and the sixth data (which can be represented as H4) NWThe terminal side trains the model based on the first dataset. The fifth data point can be either compressed CSI (i.e., unquantized CSI) or quantized CSI, without limitation. For ease of understanding, the fourth data point will be denoted as the first input CSI, the fifth data point as the first CSI feedback information, and the sixth data point as the target CSI.

[0455] For example, before sending the first dataset, the network can first perform joint training of encoder 1 and decoder 1 (an example of the second model) to obtain the first dataset. The method for the network to perform joint training of encoder 1 and decoder 1 can be referred to the description in Example 1 above.

[0456] For example, after the network side performs joint training of encoder 1 and decoder 1, it sends a first dataset containing the first input CSI, the first CSI feedback information and the target CSI to the terminal side.

[0457] Based on the first dataset, the terminal side can use the following training methods to train the model.

[0458] Training Method 3: The terminal first trains a third model (e.g., decoder 3) based on the first dataset, and then trains a fourth model (e.g., a combination of encoder 2 and decoder 3) based on the third model. The fourth model includes the first model (e.g., encoder 2) and the third model. In other words, in training method 3, the first CSI feedback information and target CSI included in the first dataset are used to train the third model, and the third model is used to train the fourth model to obtain the first model.

[0459] In the training of the third model, the first CSI feedback information is used as the input of the third model, and the target CSI is used as the ground truth value corresponding to the output of the third model.

[0460] In the training of the fourth model, the ninth data is used as the input to the first model (or replaced with the input to the fourth model), the tenth data is used as the ground truth corresponding to the output of the third model (or replaced with the ground truth corresponding to the output of the fourth model), and the output of the first model is used as the input to the third model.

[0461] Optionally, the ninth data includes one or more of the following: the first input CSI, the target CSI, or CSI other than the first input CSI and the target CSI. The tenth data includes one or more of the following: the first input CSI, the target CSI, or CSI other than the first input CSI and the target CSI. The CSI other than the first input CSI and the target CSI can be the terminal's own CSI, such as CSI obtained through measurement or simulation.

[0462] For example, the ninth data point is the first input CSI, and the tenth data point is either the first input CSI or the target CSI. Alternatively, the ninth data point may be the target CSI, and the tenth data point may also be the target CSI. The ninth data point will be referred to as the second input CSI below.

[0463] Referring to the description in Training Method 1 above, there are two possible designs for the encoder on the terminal side. Design 1: The encoder 2 to be trained on the terminal side can be used for both compression and quantization. Design 2: The encoder 2 to be trained on the terminal side can be used for compression but not for quantization.

[0464] The following sections will explain these two possible designs.

[0465] In Design 1, the encoder can be used for compression and quantization, and correspondingly, the decoder can be used for dequantization and decompression. The first CSI feedback information in the first dataset can be the quantized CSI.

[0466] For the process of model training on the terminal side, please refer to [link / reference]. Figure 14 This may include the following steps.

[0467] Training 3-1: Train decoder 3 independently based on the first dataset. Decoder 3 can be a nominal decoder. On the terminal side, the first CSI feedback information (i.e., C2) from the first dataset is used as the input to decoder 3, and the target CSI (i.e., H4) from the first dataset is used as the input. NW The first CSI feedback information (i.e., C2) is used as a label to train decoder 3. Decoder 3 can be used to dequantize and decompress the first CSI feedback information (i.e., C2) to output the reconstructed CSI (i.e., C2). By training decoder 3, the output of decoder 3 (i.e., ) and tags (i.e., H4) NW They tend to approach each other. For example, training 3-1 can be done by targeting H4. NW This is achieved through supervised learning.

[0468] It should be understood that in the output of decoder 3 With tag H4 NW When the performance metrics are met, the training of the decoder 3 model on the terminal side can be stopped, thus obtaining the trained decoder 3.

[0469] Training 3-2: Train encoder 2 based on decoder 3. Encoder 2 can be a model built by the terminal itself, for example, the model's structure is designed by the terminal. The terminal can fix the parameters of decoder 3 (i.e., use a frozen decoder 3; in other words, do not change the parameters of decoder 3 during encoder 2 training), and input the second input CSI (i.e., H... UEThe second input CSI used as input to encoder 2 is used to train encoder 2 and decoder 3 end-to-end. It should be understood that the second input CSI used in training 3-2 can be the first input CSI from the first dataset on the network side (i.e., H...). UE It can also be H1 NW ) or target CSI (i.e., H UE It can also be H4 NW It can be CSI obtained by the terminal itself, or it can be CSI obtained by the terminal itself. This application does not limit this.

[0470] For example, the end-to-end training process is as follows: The second input CSI (i.e., H) is... UE The second input CSI is input to encoder 2, which can be used to compress and quantize the input CSI, outputting CSI feedback information (i.e., quantized CSI). The output of encoder 2 is used as input to decoder 3, which can be used to dequantize and decompress the CSI feedback information output by encoder 2, outputting reconstructed CSI (i.e., quantized CSI). End-to-end training of encoder 2 and decoder 3 can also be referred to as joint training of encoder 2 and decoder 3. The label for this joint training is the second input CSI (i.e., H). UE ) or target CSI (i.e., H4) NW Through end-to-end training, the output of decoder 3 (i.e., ) and tags (i.e., H UE or H4 NW The training will tend to converge. For example, training 3-2 can be performed by targeting H. UE This can be achieved through self-supervised learning, or training 3-2 can be done by targeting H4. NW This is achieved through supervised learning.

[0471] For example, the second input CSI is the first input CSI included in the first dataset (i.e., H1). NW If the label can be the first input CSI (i.e., H1), then the label can be the first input CSI (i.e., H1). NW ) or target CSI (i.e., H4) NW ).

[0472] For example, the second input CSI is the target CSI (i.e., H4) included in the first dataset. NW If the label can be the target CSI (i.e., H4), then the label can be the target CSI. NW ).

[0473] For example, if the second input CSI is a CSI obtained by the terminal itself, then the tag can be either the CSI obtained by the terminal itself or the target CSI (i.e., H4). NW ).

[0474] It should be understood that in the output of decoder 3 Once the labels meet the performance metrics, training of encoder 2 on the terminal side can be stopped, thus obtaining the trained encoder 2. The terminal side can then deploy the trained encoder 2 to a specific device to perform model inference.

[0475] For example, corresponding to training method three, during model training, the set of functions used to monitor model performance (referred to as function set 3 for ease of distinction and explanation) may include one or more of the following functions: NMSE between the output of decoder 3 (i.e., an example of the third model) and the label, MSE between the output of decoder 3 and the label, MAE between the output of decoder 3 and the label, GCS between the output of decoder 3 and the label, SGCS between the output of decoder 3 and the label, or a weighted sum or weighted average of two or more of the above. The weighting coefficients of each item may be predefined or indicated by the network side; this application does not limit this.

[0476] For a more detailed description of the range that the values ​​of NMSE, MSE, MAE, GCS, or SGCS should satisfy, please refer to the description in step 1020 above. For example, the NMSE / SGCS between the output of decoder 3 and the tag may include the average value and / or distribution of the NMSE / SGCS between the output of decoder 3 and the tag.

