Method for determining channel state information

Abstract

The invention provides a method for determining channel state information, which comprises the following steps that: terminal equipment receives a reference signal and carries out channel estimation according to the reference signal; the terminal device determines a target CSI parameter based on the estimated channel through an AI model, and feeds back the target CSI parameter to the transmitting device; the sending end equipment adopts the target CSI parameter to carry out data sending; the terminal equipment receives the sent data, demodulates and decodes the sent data, and judges the correctness of decoding; the CSI parameter is calculated and determined by using the AI model, so that the operation complexity can be reduced and the transmission rate can be improved.

Description

A method for determining channel state information Technical Field

[0001] This application relates to the field of communications, and more particularly to a method for determining channel state information. Background Technology

[0002] In wireless communication systems, Channel Quality Indicator (CSI) measurements are typically performed based on the fading channel H obtained from channel estimation to determine key parameters such as out-of-rank (RI), precoding codebook precision (PMI), and channel quality indicator (CQI). These parameters are then fed back to the transmitter, which uses them to transmit data accordingly. Traditional methods for determining RI, PMI, and CQI through CSI measurements are complex, requiring numerous computational steps, which increases the complexity of product implementation and affects area and power consumption. Summary of the Invention

[0003] The following is an overview of the topics described in detail in this article.

[0004] The purpose of this application is to at least partially solve one of the technical problems existing in the related technologies. The embodiments of this application provide a method for determining channel state information, which can reduce computational complexity and improve transmission rate.

[0005] An embodiment of this application provides a method for determining channel state information, comprising: a terminal device receiving a reference signal and performing channel estimation based on the reference signal; the terminal device determining target CSI parameters based on the estimated channel using an AI model and feeding the target CSI parameters back to a transmitting device; the transmitting device using the target CSI parameters to transmit data; and the terminal device receiving the transmitted data, demodulating and decoding the transmitted data, and determining the correctness of the decoding.

[0006] According to certain embodiments of this application, the terminal device determines target CSI parameters based on an estimated channel using an AI model, including: the terminal device determining a first CSI parameter based on an estimated channel using an AI model; the terminal device determining a second CSI parameter based on an estimated channel using a standard CSI measurement method; and using the first CSI parameter and the second CSI parameter as target CSI parameters.

[0007] According to certain embodiments of this application, the terminal device determines the target CSI parameter based on the estimated channel using an AI model, including: when the signal-to-noise ratio (SNR) is greater than a preset SNR threshold, the terminal device determines the target CSI parameter based on the estimated channel using an AI model; when the SNR is less than or equal to the preset SNR threshold, the terminal device determines the target CSI parameter based on the estimated channel using a standard CSI measurement method.

[0008] According to certain embodiments of this application, the terminal device determines a target CSI parameter based on an estimated channel using an AI model, including: the terminal device determining a third CSI parameter based on an estimated channel using an AI model; the terminal device determining a fourth CSI parameter based on an estimated channel using a standard CSI measurement method; when the RI parameter of the third CSI parameter and the RI parameter of the fourth CSI parameter are the same, averaging the PMI parameter of the third CSI parameter and the PMI parameter of the fourth CSI parameter to obtain an average PMI parameter, averaging the CQI parameter of the third CSI parameter and the CQI parameter of the fourth CSI parameter to obtain an average CQI parameter, and using the RI parameter of the third CSI parameter, the average PMI parameter, and the average CQI parameter as the target CSI parameter; when the RI parameter of the third CSI parameter and the RI parameter of the fourth CSI parameter are different, selecting either the third CSI parameter or the fourth CSI parameter as the target CSI parameter.

[0009] According to certain embodiments of this application, training the AI ​​model includes: acquiring channel information and CSI parameter labels in multiple scenarios; inputting the channel information and the CSI parameter labels into the AI ​​model to be trained for training to obtain a trained AI model; configuring the communication system with a preferred model type and the terminal device with a configured capability level; selecting a target AI model from the trained AI model according to the preferred model type and / or capability level, and loading the target AI model onto the terminal device.

[0010] According to certain embodiments of this application, data preprocessing of the channel information includes: extracting feature parameters from the channel information as auxiliary input data; unifying the channel information in the subcarrier frequency domain dimension; converting the imaginary part of the channel information into a real number; and normalizing the channel information.

[0011] According to certain embodiments of this application, the AI ​​model includes a RIPMI joint mapping classification module, a RICQI joint mapping classification module, and an RI retraining module; the RIPMI joint mapping classification module maps the channel information to a RIPMI label composed of the RI parameter and PMI parameter in the CSI parameter label, and outputs the classification probability value of the RIPMI label; the RICQI joint mapping classification module maps the channel information to a RICQI label composed of the RI parameter and CQI parameter in the CSI parameter label, and outputs the classification probability value of the RICQI label; the RI retraining module determines the values ​​of the RI parameter, PMI parameter, and CQI parameter based on the classification probability values ​​of the RIPMI label and the RICQI label.

