Receiver model training method, electronic equipment and storage medium

Abstract

The embodiment of the invention provides a receiver model training method, electronic equipment and a storage medium, which are applied to the technical field of wireless communication, and the method comprises the following steps: obtaining a training starting request transmitted by a receiver, and training a receiver model; and receiving a training stop request transmitted by the receiver, and ending the training of the receiver model. According to the embodiment of the invention, the receiver model is finely adjusted in the model application environment, so that the adaptation degree of the transceiver and the agent receiving model can be improved, the performance of the transceiver can be improved, and the wireless communication quality is improved.

Description

A receiver model training method, electronic device and storage medium Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to a receiver model training method, an electronic device, and a storage medium. Background Technology

[0002] An end-to-end Artificial Intelligence (AI) transceiver replaces the physical layer transceiver process in existing communication systems by using deep learning networks. In the AI ​​transceiver, a deep learning neural network replaces the channel estimation, channel equalization, and demodulation processes in the downlink process. The input to the AI ​​receiver model is the frequency domain signal received by the receiver, and the output is the log-likelihood ratio (LLR) in bit information form. The LLR is then fed into the decoder for decoding. The transceiver process implemented based on the AI ​​receiver model has better performance and generalization ability compared to the existing downlink process. The performance of the AI ​​receiver model directly affects the communication transceiver process, and the performance of the AI ​​receiver model largely depends on the richness of the dataset on which the model is built. Providing the model with sufficiently rich training data during the model pre-training stage can improve the generalization ability of the AI ​​receiver. However, due to limitations in data acquisition methods, the training data cannot possibly cover all application scenarios. Therefore, the online learning capability of the AI ​​receiver is particularly important. In some application scenarios where the training set data cannot cover them, the performance of the AI ​​receiver shows a significant decline compared to scenarios covered by the training set data. Similarly, when the matching degree between the receiver and transmitter is not high, the performance of the AI ​​receiver will also decrease significantly. In these application scenarios, the AI ​​receiver needs to quickly learn the current channel environment and fine-tune its model parameters to achieve better performance. How to fine-tune the training of the AI ​​receiver has become an urgent problem to be solved. Summary of the Invention

[0003] This application provides a receiver model training method, electronic device, and storage medium, which aim to enable fine-tuning of AI receiver parameters, improve the receiver's applicability to real-world application environments, enhance receiver performance, and improve communication quality.

[0004] This application provides a receiver model training method, wherein the method includes:

[0005] Obtain the training start request from the receiver and train the receiver model;

[0006] The training stop request transmitted by the receiver is received, and the training of the receiver model is terminated.

[0007] This application also provides a receiver model training method, wherein the method includes:

[0008] If the communication throughput is determined to be lower than the preset target, a training start request is sent to the transmitter.

[0009] Train the receiver model;

[0010] Once the performance metrics of the receiver model are determined to meet the preset requirements, a training stop request is sent to the transmitter, and the training of the receiver model is stopped.

[0011] This application also provides an electronic device, wherein the electronic device includes:

[0012] One or more processors;

[0013] Memory, used to store one or more programs;

[0014] When the one or more programs are executed by the one or more processors, the one or more processors implement the receiver model training method as described in any of the embodiments of this application.

[0015] This application also provides a computer-readable storage medium storing one or more programs that are executed by one or more processors to implement the receiver model training method as described in any of the embodiments of this application.

[0016] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 is a structural schematic diagram of an intelligent transceiver provided in an embodiment of this application;

[0019] Figure 2 is a flowchart of a receiver model training method provided in an embodiment of this application;

[0020] Figure 3 is a flowchart of another receiver model training method provided in an embodiment of this application;

[0021] Figure 4 is a flowchart of another receiver model training method provided in an embodiment of this application;

[0022] Figure 5 is a flowchart of another receiver model training method provided in an embodiment of this application;

[0023] Figure 6 is a flowchart of another receiver model training method provided in an embodiment of this application;

[0024] Figure 7 is a flowchart of another receiver model training method provided in an embodiment of this application;

[0025] Figure 8 is a flowchart of another receiver model training method provided in an embodiment of this application;

[0026] Figure 9 is a schematic diagram of another receiver model training architecture provided in an embodiment of this application;

[0027] Figure 10 is a signaling flowchart for another receiver online training provided in an embodiment of this application;

[0028] Figure 11 is another receiver online training signaling flowchart provided in an embodiment of this application;

[0029] Figure 12 is a schematic diagram of a receiver model training device provided in an embodiment of this application;

[0030] Figure 13 is a schematic diagram of another receiver model training device provided in an embodiment of this application;

[0031] Figure 14 is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0032] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of this application.

[0033] In the following description, the use of suffixes such as “module,” “part,” or “unit” to denote elements is solely for the purpose of illustration in this application and has no particular meaning in itself. Therefore, “module,” “part,” or “unit” may be used interchangeably.