[0477] The performance metric corresponding to the aforementioned function set 3 can be the range that the values ​​of each function in function set 3 should satisfy. That is, the performance metric can include the range that the values ​​of one or more of the following functions should satisfy: NMSE between the output of decoder 3 and the label, MSE between the output of decoder 3 and the label, MAE between the output of decoder 3 and the label, GCS between the output of decoder 3 and the label, SGCS between the output of decoder 3 and the label, or a weighted sum or weighted average of two or more of the above. The weighting coefficients for each item can be predefined or indicated by the network side; this application does not limit this.

[0478] As can be seen from Training 3-1 and Training 3-2 above, the output of Decoder 3 is not necessarily the same. Therefore, corresponding to Training 3-1, the set of functions used to monitor model performance (denoted as Function Set 3-1 for ease of distinction and explanation) may include one or more of the following functions.

[0479] The output of decoder 3 (i.e., ) and tags (i.e., H4) NW The NMSE between (example of the fourth NMSE), i.e.

[0480] The output of decoder 3 (i.e., ) and tags (i.e., H4)NW The MSE between (example of the fourth MSE), i.e.

[0481] The output of decoder 3 (i.e., ) and tags (i.e., H4) NW MAE between ) (example of the fourth MAE), i.e.

[0482] The output of decoder 3 (i.e., ) and tags (i.e., H4) NW GCS between )

[0483] The output of decoder 3 (i.e., ) and tags (i.e., H4) NW SGCS between ) or,

[0484] The output of decoder 3 (i.e., ) and tags (i.e., H4) NW The weighted sum or weighted average of two or more of the NMSE, MSE, MAE, GCS or SGCS among the NMSE, MSE, MAE, GCS or SGCS.

[0485] The output of decoder 3 is the reconstructed CSI obtained by processing the first CSI feedback information (i.e., C2) through decoder 3 (such as decompression and prediction, or decompression, dequantization, and prediction). The label is the target CSI (i.e., H4). NW ).

[0486] The performance metric corresponding to the function set 3-1 above can be the range that the values ​​of each function in function set 3-1 should satisfy. That is, the performance metric can include the range that the values ​​of one or more of the following functions should satisfy: Or a weighted sum or weighted average of two or more of the above terms. For ease of distinction and explanation, the performance index corresponding to function set 3-1 will be referred to as performance index 3-1 below.

[0487] Corresponding to Training 3-2, the set of functions used to monitor model performance (referred to as Function Set 3-2 for ease of distinction and explanation) may include one or more of the following functions:

[0488] The output of decoder 3 (i.e. The NMSE between the label and the tag (example of the fifth NMSE), with the label being H UE For example, that is

[0489] The output of decoder 3 (i.e. The MSE between the tag and the label (example of the fifth MSE), with the tag being H UE For example, that is

[0490] The output of decoder 3 (i.e. MAE between the tag and the label (example of the fifth MAE), with the tag being H UE For example, that is

[0491] The output of decoder 3 (i.e. The GCS between the label and the tag (example of the fifth GCS), with the label being H UE For example, that is

[0492] The output of decoder 3 (i.e. The SGCS between the label and the tag (example of the fifth SGCS), with the label being H UE For example, that is or,

[0493] The output of decoder 3 (i.e. The weighted sum or weighted average of two or more of the following: NMSE, MSE, MAE, GCS, or SGCS, between the label and the tag. The weighting coefficients for each item can be predefined or indicated by the network side, and this application does not limit them.

[0494] The output of decoder 3 is the result of the second input CSI (i.e., H) processed by encoder 2 and decoder 3 (the combination of encoder 2 and decoder 3 is an example of the fourth model). UE The reconstructed CSI obtained by processing ) ).

[0495] The performance metric corresponding to the function set 3-2 above can be the range that the values ​​of each function in function set 3-2 should satisfy. That is, the performance metric can include the range that the values ​​of one or more of the following functions should satisfy: Or a weighted sum or weighted average of two or more of the above terms. For ease of distinction and explanation, the performance index corresponding to function set 3-2 will be referred to as performance index 3-2 below.

[0496] As can be seen, the function sets corresponding to training method three include function set 3-1 and function set 3-2, and the corresponding performance metrics include performance metric 3-1 and performance metric 3-2. Function set 3-1 and performance metric 3-1 correspond to training 3-1, and function set 3-2 and performance metric 3-2 correspond to training 3-2. Therefore, the performance metric corresponding to the first dataset can be {performance metric 3-1}, or {performance metric 3-2}, or {performance metric 3-1, performance metric 3-2}, and the function set corresponding to the first dataset can be {function set 3-1}, or {function set 3-2}, or {function set 3-1, function set 3-2}.

[0497] For example, if the terminal side wants to monitor training 3-1, it can train the model based on function set 3-1 and performance metric 3-1. If the terminal side wants to monitor training 3-2, it can train the model based on function set 3-2 and performance metric 3-2. If the terminal side wants to monitor the entire training process, it can train the model for training 3-1 based on function set 3-1 and performance metric 3-1, and train the model for training 3-2 based on function set 3-2 and performance metric 3-2.

[0498] The situations where performance metrics are met and not met have been explained in detail with multiple examples in Training Method 1 above, which can be found in the relevant description above, and will not be repeated here.

[0499] In Design 2, the encoder can be used for compression, not quantization. Correspondingly, the decoder can be used for decompression, not dequantization. The first CSI feedback information can be either compressed CSI or quantized CSI. The corresponding processes differ depending on the first CSI feedback information.

[0500] If the first CSI feedback information is compressed CSI, the process of model training on the terminal side is the same as described above. Figure 14 The illustrated process is similar, except that decoder 3 is used to decompress the CSI feedback information, so the CSI feedback information input to decoder 3 is compressed CSI; encoder 2 is used to compress the input CSI, so the CSI feedback information output by encoder 2 is compressed CSI. Other processes can be found in the above text. Figure 14 The relevant descriptions will not be repeated here.

[0501] If the first CSI feedback information is a quantized CSI, the process of model training on the terminal side can include the following steps.

[0502] Training 3-0: Dequantize the first CSI feedback information (i.e., C2) in the first dataset to obtain a compressed CSI that includes quantization loss.

[0503] Training 3-1': Similar to the process of Training 3-1 above, the difference is that the input of decoder 3 in Training 3-1 is the first CSI feedback information (i.e., C2) in the first dataset, while the input of decoder 3 in Training 3-1' is the compressed CSI containing quantization loss.

[0504] Training 3-2': Based on decoder 3, train encoder 2. Encoder 2 can be a model built by the terminal itself, for example, the structure of which is designed by the terminal. Exemplarily, the terminal can fix the parameters of decoder 3, i.e., use a fixed decoder 3, and train the second input CSI (i.e., H...). UE The second input CSI (i.e., H) is used as the input to encoder 2 and the training labels to perform end-to-end training on encoder 2 and decoder 3. End-to-end training can reuse the procedure provided in training 3-2 above; in this case, training 3-2' is the same as training 3-2. Alternatively, the end-to-end training procedure can differ from that provided in training 3-2: the second input CSI (i.e., H) is used as the input to encoder 2 and the training labels to perform end-to-end training on encoder 2 and decoder 3. UE Input encoder 2, encoder 2 can be used to input the second input CSI (i.e., H UE The CSI is compressed to obtain compressed CSI; the compressed CSI is quantized to obtain quantized CSI; the quantized CSI is dequantized to obtain compressed CSI including quantization loss; the compressed CSI including quantization loss is used as input to decoder 3, which can be used to decompress the compressed CSI including quantization loss to obtain reconstructed CSI (i.e., Through end-to-end training, the output of decoder 3 (i.e., ) and tags (i.e., H UE () tends to approach, for example, training 3-2' can be achieved by targeting H UE This is achieved through self-supervised learning.