[0012] According to certain embodiments of this application, the RI retraining module determines the values ​​of RI parameters, PMI parameters, and CQI parameters based on the classification probability values ​​of the RIPMI label and the RICQI label, including: concatenating the classification probability values ​​of the RIPMI label and the RICQI label into a fusion vector; training a multi-label loss function based on the fusion vector, and outputting the index value of the parameter set composed of the RI parameters, PMI parameters, and CQI parameters with the largest classification probability value; and determining the values ​​of the RI parameters, PMI parameters, and CQI parameters based on the index value of the parameter set composed of the RI parameters, PMI parameters, and CQI parameters with the largest classification probability value.

[0013] According to certain embodiments of this application, the RI retraining module determines the values ​​of RI parameters, PMI parameters, and CQI parameters based on the classification probability values ​​of the RIPMI label and the RICQI label, including: demapping the RIPMI label classification probability value to obtain a first classification probability value of the RI parameter; demapping the RICQI label classification probability value to obtain a second classification probability value of the RI parameter; training based on the first classification probability value of the RI parameter, the second classification probability value of the RI parameter, and the correct RI parameter to determine the value of the RI parameter; determining the value of the PMI parameter corresponding to the value of the RI parameter based on the value of the RI parameter and the classification probability value of the RIPMI label; and determining the value of the CQI parameter corresponding to the value of the RI parameter based on the value of the RI parameter and the classification probability value of the RICQI label.

[0014] According to certain embodiments of this application, the AI ​​model is updated in the following manner: setting an update trigger mechanism for the AI ​​model; when the update trigger mechanism is met, using supervised learning to perform a lightweight update of the AI ​​model based on the existing model, or using reinforcement learning to learn a strategy to adjust the CSI parameters based on the rewards corresponding to the outputs of the existing model and the AI ​​model to update the AI ​​model.

[0015] The above scheme has at least the following beneficial effects: the terminal device receives a reference signal and performs channel estimation based on the reference signal; the terminal device determines the target CSI parameters based on the estimated channel using an AI model and feeds the target CSI parameters back to the transmitting device; the transmitting device uses the target CSI parameters to transmit data; the terminal device receives the transmitted data, demodulates and decodes the transmitted data, and determines the correctness of the decoding; by using an AI model to calculate and determine the CSI parameters, the computational complexity can be reduced and the transmission rate can be improved. Attached Figure Description

[0016] The accompanying drawings are used to provide a further understanding of the technical solutions of this application and constitute a part of the specification. They are used together with the embodiments of this application to explain the technical solutions of this application and do not constitute a limitation on the technical solutions of this application.

[0017] Figure 1 is a structural diagram of the communication system for CSI measurement and feedback; Figure 2 is a step diagram of the method for determining channel state information; Figure 3 is a step diagram of one case where the terminal device determines the target CSI parameters based on the estimated channel; Figure 4 is a step diagram of another case where the terminal device determines the target CSI parameters based on the estimated channel; Figure 5 is a step diagram of yet another case where the terminal device determines the target CSI parameters based on the estimated channel; Figure 6 is a step diagram of one case where the RI retraining module determines the values ​​of RI parameters, PMI parameters, and CQI parameters; Figure 7 is a step diagram of another case where the RI retraining module determines the values ​​of RI parameters, PMI parameters, and CQI parameters; Figure 8 is a structural diagram of the AI ​​model. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, or the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0020] In wireless communication systems, Channel Quality Indicator (CSI) measurements are typically performed based on the fading channel H obtained from channel estimation to determine key parameters such as out-of-rank (RI), precoding codebook (PMI), and channel quality indicator (CQI). These parameters are then fed back to the transmitter, which uses the fed-back CSI parameters to transmit data accordingly. Therefore, CSI measurement and feedback help ensure the rational scheduling and resource allocation of the communication system. Referring to Figure 1, which is a structural diagram of a communication system with CSI measurement and feedback, when CSI measurements indicate good channel quality, a higher number of transport streams (corresponding to a larger RI), a higher modulation order, and a higher code rate (corresponding to a larger CQI) can be used for high-speed transmission of multiple data streams.

[0021] When CSI measurements indicate poor channel quality, a lower number of transport streams (corresponding to a smaller RI), a lower modulation order, and a lower code rate (corresponding to a smaller CQI) can be used to ensure robust data transmission.

[0022] In CSI measurements, when the codebook direction selected for signal transmission can accurately point to the receiver or avoid interference obstacles (i.e., selecting an accurate precoded codebook PMI), more signal energy is ensured to be received, thereby improving the receiver's reception capability and the system's communication performance.

[0023] Therefore, the accurate selection of CSI parameters is crucial for improving the communication performance of the system.

[0024] The traditional measurement method (standard CSI measurement method) is as follows.

[0025] The value of RI represents the number of data streams that a communication system can transmit in parallel, and it is related to the number of antennas of the transmitter and receiver in the communication system: the maximum value of RI = min{number of transmitter antennas, number of receiver antennas}.

[0026] For example, if the number of transmitting antennas supported by the communication system is {1,2,4}, and the number of receiving antennas supported is {1,2,4}.

[0027] The candidate RI values ​​are shown in Table 1, which is a table of candidate RI values ​​for different transmit and receive antennas.

[0028] Table 1. Candidate RI values ​​for different numbers of transmit and receive antennas.

[0029] The more PMI precoding codebook index values ​​there are, the more the signal quality can be enhanced, interference avoided, and transmission distance and speed increased by adjusting the amplitude and phase of the signal on the transmit antenna and concentrating the energy of the synthesized signal beam in a specific beam direction. Candidate PMI values ​​are shown in Table 2, which lists candidate PMI values ​​for different transmit and receive antennas.

[0030] Table 2. Candidate PMI values ​​for different transmit and receive antennas.

[0031] Different CQI values ​​correspond to different modulation orders and coding rates. Candidate PMI values ​​are shown in Table 3, which corresponds to different CQI index values ​​and modulation orders and coding rates.

[0032] Table 3. Correspondence between different CQI index values ​​and modulation order and code rate

[0033] Therefore, by measuring and feeding back CSI data, appropriate transmission parameters can be determined according to the actual fading of the channel, ensuring stable and efficient data transmission in the communication system.

[0034] Traditional methods for calculating CSI parameters determine (RI, PMI) and CQI independently as two separate parts. This is a suboptimal method considering complexity, and even so, its complexity is still considerable. For example, in a 4T4R system, the computational complexity can reach 5M FLOPs to 100M FLOPs under different bandwidths. If all combinations of (RI, PMI, CQI) are iterated to select the optimal value, the complexity increases exponentially, by more than ten times. Therefore, traditional methods for calculating CSI parameters suffer from high complexity and suboptimal performance.

[0035] Traditional CSI measurements for determining RI, PMI, and CQI are complex, requiring numerous computational steps, which increases product implementation complexity, area, and power consumption. Furthermore, commonly used traditional CSI measurement methods for determining RI, PMI, and CQI are usually not optimal CSI parameters, as theoretically optimal CSI parameters require extremely complex calculations that are difficult to implement in products.

[0036] To address the above problems, embodiments of this application provide a method for determining channel state information.

[0037] The embodiments of this application will be further described below with reference to the accompanying drawings.

[0038] Referring to Figure 2, the method for determining channel state information includes the following steps: Step S110, the terminal device receives a reference signal and performs channel estimation based on the reference signal; Step S120, the terminal device determines the target CSI parameters based on the estimated channel using an AI model and feeds back the target CSI parameters to the transmitting device; Step S130, the transmitting device transmits data using the target CSI parameters; Step S140, the terminal device receives the transmitted data, demodulates and decodes the transmitted data, and determines the correctness of the decoding.

[0039] It's important to note that the selection and use of AI models must match the terminal's capabilities or the scenario's requirements. The communication system configures preferred model types, and the terminal device configures its capability level. Based on the preferred model type and / or capability level, a target AI model is selected from the trained AI models, and the target AI model is loaded onto the terminal device. The communication system indicates / configures the preferred model type, such as a robust model or an aggressive model; alternatively, the terminal reports its own capability level, such as low capability / low cost or high performance. Different models are loaded onto the terminal according to the above model preferences or terminal capability instructions / configurations.

[0040] Specifically, different models have different characteristics and need to be selected according to the capabilities of the terminal and the requirements of the scenario. When deploying AI models, it is necessary to consider the capabilities, requirements, and actual environment of the terminal products for deployment and updates.

[0041] For example, if the terminal is a low-capacity terminal and prefers low-complexity models, then only the low-complexity model B can be selected to be deployed to the terminal for loading model parameters; if the terminal is a high-capacity terminal and prioritizes performance, then the high-complexity model A can be selected to be deployed to the terminal for loading model parameters.

[0042] In communication system scenarios requiring higher performance transmission and favoring more aggressive CSI parameters, the corresponding model D is selected for deployment, and the model parameters are loaded. In communication system scenarios requiring more stable and reliable link transmission and favoring more conservative CSI parameters, the corresponding model C is selected for deployment, and the model parameters are loaded.