[0505] Similar to Training 3-2, the second input CSI used in Training 3-2' can be the first input CSI from the first dataset on the network side (i.e., H). UE It can also be H1 NW ) or target CSI (i.e., H UE It can also be H4 NW It can be CSI obtained by the terminal itself, or it can be CSI obtained by the terminal itself. This application does not limit this.

[0506] For a more detailed description of the set of functions and performance metrics used to monitor model performance for Training 3-1' and Training 3-2', please refer to the description of the set of functions and performance metrics used to monitor model performance for Training 3-1 and Training 3-2 above.

[0507] Example 3: The network side sends the first dataset, which includes the fifth data (which can be represented as C2) and the sixth data (which can be represented as H4). NW ) and the seventh data (which can be represented as H2) NW The terminal side trains the model based on the first dataset. The fifth data point can be either compressed CSI (i.e., unquantized CSI) or quantized CSI, without limitation. For ease of understanding, the fifth data point will be referred to as the first CSI feedback information, the sixth data point as the target CSI, and the seventh data point as the SRS measurement value.

[0508] For example, before sending the first dataset, the network can first perform joint training of encoder 1 and decoder 1 (an example of the second model) to obtain the first dataset. The method for the network to perform joint training of encoder 1 and decoder 1 can be referred to the description in Example 1 above.

[0509] For example, after the network side performs joint training of encoder 1 and decoder 1, it sends a first dataset containing first CSI feedback information, target CSI and SRS measurements to the terminal side.

[0510] Based on the first dataset, the terminal side can use the following training methods to train the model.

[0511] Training Method 4: The terminal first trains the third model (e.g., decoder 4) based on the first dataset, and then trains the fourth model (e.g., a combination of encoder 2 and decoder 4) based on the third model. The fourth model includes the first model (e.g., encoder 2). In other words, in Training Method 4, the first CSI feedback information, SRS measurement values, and target CSI included in the first dataset are used to train the third model, and the third model is used to train the fourth model to obtain the first model.

[0512] In the training of the third model, the input of the third model includes the first CSI feedback information and the SRS measurement value, and the output of the third model corresponds to the target CSI.

[0513] In the training of the fourth model, data #2 is used as the input of the first model (or replaced with the input of the fourth model), data #2 is used as the ground truth value corresponding to the output of the third model (or replaced with the ground truth value corresponding to the output of the fourth model), and the output of the first model is used as the input of the third model.

[0514] Optionally, data #2 includes one or more of the following: the target CSI or a CSI other than the target CSI. The CSI other than the target CSI can be a CSI acquired by the terminal itself, such as a CSI obtained through measurement or simulation. Hereinafter, data #2 will be referred to as the second input CSI. The channel information represented by data #2 corresponds to the same channel as the channel information represented by the target CSI.

[0515] Referring to the description in Training Method 1 above, there are two possible designs for the encoder on the terminal side. Design 1: The encoder 2 to be trained on the terminal side can be used for both compression and quantization. Design 2: The encoder 2 to be trained on the terminal side can be used for compression but not for quantization.

[0516] The following sections will explain these two possible designs.

[0517] In Design 1, the encoder can be used for compression and quantization, and correspondingly, the decoder can be used for dequantization and decompression. The first CSI feedback information in the first dataset can be the quantized CSI.

[0518] For the process of model training on the terminal side, please refer to [link / reference]. Figure 15 This may include the following steps.

[0519] Training 4-1: Train decoder 4 separately based on the first dataset. This decoder 4 can be a nominal decoder. The terminal side will use the first CSI feedback information (i.e., C2) and SRS measurement (i.e., H2) from the first dataset. NW ) as input to decoder 4, taking the target CSI (i.e., H4) from the first dataset as input. NW The first CSI feedback information (i.e., C2) is used as a label to train decoder 4. Decoder 4 can be used to dequantize and decompress the first CSI feedback information (i.e., C2) and the SRS measurement (i.e., H2). NW The process is performed to output the reconstructed CSI (i.e., By training decoder 4, the output of decoder 4 (i.e., ) and tags (i.e., H4) NW They tend to approach each other. For example, training 4-1 can be done by targeting H4. NW This is achieved through supervised learning.

[0520] It should be understood that in the output of decoder 4 With tag H4 NW When the performance metrics are met, the training of the decoder 4 model on the terminal side can be stopped, thus obtaining the trained decoder 4.

[0521] Training 4-2: Train encoder 2 based on decoder 4. Encoder 2 can be a model built by the terminal itself, for example, the model's structure is designed by the terminal. The terminal can fix the parameters of decoder 4 (i.e., use a frozen decoder 4; in other words, do not change the parameters of decoder 4 during encoder 2 training), and input the second input CSI (i.e., H... UE The second input CSI used in training 4-2 is used as input to encoder 2 to perform end-to-end training of encoder 2 and decoder 4. It should be understood that the second input CSI used in training 4-2 can be the target CSI (i.e., H) from the first dataset on the network side. UE It can be H4 NW It can also be a CSI with the same format as the target CSI that is obtained by the terminal itself, and this application does not limit it.

[0522] For example, the end-to-end training process is as follows: The second input CSI (i.e., H) is... UE The second input CSI is input to encoder 2, which can be used to compress and quantize the CSI, outputting CSI feedback information (i.e., quantized CSI); the output of encoder 2 is used as input to decoder 4, which can be used to dequantize and decompress the CSI feedback information output by encoder 2, outputting reconstructed CSI (i.e., quantized CSI). End-to-end training of encoder 2 and decoder 4 can also be referred to as joint training of encoder 2 and decoder 4. The label for this joint training is the second input CSI (i.e., H). UE Through end-to-end training, the output of decoder 4 (i.e., ) and tags (i.e., H UE The training will tend to converge. For example, training 4-2 can be performed by targeting H. UE This is achieved through self-supervised learning.

[0523] It should be understood that in the output of decoder 4 Once the labels meet the performance metrics, training of encoder 2 on the terminal side can be stopped, thus obtaining the trained encoder 2. The terminal side can then deploy the trained encoder 2 to a specific device to perform model inference.

[0524] For example, corresponding to training method four, during model training, the set of functions used to monitor model performance (referred to as function set 4 for ease of distinction and explanation) may include one or more of the following functions: NMSE between the output of decoder 4 (i.e., an example of the third model) and the label, MSE between the output of decoder 4 and the label, MAE between the output of decoder 4 and the label, GCS between the output of decoder 4 and the label, SGCS between the output of decoder 4 and the label, or a weighted sum or weighted average of two or more of the above. The weighting coefficients of each item may be predefined or indicated by the network side; this application does not limit this.