[0043] The entity being loaded is either the LCM, the base station, or the terminal itself. If it is the LCM or the base station, the LCM or the base station needs to transmit the appropriate model parameters, results, and values ​​to the terminal through the interface; if the terminal loads and selects the model itself, the terminal needs to retain multiple sets of models.

[0044] Using AI models to calculate and determine CSI is intended for use in certain scenarios / conditions or under all conditions. Therefore, AI models can be used either as a substitute or as an auxiliary tool.

[0045] When using AI models in a replacement manner, in all cases, AI models are used to calculate and determine CSI parameters (RI, PMI, CQI), instead of traditional methods. For example, it has been thoroughly demonstrated that AI models can provide better communication performance in all scenarios.

[0046] When using AI models, their application is supplementary. AI model calculations to determine CSI parameters (RI, PMI, CQI) serve only as a complement to traditional methods. For example, when traditional methods cannot meet performance requirements, or when AI models can achieve better communication performance in certain scenarios / conditions. Possible scenarios / conditions include: when the block error rate or packet loss rate of the communication system is too high using traditional methods, a suitable AI model can be switched to calculate CSI; or, simulations show that under medium-to-high SNR conditions, CSI parameters obtained based on AI models can achieve significantly higher transmission rates compared to traditional methods, therefore, AI models can be used to calculate CSI parameters under medium-to-high SNR conditions.

[0047] Referring to Figure 3, in one scenario, the terminal device can report two sets of CSI parameters (RI, PMI, CQI) for the transmitting end to reference. Specifically, the terminal device determines the target CSI parameters based on the estimated channel using an AI model, including the following steps: Step S211, the terminal device determines the first CSI parameter based on the estimated channel using an AI model; Step S212, the terminal device determines the second CSI parameter based on the estimated channel using a standard CSI measurement method; Step S213, the first CSI parameter and the second CSI parameter are used as the target CSI parameters.

[0048] Referring to Figure 4, in another scenario, the terminal device calculates CSI parameters using an AI model when the first condition is met, and calculates CSI parameters using a traditional method (i.e., the standard CSI measurement method) when the first condition is not met, and reports a set of CSI parameters (RI, PMI, CQI). The first condition can be: when the SNR is greater than or equal to the signal-to-noise ratio threshold. Specifically, the terminal device determines the target CSI parameters based on the estimated channel using an AI model, including the following steps: Step S221, when the signal-to-noise ratio is greater than a preset signal-to-noise ratio threshold, the terminal device determines the target CSI parameters based on the estimated channel using an AI model; Step S222, when the signal-to-noise ratio is less than or equal to the preset signal-to-noise ratio threshold, the terminal device determines the target CSI parameters based on the estimated channel using the standard CSI measurement method.

[0049] Referring to Figure 5, in another scenario, the terminal device simultaneously calculates two sets of CSI using both an AI model and a traditional method. After a first processing step, a single set of reported CSI parameter values ​​is obtained. This first processing step can be: when the RI values ​​of the two sets of CSI parameters are the same, the PMI and CQI are averaged, rounded up or down. When the RI values ​​of the two sets of CSI parameters are different, the CSI parameter corresponding to the smaller RI value (RI, PMI, CQI) is selected; or the CSI parameter corresponding to the larger RI value (RI, PMI, CQI) is selected; or other methods are used. Specifically, the terminal device determines the target CSI parameter based on the estimated channel using an AI model, including the following steps: Step S231, the terminal device determines the third CSI parameter based on the estimated channel using an AI model; Step S232, the terminal device determines the fourth CSI parameter based on the estimated channel using a standard CSI measurement method; Step S233, when the RI parameter of the third CSI parameter and the RI parameter of the fourth CSI parameter are the same, the PMI parameter of the third CSI parameter and the PMI parameter of the fourth CSI parameter are averaged to obtain an average PMI parameter, the CQI parameter of the third CSI parameter and the CQI parameter of the fourth CSI parameter are averaged to obtain an average CQI parameter, and the RI parameter of the third CSI parameter, the average PMI parameter, and the average CQI parameter are used as the target CSI parameter; Step S234, when the RI parameter of the third CSI parameter and the RI parameter of the fourth CSI parameter are different, either the third CSI parameter or the fourth CSI parameter is selected as the target CSI parameter.

[0050] To ensure that the AI ​​model can infer and determine appropriate CSI parameters, it is necessary to train and test the AI ​​model. The main purpose is to use the AI ​​model to learn the accurate mapping relationship between sample data (e.g., channel estimation H) and sample label CSI parameters (e.g., RI, PMI, CQI).

[0051] Comprehensive data collection. To ensure the applicability and generalizability of AI models, the datasets used for training AI models need to have the following characteristics: they should cover common scenarios and configurations.

[0052] The scenario configuration and collection method for sample data collection are shown in Table 4. Table 4 is a table of scenario configuration and collection method for sample data collection.

[0053] Table 4. Scenarios and Collection Methods for Sample Data Acquisition

[0054] Data collected under different configurations needs to be adapted to a unified AI model for training. Therefore, format fusion is required to determine a unified input parameter, format, and value range.

[0055] For example, feature parameters can be extracted based on channel information as auxiliary input data. For instance, besides directly using H as input to the AI ​​model, feature parameters can also be manually extracted based on H, using known theoretical knowledge, as auxiliary input for AI model training to help the AI ​​model better learn the mapping relationship between data and labels. This includes: estimated channel quality parameter SNR, feature values ​​calculated based on H, etc.

[0056] Channel information needs to be standardized across the subcarrier frequency domain. For example, the fading channel H obtained from pilot estimation differs across subcarrier dimensions under different bandwidth configurations. Therefore, the input data H from different subcarrier dimensions needs to be formatted uniformly to obtain a fixed-dimensional input in the subcarrier frequency domain. For instance, the subcarrier dimension can be K. When K > 1, H under different bandwidths needs to be processed along the subcarrier dimension (subcarrier selection or expansion) until it becomes dimension K. Alternatively, K = 1, which represents obtaining the wideband averaged H.

[0057] The imaginary part of the channel information is converted to a real number. For example, the fading channel H based on pilot estimation is a complex signal containing amplitude and phase information, while AI models can usually only process real data. Therefore, the input data H needs to be converted from a complex number to a real number. For example, various conversion forms such as the real and imaginary parts of H, and the amplitude and phase of H.

[0058] The channel information is normalized. For example, a reasonable range of input data values ​​is essential for maintaining a reasonable gradient range and gradient propagation during AI model training. Therefore, H can be scaled, for example, by normalization or multiplying by a uniform scaling factor.

[0059] In addition, to improve the training performance of the model, the following label processing and data processing can be performed.

[0060] Joint mapping of labels. For multi-label classification problems with CSI parameters (RI, PMI, CQI), considering that the values ​​of PMI and RI are strongly correlated, and the values ​​of CQI and RI are also strongly correlated, RI and PMI can be jointly mapped to a new label RIPMI, and RI and CQI can be jointly mapped to a new label RICQI, thus reducing the number of labels to be classified.

[0061] Optimal selection of data samples. If multiple sets of data and label samples can be obtained through various algorithms, the performance of different algorithms can be compared, and the data and label samples with better performance can be selected and retained to improve the upper limit of the AI ​​model's capabilities. For example, fully joint or semi-joint algorithms can be used to select the best data samples.

[0062] Abnormal data sample deletion. When abnormal data is found in a data sample, it needs to be deleted. For example, if the communication performance of a certain data sample deteriorates when channel conditions are better, then that data sample is unreliable; if the receiver cannot correctly demodulate and decode the link transmission corresponding to a certain data sample, then that data sample is considered unreliable. Unreliable abnormal data needs to be deleted from the dataset.

[0063] If the throughput decreases when the SNR increases, then the abnormal data is deleted. NACK data that failed to decode is deleted, while ACK data samples are retained.

[0064] Different post-processing methods for sample data also include the following.

[0065] For the signal-to-noise ratio (SNR), the SNR value is estimated based on the channel estimation H. The SNR value is the ratio of the signal power to the noise power in H, which can reflect the channel transmission quality to a certain extent and is usually related to the tag CSI parameter.

[0066] For eigenvalues If the feature values ​​are relatively uniform, it indicates the absence of a dominant component, suggesting potential multi-stream data transmission; the corresponding label RI value is generally large. Conversely, if some feature values ​​are significantly larger than others, it indicates the presence of a dominant component; the corresponding RI value is generally small, and data transmission is primarily concentrated along the eigenvector directions corresponding to specific feature values. There are two specific calculation methods: The broadband feature value is: The sub-band eigenvalues ​​are: .

[0067] Multi-bandwidth H-format fusion. The format of the fading channel H based on pilot estimation is as follows: Under 5 / 10 / 20 / 40 / 80 MHz bandwidth configurations, the corresponding The values ​​are 20 / 40 / 80 / 160 / 320. Now, the following method is needed to unify the data samples H under different bandwidths to [value missing]. Take a value, for example When high bandwidth Subcarrier signal extraction is performed at >20: (Result) satisfy and will By concatenating along the TTI dimension, we obtain... Accordingly, labels and other data need to be repeated in the sample dimension. times, satisfy It is an integer. When bandwidth is small... When H is less than 20, it needs to be repeated along the subcarrier dimension, becoming... The format.

[0068] To determine the number of channels in H, we need to extract the real and imaginary parts of H and convert them into independent dimensions to obtain the dimension of H. Here, 2 represents the two parts real(H) and imag(H). The magnitude abs(H) and phase phase(H) of H are extracted and transformed into independent dimensions to obtain the dimension of H. Alternatively, combine the two numbers to obtain the dimension of H. .