[0525] For a more detailed description of the range that the values ​​of NMSE, MSE, MAE, GCS, or SGCS should satisfy, please refer to the description in step 1020 above. For example, the NMSE / SGCS between the output of decoder 4 and the label may include the average and / or distribution of the NMSE / SGCS between the output of decoder 4 and the label. The performance metric corresponding to the above function set 4 may be the range that the values ​​of each function in function set 4 should satisfy. That is, the performance metric may include the range that the values ​​of one or more of the following functions should satisfy: the NMSE between the output of decoder 4 and the label, the MSE between the output of decoder 4 and the label, the MAE between the output of decoder 4 and the label, the GCS between the output of decoder 4 and the label, the SGCS between the output of decoder 4 and the label, or a weighted sum or weighted average of two or more of the above. The weighting coefficients of each item may be predefined or indicated by the network side, and this application does not limit them.

[0526] As can be seen from Training 4-1 and Training 4-2 above, the output of Decoder 4 is not necessarily the same. Therefore, corresponding to Training 4-1, the set of functions used to monitor model performance (denoted as Function Set 4-1 for ease of distinction and explanation) may include one or more of the following functions.

[0527] The output of decoder 4 (i.e., ) and tags (i.e., H4) NW The NMSE between (example of the sixth NMSE), i.e.

[0528] The output of decoder 4 (i.e., ) and tags (i.e., H4) NW The MSE between (example of the sixth MSE), i.e.

[0529] The output of decoder 4 (i.e., ) and tags (i.e., H4) NW MAE between ) (example of the sixth MAE), i.e.

[0530] The output of decoder 4 (i.e., ) and tags (i.e., H4) NW GCS between )

[0531] The output of decoder 4 (i.e., ) and tags (i.e., H4) NW SGCS between ) or,

[0532] The output of decoder 4 (i.e., ) and tags (i.e., H4) NW The weighted sum or weighted average of two or more of the NMSE, MSE, MAE, GCS or SGCS among the NMSE, MSE, MAE, GCS or SGCS.

[0533] The output of decoder 4 is the first CSI feedback information (i.e., C2) and the SRS measurement value (i.e., H2) obtained by decoder 4. NW The reconstructed CSI obtained by processing ) is, The label is the target CSI (i.e., H4). NW ).

[0534] The performance metric corresponding to the function set 4-1 above can be the range that the values ​​of each function in function set 4-1 should satisfy. That is, the performance metric can include the range that the values ​​of one or more of the following functions should satisfy: Or a weighted sum or weighted average of two or more of the above terms. For ease of distinction and explanation, the performance index corresponding to function set 4-1 will be referred to as performance index 4-1 below.

[0535] Corresponding to Training 4-2, the set of functions used to monitor model performance (referred to as Function Set 4-2 for ease of distinction and explanation) may include one or more of the following functions.

[0536] The output of decoder 4 (i.e., ) and tags (i.e., H UE The NMSE between ) (example of the seventh NMSE), i.e.

[0537] The output of decoder 4 (i.e., ) and tags (i.e., H UE The MSE between ) (example of the seventh MSE), i.e.

[0538] The output of decoder 4 (i.e., ) and tags (i.e., H UEThe MAE between ) (example of the seventh MAE), i.e.

[0539] The output of decoder 4 (i.e., ) and tags (i.e., H UE The GSC (example of the seventh GCS) between ) is

[0540] The output of decoder 4 (i.e., ) and tags (i.e., H UE The SGCS between (example of the seventh SGCS), i.e. or,

[0541] The output of decoder 4 (i.e., ) and tags (i.e., H UE The weighted sum or weighted average of two or more of the following: NMSE, MSE, MAE, GCS, or SGCS. The weighting coefficients for each item can be predefined or indicated by the network side, and this application does not limit them.

[0542] The output of decoder 4 is the result of the second input CSI (i.e., H) processed by encoder 2 and decoder 4 (the combination of encoder 2 and decoder 4 is an example of the fourth model). UE The reconstructed CSI obtained by processing ) is, ).

[0543] The performance metric corresponding to the function set 4-2 above can be the range that the values ​​of each function in function set 4-2 should satisfy. That is, the performance metric can include the range that the values ​​of one or more of the following functions should satisfy: Or a weighted sum or weighted average of two or more of the above terms. For ease of distinction and explanation, the performance index corresponding to function set 4-2 will be referred to as performance index 4-2 below.

[0544] As can be seen, the function sets corresponding to training method four include function set 4-1 and function set 4-2, and the corresponding performance metrics include performance metric 4-1 and performance metric 4-2. Function set 4-1 and performance metric 4-1 correspond to training 4-1, and function set 4-2 and performance metric 4-2 correspond to training 4-2. Therefore, the performance metric corresponding to the first dataset can be {performance metric 4-1}, or {performance metric 4-2}, or {performance metric 4-1, performance metric 4-2}, and the function set corresponding to the first dataset can be {function set 4-1}, or {function set 4-2}, or {function set 4-1, function set 4-2}.

[0545] For example, if the terminal side wants to monitor training 4-1, it can train the model based on function set 4-1 and performance metric 4-1. If the terminal side wants to monitor training 4-2, it can train the model based on function set 4-2 and performance metric 4-2. If the terminal side wants to monitor the entire training process, it can train the model for training 4-1 based on function set 4-1 and performance metric 4-1, and train the model for training 4-2 based on function set 4-2 and performance metric 4-2.

[0546] The situations where performance metrics are met and not met have been explained in detail with multiple examples in Training Method 1 above, which can be found in the relevant description above, and will not be repeated here.

[0547] In Design 2, the encoder can be used for compression, not quantization. Correspondingly, the decoder can be used for decompression, not dequantization. The first CSI feedback information can be either compressed CSI or quantized CSI. The corresponding processes differ depending on the first CSI feedback information.

[0548] If the first CSI feedback information is compressed CSI, the process of model training on the terminal side is the same as described above. Figure 15 The illustrated process is similar, except that decoder 4 is used to decompress the CSI feedback information, so the CSI feedback information input to decoder 4 is compressed CSI; encoder 2 is used to compress the input CSI, so the CSI feedback information output by encoder 2 is compressed CSI. Other processes can be found in the above text. Figure 15 The relevant descriptions will not be repeated here.

[0549] If the first CSI feedback information is a quantized CSI, the process of model training on the terminal side can include the following steps.

[0550] Training 4-0: Dequantize the first CSI feedback information in the first dataset to obtain a compressed CSI containing quantization loss.

[0551] Training 4-1': Similar to the process of Training 4-1 above, the difference is that the input of decoder 4 in Training 4-1 is the first CSI feedback information (i.e., C2) and SRS measurement value (i.e., H2) from the first dataset. NW The input to decoder 4 in training 4-1' is compressed CSI and SRS measurements (i.e., H2) including quantization loss. NW ).

[0552] Training 4-2': Based on decoder 4, train encoder 2. Encoder 2 can be a model built by the terminal itself, for example, the structure of which is designed by the terminal. Exemplarily, the terminal can fix the parameters of decoder 4, i.e., use a fixed decoder 4, and train the second input CSI (i.e., H...). UE The second input CSI (i.e., H) is used as the input to encoder 2 and the training labels to perform end-to-end training on encoder 2 and decoder 4. End-to-end training can reuse the procedure provided in training 4-2 above; in this case, training 4-2' is the same as training 4-2. Alternatively, the end-to-end training procedure can differ from that provided in training 4-2: the second input CSI (i.e., H) is used as the input to encoder 2 and the training labels to perform end-to-end training on encoder 2 and decoder 4. UE Input encoder 2, encoder 2 can be used to input the second input CSI (i.e., H UE The CSI is compressed to obtain a compressed CSI; the compressed CSI is quantized to obtain a quantized CSI; the quantized CSI is dequantized to obtain a compressed CSI containing quantization loss; the compressed CSI containing quantization loss is used as input to decoder 4, which can be used to decompress the compressed CSI containing quantization loss to obtain the reconstructed CSI (i.e., Through end-to-end training, the output of decoder 4 (i.e., ) and tags (i.e., H UE () tends to approach, for example, training 4-2' can be achieved by targeting H UE This is achieved through self-supervised learning.