[0069] To normalize H, divide H by the quantization bit width, we have: .or, , This is the maximum value of perSNR. Or, , It is the maximum value of all SNRs.

[0070] For the 62 categories of the label (RI, PMI), refer to Table 5. Table 5 is a 62-category mapping table for the RIPMI label.

[0071] Table 5. 62-category mapping table of RIPMI labels

[0072] For the 96 categories of labels (RI, CQI), refer to Table 6. Table 6 is a mapping table of the 96 categories of RICQI labels.

[0073] Table 6. RICQI Tag 96-Category Mapping Table

[0074] Additional input parameters help the model learn the mapping relationship between data and labels more effectively; standardized data format and values ​​enable training of consistent AI models, reducing their complexity; label mapping and transformation reduce the number of labels to be classified, simplifying the model's learning requirements; data cleaning and optimization improve the quality of training data, contributing to better AI model performance. Comprehensive data collection and post-processing enhance the quality of input data for AI models, leading to improved training performance.

[0075] Based on the training set and model structure, AI models with different complexities and parameters can be designed. For example, Model A has higher complexity (more convolutional layers, more fully connected kernels, etc.) and better overall classification accuracy; Model B has lower complexity (fewer convolutional layers, fewer fully connected kernels, etc.) and moderate overall classification accuracy; Model C uses a different structure, making the model's output CSI parameters more conservative; Model D uses a different structure, making the model's output CSI parameters more aggressive. It is understandable that Models A, B, C, and D all use the same AI model architecture.

[0076] By fully utilizing the structure of AI models to extract data features, accurate classification of labels on a sample set can be achieved.

[0077] Referring to Figure 8, the AI ​​model includes a RIPMI joint mapping classification module, a RICQI joint mapping classification module, and an RI retraining module.

[0078] For the RIPMI joint mapping classification module, the inputs are the estimated fading channel H, the estimated signal-to-noise ratio (SNR) (optional), and the calculated eigenvalue λ (optional). The RIPMI joint mapping classification module maps the channel information to a RIPMI label jointly generated by the RI and PMI parameters in the CSI parameter label, and outputs the classification probability value of the RIPMI label, i.e., the likelihood or probability value of the (RI, PMI) label for 62 categories. The labels trained by the module are RI and PMI.

[0079] For the RICQI joint mapping classification module, the inputs are the estimated fading channel H, the estimated signal-to-noise ratio (SNR) (optional), and the calculated eigenvalue λ (optional). The RICQI joint mapping classification module maps the channel information to a RICQI label jointly generated by the RI and CQI parameters in the CSI parameter label, and outputs the classification probability value of the RICQI label, i.e., the likelihood or probability value of the (RI, CQI) label for 96 classifications. The labels trained by the module are RI and CQI.

[0080] The RI retraining module determines the values ​​of the RI, PMI, and CQI parameters based on the classification probability values ​​of the RIPMI and RICQI labels. Since the outputs of the RIPMI and RICQI joint mapping classification modules have an intersection term RI, this output is used as input to the RI retraining module. Based on the known labels, the model is then ensembled and trained again to obtain a more accurate RI.

[0081] Referring to Figure 6, in one scenario, the RI retraining module determines the values ​​of the RI parameters, PMI parameters, and CQI parameters based on the classification probability values ​​of the RIPMI and RICQI labels, including the following steps: Step S311, concatenating the classification probability values ​​of the RIPMI and RICQI labels into a fusion vector; Step S312, training a multi-label loss function based on the fusion vector, and outputting the index value of the parameter set consisting of the RI, PMI, and CQI parameters with the largest classification probability value; Step S313, determining the values ​​of the RI, PMI, and CQI parameters based on the index value of the parameter set consisting of the RI, PMI, and CQI parameters with the largest classification probability value.

[0082] Referring to Figure 7, in another scenario, the RI retraining module determines the values ​​of the RI parameter, PMI parameter, and CQI parameter based on the classification probability values ​​of the RIPMI label and the RICQI label, including the following steps: Step S321, demapping the RIPMI label classification probability value to obtain the first classification probability value of the RI parameter; Step S322, demapping the RICQI label classification probability value to obtain the second classification probability value of the RI parameter; Step S323, training based on the first classification probability value, the second classification probability value, and the correct RI parameter to determine the value of the RI parameter; Step S324, determining the value of the PMI parameter corresponding to the value of the RI parameter based on the value of the RI parameter and the classification probability value of the RIPMI label, and determining the value of the CQI parameter corresponding to the value of the RI parameter based on the value of the RI parameter and the classification probability value of the RICQI label.

[0083] Specifically, the input to the RI retraining module can be: directly concatenating the outputs of the RIPMI joint mapping classification module and the RICQI joint mapping classification module into a vector of length 62+96 as input; or it can be: demapping the output of the RIPMI joint mapping classification module to obtain the RI 4-class logical values, demapping the output of the RICQI joint mapping classification module to obtain the RI 4-class logical values, and using the above two RI 4-class logical values, the output of the RIPMI joint mapping classification module (optional), and the output of the RICQI joint mapping classification module (optional) as input.

[0084] The implementation of the RI retraining module for training labels can be as follows: Based on the outputs of the RIPMI joint mapping classification module and the RICQI joint mapping classification module, a multi-label loss function is trained, directly outputting the (RI, PMI, CQI) index value with the highest probability. Alternatively, training can be performed based on the 4-class logistic values / likelihood probability values ​​of the two input RIs and the correct RI label, outputting only the most probable RI value. Then, based on this RI value, the PMI value with the highest probability corresponding to that RI is found from the output of the RIPMI joint mapping classification module, and the CQI value with the highest probability corresponding to that RI is found from the output of the RICQI joint mapping classification module.

[0085] This AI model decouples different labels; for example, PMI and CQI have little correlation and can be learned independently. The same label can be learned in different models, and then further improved through subsequent model ensemble training to enhance the classification accuracy of the publicly available label. For instance, RI is learned in both the RIPMI joint mapping classification module and the RIPMI joint mapping classification module. When the maximum probability RI values ​​output by the two models differ, the correct label RI and the RI retraining module can be used for further learning and correction. By jointly processing labels and mapping them to new classification labels, the multi-label problem is transformed into a single-label problem, which helps to uncover the relationships between labels and simplifies the model. For example, RI and PMI are jointly mapped to RIPMI, and RI and CQI are jointly mapped to RICQI.

[0086] When deploying and calculating CSI parameters using a pre-trained AI model, the distribution characteristics of the pre-trained data may not match the current environment, resulting in the CSI calculated by the AI ​​model not being a perfect match for the current fading environment. In such cases, a mechanism-based model micro-training and update can be performed.

[0087] The AI ​​model is updated in the following ways: an update trigger mechanism is set for the AI ​​model; when the update trigger mechanism is met, supervised learning is used to perform a lightweight update of the AI ​​model based on the existing model, or reinforcement learning is used to learn and adjust the CSI parameters based on the rewards corresponding to the outputs of the existing model and the AI ​​model to update the AI ​​model.

[0088] The update trigger mechanism can be periodic or non-periodic.

[0089] The model is fine-tuned and updated periodically, for example, periodically updated after K ms; K can be a configured value, or K can be N times the period of the pilot reference signal.

[0090] Non-periodic model fine-tuning and updates are performed. For example, if the CSI parameters determined by the AI ​​model are found to cause an excessively high block error rate or packet loss rate in the communication system, then micro-training and updates of the model are triggered.

[0091] The terminal collects H, SNR (optional), eigenvalues ​​(optional), corresponding high-quality CSI parameters, and corresponding sample ACK / NACK results (optional) based on pilot signals over a period of time, and updates the model weight coefficients based on the existing model.

[0092] The model weights can be updated using supervised learning, which iteratively trains based on the input data and labels; or using reinforcement learning, which fine-tunes the CSI parameters by updating the policy based on the reward values ​​corresponding to ACK / NACK after multiple determinations of the CSI parameters. The goal is to maximize the reward, i.e., to expect more ACKs.

[0093] It can select and deploy appropriate models based on terminal capabilities and scenario requirements, ensuring both complexity and performance metrics. Through model updates, it can update pre-trained models with simple, low-complexity training (few repetitions) to create models more suited to real-world environments. This ensures the model's characteristics are adapted to terminal capabilities, scenario requirements, and the actual environment, achieving greater improvements in communication performance at a lower complexity cost.

[0094] The above is a detailed description of the preferred embodiments of this application, but this application is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of this application, and these equivalent modifications or substitutions are all included within the scope defined by the claims of this application.

Claims

1. A method for determining channel state information, characterized in that, include: The terminal device receives a reference signal and performs channel estimation based on the reference signal; The terminal device determines the target CSI parameters based on the estimated channel using an AI model, and feeds back the target CSI parameters to the transmitting device. The transmitting device uses the target CSI parameters to send data; the terminal device receives the transmitted data, demodulates and decodes the transmitted data, and determines the correctness of the decoding.

2. The method for determining channel state information according to claim 1, characterized in that, The terminal device determines the target CSI parameter based on the estimated channel using an AI model, including: the terminal device determining a first CSI parameter based on the estimated channel using an AI model; the terminal device determining a second CSI parameter based on the estimated channel using a standard CSI measurement method; and using the first CSI parameter and the second CSI parameter as the target CSI parameter.