[0553] Similar to Training 4-2, the second input CSI used in Training 4-2' can be the target CSI from the first dataset on the network side (i.e., H). UE It can also be H4 NW It can also be a CSI with the same format as the target CSI that is obtained by the terminal itself, and this application does not limit it.

[0554] For a more detailed description of the set of functions and performance metrics used to monitor model performance for Training 4-1' and Training 4-2', please refer to the description of the set of functions and performance metrics used to monitor model performance for Training 4-1 and Training 4-2 above.

[0555] The foregoing description, in conjunction with several accompanying drawings, illustrates model training based on different datasets and different models. These examples are illustrative only and should not be construed as limiting the scope of this application. Based on the examples provided above, those skilled in the art can make simple combinations and / or modifications to obtain more possible embodiments.

[0556] It should also be understood that the above text, in combination with... Figures 10 to 15In the illustrated embodiment, the terminal side can be a terminal device, and the network side can be a network device. In another design, the terminal side can also include a terminal device and the host or cloud server of the OTT system; the network side can include network devices and intelligent network elements. In this case, devices on the terminal side can communicate with each other, and devices on the network side can also communicate with each other. The following will explain... Figure 16 The method 1600 shown describes the specific implementation process of the method provided in this application in an architecture consisting of a host or cloud server (hereinafter referred to as an OTT system server), terminal devices, network devices, and intelligent network elements that are deployed with an OTT system. This process assumes that the training device is an OTT system server.

[0557] Figure 16 Another illustrative flowchart of the communication method provided in the embodiments of this application. Figure 16 The method 1600 shown may include the following steps:

[0558] Step 1601: The intelligent network element sends a dataset to the network device, which is used for model training.

[0559] Step 1601 is an optional step. If the dataset is obtained by a network device, then method 1600 may not include step 1601.

[0560] Step 1602: The network device sends the dataset to the terminal device.

[0561] In one possible implementation, if the dataset is obtained by an intelligent network element and the model training is performed by an OTT system server on the terminal side, then method 1600 can execute step 1604 below instead of step 1602.

[0562] Step 1603: The terminal device sends the dataset to the OTT system server.

[0563] Step 1603 is an optional step. If the terminal device is a device that performs model training, then method 1600 may not include step 1603.

[0564] Steps 1601-1603 illustrate the process by which the intelligent network element transmits data sets over the air interface between the network device and the terminal device. In another implementation, the intelligent network element can send the data sets to the OTT system server via wired transmission, as shown in step 1604, which is indicated by a dashed line in the figure.

[0565] Step 1605: The network device sends first information to the terminal device, which is used for performance indicators.

[0566] Step 1606: The terminal device sends information indicating the performance metric to the OTT system server. The information indicating the performance metric sent by the terminal device to the OTT system server can be primary information or other information, without limitation.

[0567] In another implementation, the terminal device can also determine the performance indicators based on the first information, and then send the information indicating the performance indicators to the OTT system server. That is, step 1606 can also be replaced by: the terminal device sending the information indicating the performance indicators to the OTT system server.

[0568] Step 1606 is an optional step. If the terminal device is a device that performs model training, then method 1600 may not include step 1606.

[0569] Step 1607: The OTT system server trains the model based on performance metrics and the received dataset.

[0570] Step 1608: The OTT system server determines whether the performance indicators are met.

[0571] Since the OTT system server is a training device, it can determine whether the performance indicators are met on its own.

[0572] Step 1609: If the performance indicators are met, the OTT system server executes the inference task of the first model.

[0573] In one possible implementation, steps 1607 to 1609 described above can be executed by a terminal device.

[0574] Step 1610: If the performance indicators are met, the OTT system server sends information indicating that the performance indicators are met.

[0575] Step 1610 is an optional step. If step 1608 is executed by the terminal device, then method 1600 may not include step 1610. Step 1608 being executed by the terminal device may include the following situations: (1) The terminal device performs model training and determines whether the performance indicators are met during the model training process; (2) The OTT system server performs model training, and the terminal device can predict whether the performance indicators are met based on the capabilities of the OTT system server; (3) The OTT system server performs model training, and the OTT system server can send the output of the trained model during the model training process to the terminal device, so that the terminal device can determine whether the performance indicators are met.

[0576] Step 1611: The terminal device sends a fourth message to the network device, which indicates that the performance indicators have been met.

[0577] The information received by the terminal device from the OTT system server in step 1610 to indicate that the performance indicators have been met can be the same as or different from the second information sent by the terminal device in step 1611. This application does not limit this.

[0578] Step 1612: The network device sends information to the intelligent network element to indicate that the performance indicators have been met.

[0579] The fourth information received by the network device from the terminal device in step 1611 and the information sent by the network device in step 1612 to indicate that the performance indicators have been met can be the same information or different information. This application does not limit this.

[0580] Steps 1610-1612 illustrate the process by which the OTT system server transmits information indicating that performance indicators have been met via the air interface between the network device and the terminal device. In another implementation, the OTT system server can send the information indicating that performance indicators have been met to the intelligent network element via wired transmission. In other words, steps 1610-1612 are optional.

[0581] In one possible implementation, if step 1608 is executed by the terminal device, method 1600 may execute step 1611 instead of steps 1610 and 1612. In other words, steps 1610 and 1612 are optional steps.

[0582] Step 1613: If the performance indicators are not met, the OTT system server sends information to the terminal device indicating that the performance indicators are not met.

[0583] Step 1613 is an optional step. If step 1608 above is performed by the terminal device, then method 1600 may not include step 1613.

[0584] Step 1614: The terminal device sends third information to the network device, which can be used for one or more of the following: indicating that the performance metric is not met; requesting a change of dataset; requesting a change of performance metric; or requesting to shut down model training.

[0585] Terminal devices can send third-party information if performance indicators are not met.

[0586] The information received by the terminal device from the OTT system server in step 1613 to indicate that the performance indicators are not met can be the same as or different from the third information sent by the terminal device in step 1614. This application does not limit this.

[0587] Step 1615: The network device sends a fifth message to the intelligent network element. This fifth message can be used for one or more of the following: indicating that the performance metric is not met; requesting a change of the dataset; requesting a change of the performance metric; or requesting to shut down model training.

[0588] The fifth message sent by the network device in step 1615 and the third message received from the terminal device in step 1614 can be the same message or different messages; this application does not limit this.

[0589] Steps 1613-1615 illustrate the process by which the OTT system server transmits information indicating that performance metrics are not met via the air interface between the network device and the terminal device. In another implementation, the OTT system server can send this information to the intelligent network element via wired transmission. In other words, steps 1613-1615 are optional.

[0590] In one possible implementation, if step 1608 is executed by the terminal device, then method 1600 may execute step 1614 instead of steps 1613 and 1615. In other words, steps 1613 and 1615 are optional steps.

[0591] For detailed explanations of each step in Method 1600, please refer to the above text. Figure 10 The relevant descriptions in Method 1000 will not be repeated here. Furthermore, the technical solution shown in Method 1600 corresponds to the technical solution shown in Method 1000 above, and therefore the beneficial effects obtained are similar, and will not be repeated here either.

[0592] It should be understood that Figure 16 The process shown is just to Figure 10 The method 1000 shown is one possible implementation process applied to a system architecture deploying an OTT system server and intelligent network elements, and should not be construed as limiting this application in any way. Those skilled in the art can, based on the same concept, make simple substitutions or modifications to one or more steps of method 1000 to achieve the same effect. Such simple modifications or substitutions should fall within the protection scope of this application.

[0593] It should also be understood that the processes listed above are merely examples, and this application includes, but is not limited to, them. For example, Figure 16 The terminal device and OTT system server can also be replaced by the terminal device itself. In this case, the communication between the terminal device and the OTT system server is internal communication within the terminal device; or... Figure 16 The network devices and intelligent network elements mentioned can also be replaced with network devices. In this case, the communication between the network devices and intelligent network elements is internal communication between the network devices. For the sake of simplicity, no illustrations are provided in this article.

[0594] It should be understood that in the various embodiments shown above in conjunction with the accompanying drawings, the sequence number of each step does not imply the order of execution. The execution order of each step should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0595] It should also be understood that the various embodiments exemplified above can also be applied to scenarios where the terminal sends a dataset and / or model to the network side for model training via the network side. In this case, the steps on the terminal side in the above embodiments can be performed by the network side, and vice versa. Those skilled in the art can make simple modifications or substitutions based on the same concept to obtain more embodiments.

[0596] In the various embodiments of this application, unless otherwise specified or in case of logical conflict, the terminology and / or descriptions of different embodiments are consistent and can be referenced by each other. The technical features of different embodiments can be combined to form new embodiments according to their inherent logical relationship.

[0597] In the above embodiments, exemplary descriptions are mainly based on devices in the current network architecture (such as terminal devices, network devices, OTT system servers, intelligent network elements, etc.). The specific form of the devices is not limited in the embodiments of this application. For example, devices that can achieve the same function in the future can also be applied to the methods provided in the embodiments of this application.

[0598] It is understood that the methods and operations implemented by devices (such as terminal devices, network devices, OTT system servers, and smart network elements) in the above-described method embodiments can also be implemented by components of the devices (such as chips or circuits).

[0599] The methods provided in the embodiments of this application have been described in detail above with reference to several accompanying drawings. The apparatus provided in the embodiments of this application will now be described with reference to the accompanying drawings.

[0600] Figure 17 and Figure 18 This is a schematic block diagram of possible apparatuses provided in the embodiments of this application. These apparatuses can be used to implement the functions on the terminal side or network side in the above method embodiments, and thus can also achieve the beneficial effects of the above method embodiments.

[0601] Figure 17 A schematic block diagram of the apparatus provided in the embodiments of this application. Figure 17 The device 1700 shown may include a processing module 1710 and a communication module 1720.

[0602] In one possible design, device 1700 can be used to implement Figures 10 to 16The communication method is implemented on the terminal side in any of the embodiments shown. For example, the processing module 1710 is used to implement the model training performed on the terminal side in each method embodiment, determine whether the first indicator is met, and perform the inference task of the first model, etc., and other processing-related steps; the communication module 1720 is used to implement the sending and / or receiving steps performed on the terminal side in each method embodiment, such as receiving a dataset, receiving first information, receiving second information, or sending third information, or one or more of these steps.

[0603] For example, the processing module 1710 can be used to: acquire a dataset for model training; the processing module 1710 can also be used to train a model based on the dataset and performance metrics to obtain a first model.

[0604] Optionally, the communication module 1720 can be used to receive first information, which is used to indicate performance indicators.

[0605] Optionally, the communication module 1720 can be used to receive second information, which is used to instruct the fourth and fifth data to be used for model training of the first model, or to instruct the fourth and fifth data to be used for model training of the third model.

[0606] Optionally, the processing module 1710 can be used to perform the inference task of the first model if the performance indicators are met.

[0607] Optionally, the communication module 1720 is used to send third information if performance indicators are not met.

[0608] Optionally, the communication module 1720 is used to receive a dataset.

[0609] For a more detailed description of the processing module 1710 and the communication module 1720 mentioned above, please refer to [link / reference]. Figures 10 to 16 The relevant descriptions in the method embodiments shown are directly obtained and will not be repeated here.

[0610] In another possible design, device 1700 can be used to achieve Figures 10 to 16 The communication method is implemented by the network side in any of the embodiments shown. For example, the processing module 1710 is used to implement the processing-related steps such as generating datasets and / or models performed by the network side in each method embodiment; the communication module 1720 is used to implement the sending and / or receiving steps performed by the network side in each method embodiment, such as sending datasets, sending first information, sending second information, or receiving third information, or one or more of these steps.

[0611] For example, the communication module 1720 can be used to: send a dataset for model training; the communication module 1720 can also be used to send first information indicating performance metrics.

[0612] Optionally, the communication module 1720 can be used to send second information, which is used to instruct the fourth and fifth data to be used for model training of the first model, or to instruct the fourth and fifth data to be used for model training of the third model.

[0613] Optionally, the communication module 1720 is used to receive third information.

[0614] For a more detailed description of the processing module 1710 and the communication module 1720 mentioned above, please refer to [link / reference]. Figures 10 to 16 The relevant descriptions in the method embodiments shown are directly obtained and will not be repeated here.

[0615] It should be noted that the communication module can also be called a transceiver module, transceiver unit, transceiver, transceiver device, or transceiver apparatus, etc. The processing module can also be called a processor, processing board, processing unit, or processing apparatus, etc. Optionally, the communication module is used to perform the sending and receiving operations on the terminal side or network side in the above method. The device in the communication module that implements the receiving function can be considered as the receiving module, and the device in the communication module that implements the sending function can be considered as the sending module; that is, the communication module can include both a receiving module and a sending module.

[0616] It should also be noted that, in one possible design, the aforementioned processing module and / or communication module can be implemented through virtual modules. For example, the processing module can be implemented through software functional units or virtual devices, and the communication module can be implemented through software functions or virtual devices. In another possible design, the processing module or communication module can also be implemented through physical devices. For example, if the device is implemented using a chip / chip circuit, the communication module can be an input / output circuit and / or a communication interface, performing input operations (corresponding to the aforementioned receiving operation) and output operations (corresponding to the aforementioned sending operation); the processing module can be an integrated processor, a microprocessor, or an integrated circuit.

[0617] The module division in this embodiment is illustrative and represents only one logical functional division; in actual implementation, other division methods may be used. Furthermore, the functional modules in the various examples of this embodiment can be integrated into a single processor, exist as separate physical entities, or be integrated into a single module. The integrated modules described above can be implemented in hardware or as software functional modules.

[0618] Figure 18 This is a schematic diagram of the structure of a communication device provided in yet another embodiment of this application. Figure 18 As shown, the device 1800 includes a processing circuit 1810 and a communication circuit 1820. The processing circuit 1810 and the communication circuit 1820 are coupled to each other.

[0619] It can be understood that the processing circuit 1810 can be one or more processors, or it can be all or part of the processing functions of one or more processors.

[0620] Understandably, the communication circuit 1820 can be a transceiver or an input / output interface.