3. The method for determining channel state information according to claim 1, characterized in that, The terminal device determines the target CSI parameters based on the estimated channel using an AI model, including: when the signal-to-noise ratio (SNR) is greater than a preset SNR threshold, the terminal device determines the target CSI parameters based on the estimated channel using an AI model; when the SNR is less than or equal to the preset SNR threshold, the terminal device determines the target CSI parameters based on the estimated channel using a standard CSI measurement method.

4. The method for determining channel state information according to claim 1, characterized in that, The terminal device determines the target CSI parameter based on the estimated channel using an AI model, including: the terminal device determining a third CSI parameter based on the estimated channel using an AI model; the terminal device determining a fourth CSI parameter based on the estimated channel using a standard CSI measurement method; when the RI parameter of the third CSI parameter and the RI parameter of the fourth CSI parameter are the same, the PMI parameter of the third CSI parameter and the PMI parameter of the fourth CSI parameter are averaged to obtain an average PMI parameter, the CQI parameter of the third CSI parameter and the CQI parameter of the fourth CSI parameter are averaged to obtain an average CQI parameter, and the RI parameter of the third CSI parameter, the average PMI parameter, and the average CQI parameter are used as the target CSI parameter; when the RI parameter of the third CSI parameter and the RI parameter of the fourth CSI parameter are different, either the third CSI parameter or the fourth CSI parameter is selected as the target CSI parameter.

5. The method for determining channel state information according to claim 1, characterized in that, Training the AI ​​model includes: acquiring channel information and CSI parameter labels in multiple scenarios; inputting the channel information and CSI parameter labels into the AI ​​model to be trained for training to obtain a trained AI model; configuring the communication system with a preferred model type and the terminal device with a configured capability level; selecting a target AI model from the trained AI model according to the preferred model type and / or capability level, and loading the target AI model onto the terminal device.

6. The method for determining channel state information according to claim 1, characterized in that, The channel information is preprocessed, including: extracting feature parameters from the channel information as auxiliary input data; unifying the channel information in the subcarrier frequency domain dimension; converting the imaginary part of the channel information to a real number; and normalizing the channel information.

7. The method for determining channel state information according to claim 1, characterized in that, The AI ​​model includes a RIPMI joint mapping classification module, a RICQI joint mapping classification module, and an RI retraining module. The RIPMI joint mapping classification module maps the channel information to a RIPMI label composed of the RI parameter and PMI parameter in the CSI parameter label, and outputs the classification probability value of the RIPMI label. The RICQI joint mapping classification module maps the channel information to a RICQI label composed of the RI parameter and CQI parameter in the CSI parameter label, and outputs the classification probability value of the RICQI label. The RI retraining module determines the values ​​of the RI parameter, PMI parameter, and CQI parameter based on the classification probability values ​​of the RIPMI label and the RICQI label.

8. The method for determining channel state information according to claim 7, characterized in that, The RI retraining module determines the values ​​of the RI parameters, PMI parameters, and CQI parameters based on the classification probability values ​​of the RIPMI and RICQI labels, including: concatenating the classification probability values ​​of the RIPMI and RICQI labels into a fusion vector; training a multi-label loss function based on the fusion vector, and outputting the index value of the parameter set consisting of the RI, PMI, and CQI parameters with the largest classification probability value; and determining the values ​​of the RI, PMI, and CQI parameters based on the index value of the parameter set consisting of the RI, PMI, and CQI parameters with the largest classification probability value.

9. The method for determining channel state information according to claim 7, characterized in that, The RI retraining module determines the values ​​of RI parameters, PMI parameters, and CQI parameters based on the classification probability values ​​of the RIPMI label and the RICQI label, including: demapping the RIPMI label classification probability value to obtain a first classification probability value of the RI parameter; demapping the RICQI label classification probability value to obtain a second classification probability value of the RI parameter; training based on the first classification probability value of the RI parameter, the second classification probability value of the RI parameter, and the correct RI parameter to determine the value of the RI parameter; determining the value of the PMI parameter corresponding to the value of the RI parameter based on the value of the RI parameter and the classification probability value of the RIPMI label; and determining the value of the CQI parameter corresponding to the value of the RI parameter based on the value of the RI parameter and the classification probability value of the RICQI label.

10. The method for determining channel state information according to claim 1, characterized in that, The AI ​​model is updated in the following manner: an update trigger mechanism for the AI ​​model is set; when the update trigger mechanism is met, supervised learning is used to perform a lightweight update of the AI ​​model based on the existing model, or reinforcement learning is used to learn and adjust the CSI parameters based on the rewards corresponding to the outputs of the existing model and the AI ​​model to update the AI ​​model.

Images