[0621] Optionally, the device 1800 may further include a memory 1830 for storing instructions executed by the processing circuit 1810, or storing input data required for the running instructions of the processing circuit 1810, or storing data generated after the running instructions of the processing circuit 1810.

[0622] It is understood that the memory 1830 may be located outside the processing circuit 1810, or inside the processing circuit 1810.

[0623] As an example, the processing circuit 1810 is used to implement the functions of the processing module 2010, and the communication circuit 1820 is used to implement the functions of the communication module 2020.

[0624] As an example, device 1800 can be a communication device or a chip used in a communication device.

[0625] When device 1800 is a communication device, the communication circuit can be a transceiver; when device 1800 is a chip, the communication circuit can be an input / output circuit, a bus, pins, or other types of communication interfaces. The input circuit in the input / output circuit can be used for receiving, and the output interface can be used for transmitting.

[0626] When the aforementioned device is a chip or OTT device applied to the terminal side, the terminal-side chip or OTT device implements the terminal-side functions described in the above method embodiments, such as implementing terminal-side processing functions. The terminal-side chip or OTT device receiving information from the network side can be understood as the information being first received by other modules on the terminal side (such as an RF module or antenna), and then sent by these modules to the terminal-side chip or OTT device. The terminal-side chip or OTT device sending information to the network side can be understood as the information being first sent by the terminal-side chip or OTT device to other modules on the terminal side (such as an RF module or antenna), and then sent by these modules to the network side.

[0627] When the aforementioned device is a chip or OTT device applied to the network side, the network-side chip or OTT device implements the network-side functions described in the above method embodiments, such as implementing network-side processing functions. The network-side chip or OTT device receiving information from the terminal side can be understood as the information being first received by other modules on the network side (such as radio frequency modules or antennas), and then sent to the network-side chip or OTT device by these modules. The network-side chip or OTT device sending information to the terminal side can be understood as the information being first sent by the network-side chip or OTT device to other modules on the network side (such as radio frequency modules or antennas), and then sent to the terminal side by these modules.

[0628] This application also provides a computer program product that, when run on a processor, can implement the communication method executed by the terminal side or the communication method executed by the network side in the above method embodiments.

[0629] This application also provides a computer-readable storage medium containing computer instructions that, when executed on a processor, can implement the communication method executed by the terminal side or the communication method executed by the network side in the above method embodiments.

[0630] This application also provides a communication system, including the aforementioned terminal side and network side. The terminal side can be used to implement the communication method implemented by the terminal side in the above method embodiments, and the network side can be used to implement the communication method implemented by the network side in the above method embodiments.

[0631] It is understood that the processor in the embodiments of this application may be any of the following devices or all or part of the circuitry used for processing functions: a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), processors for AI, field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor may be a microprocessor or any conventional processor.

[0632] The processor used for AI can be one or more of the following: graphics processing unit (GPU), neural processing unit (NPU), tensor processing unit (TPU), and data processing unit (DPU).

[0633] For example, one possible implementation of a processor for AI could be... Figure 19 The AI ​​processor shown is 1900.

[0634] like Figure 19 As shown, the AI ​​processor 1900 may include one or more of the following: AI core, digital vision pre-processing (DVPP) module, task scheduler (TS), L3 cache, AI CPU, control CPU, L2 cache, universal serial bus (USB) interface, network card, peripheral component interconnect express (PCIe) interface (PCIe is a high-speed serial computer expansion bus standard), double data rate (DDR) / high bandwidth memory (HBM) interface, generation purpose input / output (GPIO) / inter-integrated circuit (I2C) bus, etc.

[0635] The terms “unit”, “module”, etc., used in this specification may be used to refer to computer-related entities, hardware, firmware, combinations of hardware and software, software, or software in execution.

[0636] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0637] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0638] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0639] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0640] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0641] In the above embodiments, the functions of each functional unit can be implemented entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. This computer program product includes one or more computer instructions (programs). When the computer program instructions (programs) are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., digital video discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0642] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0643] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A communication method, characterized in that, include: Obtain a dataset for model training to obtain a first model; The first model is obtained by training the model based on the dataset and performance metrics. The performance metrics correspond to the inference task of the first model. The inference task of the first model is to perform a first process on the first data to obtain the second data. The first process includes compression, or the first process includes compression and quantization. The first data is used to characterize the first channel information; The second data is used to determine third data based on the second model. The third data is used to characterize second channel information. The second channel information includes at least one of the following: part or all of the first channel information, or third channel information, wherein the third channel information is different from the first channel information. The inference task of the second model includes processing the second data to obtain the third data.

2. The method according to claim 1, characterized in that, The second channel information includes the third channel information, as well as part or all of the first channel information.

3. The method according to claim 1 or 2, characterized in that, Before training the model, the method further includes: Receive first information, which is used to indicate the performance metric.

4. The method according to any one of claims 1 to 3, characterized in that, The dataset includes one or more of the following: fourth data, fifth data, sixth data, or seventh data; The fourth data is used to characterize the fourth channel information; The fifth data is obtained by performing the first processing on the fourth data, and the fifth data is used to determine the eighth data, or the fifth data and the seventh data are used to determine the eighth data; The sixth data is the truth value corresponding to the eighth data; The eighth data is used to characterize the fifth channel information, and the seventh data is used to characterize the sixth channel information. The fifth channel information includes at least one of the following: part or all of the fourth channel information, part or all of the sixth channel information, or the seventh channel information, wherein the seventh channel information is different from the fourth channel information and different from the sixth channel information.

5. The method according to claim 4, characterized in that, The dataset includes the fourth data and the fifth data; The fourth data and the fifth data are used for model training of the first model. During model training, the input of the first model includes the fourth data, and the true value corresponding to the output of the first model includes the fifth data. or, The fourth and fifth data are used for model training of the third model, and the third model is used for model training of the fourth model to obtain the first model. The fourth model includes the first model and the third model. In the model training of the third model, the input of the third model includes the fifth data, and the true value corresponding to the output of the third model includes the fourth data. The third model is used for model training of the fourth model, which includes: in the model training of the fourth model, data #1 is used as the input of the first model and as the true value corresponding to the output of the third model, and the output of the first model is used as the input of the third model.

6. The method according to claim 5, characterized in that, The fourth and fifth data are used for training the first model. The performance metrics include one or more of the following ranges: first mean squared error (MSE); first normalized mean squared error (NMSE); first mean absolute error (MAE); or, a weighted sum of multiples of the first MSE, first NMSE, or first MAE, or a weighted average. Wherein, the first MSE is the MSE between the information obtained by performing the first processing on the fourth data through the first model and the fifth data; The first NMSE is the NMSE between the information obtained by performing the first processing on the fourth data using the first model and the fifth data. The first MAE is the MAE between the information obtained by performing the first processing on the fourth data using the first model and the fifth data.

7. The method according to claim 5, characterized in that, The fourth and fifth data are used for model training of the third model; the performance metrics include one or more of the following ranges: second MSE; second NMSE; second MAE; third MSE; third NMSE; third MAE; third generalized cosine similarity (GCS); third squared generalized cosine similarity (SGCS); or, a weighted sum or weighted average of multiples of the second MSE, second NMSE, second MAE, third MSE, third NMSE, third MAE, third GCS, or third SGCS. The second MSE is the MSE between the reconstructed CSI and the fourth data obtained by performing a second processing on the fifth data using the third model; The second NMSE is the NMSE between the reconstructed CSI and the fourth data obtained by performing a second processing on the fifth data using the third model; The second MAE is the MAE between the reconstructed CSI and the fourth data obtained by performing a second processing on the fifth data using the third model; The third MSE is the MSE between the reconstructed CSI and the data #1 obtained by processing the data #1 using the fourth model; The third NMSE is the NMSE between the reconstructed CSI and the data #1 obtained by processing the data #1 using the fourth model; The third MAE is the MAE between the reconstructed CSI and the data #1 obtained by processing the data #1 using the fourth model; The third GCS is the generalized cosine similarity GCS between the reconstructed CSI and the data #1 obtained by processing the data #1 using the fourth model; The third SGCS is the squared generalized cosine similarity SGCS between the reconstructed CSI and the data #1 obtained by processing the data #1 using the fourth model; The second process includes decompression, or the second process includes decompression and dequantization.

8. The method according to any one of claims 5 to 7, characterized in that, Before training the model based on the dataset and performance metrics, the method further includes: Receive second information, which is used to instruct the fourth data and the fifth data to be used for model training of the first model, or to instruct the fourth data and the fifth data to be used for model training of the third model.

9. The method according to claim 4, characterized in that, The dataset includes the fourth data, the fifth data, and the sixth data; Wherein, the fifth data and the sixth data are used for model training of the third model, and the third model is used for model training of the fourth model to obtain the first model, and the fourth model includes the first model and the third model; in the model training of the third model, the input of the third model includes the fifth data, and the true value corresponding to the output of the third model includes the sixth data; the model training of the third model for the fourth model includes: in the model training of the fourth model, the ninth data is used as the input of the first model, the output of the first model is used as the input of the third model, and the true value corresponding to the output of the third model includes the tenth data.

10. The method according to claim 9, characterized in that, The performance metrics include one or more of the following ranges that should be met: fourth MSE; fourth NMSE; fourth MAE; fifth MSE; fifth NMSE; fifth MAE; fifth GCS; fifth SGCS; or, a weighted sum or weighted average of multiple of the following: fourth MSE, fourth NMSE, fourth MAE, fifth MSE, fifth NMSE, fifth MAE, fifth GCS, or fifth SGCS. The fourth MSE is the MSE between the reconstructed CSI and the sixth data obtained by processing the fifth data using the third model; The fourth NMSE is the NMSE between the reconstructed CSI and the sixth data obtained by processing the fifth data using the third model; The fourth MAE is the MAE between the reconstructed CSI and the sixth data obtained by processing the fifth data using the third model; The fifth MSE is the MSE between the reconstructed CSI and the tenth data obtained by processing the ninth data using the fourth model; The fifth NMSE is the NMSE between the reconstructed CSI and the tenth data obtained by processing the ninth data using the fourth model; The fifth MAE is the MAE between the reconstructed CSI and the tenth data obtained by processing the ninth data using the fourth model; The fifth GCS is the GCS between the reconstructed CSI and the tenth data obtained by processing the ninth data using the fourth model; The fifth SGCS is the SGCS between the reconstructed CSI and the tenth data obtained by processing the ninth data using the fourth model.

11. The method according to claim 4, characterized in that, The dataset includes the fifth data, the seventh data, and the sixth data; Wherein, the fifth data, the seventh data, and the sixth data are used for model training of the third model, and the third model is used for model training of the fourth model to obtain the first model, the fourth model including the first model and the third model; in the model training of the third model, the input of the third model includes the fifth data and the seventh data, and the true value corresponding to the output of the third model includes the sixth data; the third model is used for model training of the fourth model including: in the model training of the fourth model, data #2 is used as the input of the first model and as the true value corresponding to the output of the third model, and the output of the first model is used as the input of the third model.

12. The method according to claim 11, characterized in that, The performance metrics include one or more of the following ranges that should be met: sixth MSE; sixth NMSE; sixth MAE; seventh MSE; seventh NMSE; seventh MAE; seventh GCS; seventh SGCS; or, a weighted sum or weighted average of multiples of the sixth MSE, sixth NMSE, sixth MAE, seventh MSE, seventh NMSE, seventh MAE, seventh GCS, or seventh SGCS. The sixth MSE is the MSE between the reconstructed CSI and the sixth data obtained by processing the fifth and seventh data using the third model; The sixth NMSE is the NMSE between the reconstructed CSI and the sixth data obtained by processing the fifth and seventh data using the third model; The sixth MAE is the MAE between the reconstructed CSI and the sixth data obtained by processing the fifth and seventh data using the third model; The seventh MSE is the MSE between the reconstructed CSI and the data #2 obtained by processing the data #2 using the fourth model; The seventh NMSE is the NMSE between the reconstructed CSI and the data #2 obtained by processing the data #2 using the fourth model; The seventh MAE is the MAE between the reconstructed CSI and the data #2 obtained by processing the data #2 using the fourth model; The seventh GCS is the GCS between the reconstructed CSI and the data #2 obtained by processing the data #2 using the fourth model; The seventh SGCS is the SGCS between the reconstructed CSI and the data #2 obtained by processing the data #2 using the fourth model.

13. The method according to any one of claims 6, 7, 10 or 12, characterized in that, The range that MSE should meet includes the range that the average or distribution of MSE should meet; and / or, The range that NMSE should satisfy includes the range that the average or distribution of NMSE should satisfy; and / or, The range that MAE should meet includes the range that the average or distribution of MAE should meet.

14. The method according to any one of claims 7, 10, or 12, characterized in that, The range that GCS should satisfy includes the range that the average or distribution of GCS should satisfy, and / or, The range that SGCS should satisfy includes the range that the average or distribution of SGCS should satisfy.

15. The method according to any one of claims 1 to 14, characterized in that, The performance metrics are related to the measurement configuration, which is used to configure a reference signal, and the first channel information is related to the measurement results of the reference signal.

16. The method according to any one of claims 1 to 15, characterized in that, The method further includes: If the performance metrics are met, the inference task of the first model is executed.

17. The method according to any one of claims 1 to 16, characterized in that, The method further includes: If the performance indicators are not met, a third message is sent, the third message being used for one or more of the following: This indicates that the performance metric is not met; Request that the dataset be replaced; Request a change to the aforementioned performance metric; or Request to shut down the training of the model.

18. The method according to any one of claims 1 to 17, characterized in that, The acquisition of the dataset includes: Receive the dataset.

19. A communication method, characterized in that, include: Send the dataset, which is used for model training to obtain a first model; The performance metrics of the first model correspond to the inference task of the first model. The first model is obtained by training the model based on the dataset and the performance metrics. The inference task of the first model is to perform a first process on the first data to obtain the second data. The first process includes compression, or the first process includes compression and quantization. The first data is used to characterize the first channel information; The second data is used to determine third data based on the second model. The third data is used to characterize second channel information. The second channel information includes at least one of the following: part or all of the first channel information, or third channel information, wherein the third channel information is different from the first channel information. The inference task of the second model includes processing the second data to obtain the third data.

20. The method according to claim 19, characterized in that, The method further includes: Send a first message, which is used to indicate the performance metrics of the first model.

21. A communication device, characterized in that, Includes modules or units for performing the method according to any one of claims 1 to 19.

Images