[0034] The architecture of the end-to-end trained AI transceiver is shown in Figure 1. Both the transmitter and receiver are end-to-end trained neural network models. The transmitter sends symbol-level data, and the signal is sent to the receiver through the channel. The receiver receives the symbol-level data, normalizes it, and inputs it into the receiver neural network. The output is the demodulated bit information LLR. Finally, the LLR is sent to the decoder for the decoding process.

[0035] Figure 2 is a flowchart of a receiver model training method provided in an embodiment of this application. This embodiment of the application is applicable to the scenario of fine-tuning the model parameters of a receiver model of a smart transceiver. The method can be executed by a receiver model training device. As shown in Figure 2, the method provided in this embodiment of the application specifically includes the following steps:

[0036] Step 110: Obtain the training start request for receiver transmission and train the receiver model.

[0037] The training start request can be a trigger quality transmitted by the receiver to request training and fine-tuning of the receiver model. The training start request can be sent by the receiver, and the receiver model can be implemented by a neural network. The receiver model can process the received symbol-level data into demodulated bit information LLR.

[0038] In this embodiment of the application, the receiver can transmit a training start request to the transmitter. After receiving the training start request, the transmitter can train the receiver model together with the receiver. During the training process, the transmitter can provide the receiver with training sample data or training teacher data for the receiver model training.

[0039] Step 120: Receive the training stop request transmitted by the receiver and end the training of the receiver model.

[0040] The training stop request can be a message from the receiver to instruct the transmitter to stop training the receiver model, and the training stop request can be transmitted by the receiver.

[0041] In this embodiment of the application, the receiver can transmit a training stop request to the transmitter, and the transmitter can terminate the training of the receiver model when it receives the training stop request.

[0042] In this embodiment, the receiver model is trained by acquiring the training start request sent by the receiver, and the training of the receiver model is stopped when the training end request is acquired. This enables fine-tuning of the receiver model configured in the transceiver, which can improve the adaptability of the receiver model, enhance the information demodulation accuracy, and improve the communication quality of the system.

[0043] Figure 3 is a flowchart of another receiver model training method provided in an embodiment of this application. This embodiment is a specific modification based on the above embodiment. Referring to Figure 3, the method provided in this embodiment specifically includes the following steps:

[0044] Step 210: In response to the training start request transmitted by the receiver, send a training start consent instruction back to the receiver.

[0045] Among them, the training start consent instruction can be an instruction message from the transmitter to consent to the training of the receiver model, and the training start consent instruction can be transmitted from the transmitter to the receiver.

[0046] In this embodiment of the application, when the transmitter receives a training start request, it may send a training start consent instruction to the receiver.

[0047] Step 220: Start the online training subprocess and train the receiver model based on the online training subprocess.

[0048] The online training subprocess can be a subprocess used to train the receiver model. This subprocess can be configured within the transmitter and can have its own memory space. This subprocess can include the resources and state information required to train the receiver model.

[0049] In this embodiment of the application, the transmitter can send an enable consent instruction to the receiver and start the online training subprocess to train the receiver model.

[0050] Step 230: Receive the training stop request transmitted by the receiver and end the training of the receiver model.

[0051] In this embodiment, when the transmitter receives a training start request, it sends a training start consent instruction to the receiver and initiates an online training subprocess. This subprocess trains the receiver model, and training ends when the transmitter receives a training stop request from the receiver. This subprocess-based training of the receiver model allows the transceiver to perform parameter training on the receiver model while simultaneously maintaining data transmission and reception operations. This avoids interruptions to communication transmission due to parameter fine-tuning of the receiver model, thus improving receiver model performance and ensuring stable system communication.

[0052] Figure 4 is a flowchart of another receiver model training method provided in an embodiment of this application. This embodiment is a specific modification based on the above embodiment. Referring to Figure 4, the method provided in this embodiment specifically includes the following steps:

[0053] Step 310: Obtain the training start request transmitted by the receiver, and generate training frame data according to the training frame sequence.

[0054] The training frame sequence can be a data sequence used to generate training frames. The training frame sequence can include at least two training frame data, which can be pre-agreed bit information.

[0055] In this embodiment, when the transmitter receives a training start request transmitted by the receiver, it can acquire a pre-configured training frame sequence, which can be pre-configured in the transmitter. Training frame data for training the receiver model can be determined within the training frame sequence.

[0056] Step 320: Modulate the training frame data as frequency domain symbol data.

[0057] In this embodiment of the application, the determined training frame data can be encoded and mapped to frequency domain symbols through a specific modulation method to obtain frequency domain symbol data. The frequency domain symbol data can be information represented in the frequency domain after adjustment. The frequency domain symbol data can be converted from the time domain signal to the frequency domain through Fourier transform.

[0058] Step 330: Transmit the frequency domain symbol data to the receiver so that the receiver can train the receiver model.

[0059] In this embodiment, the transmitter can send the generated frequency domain symbol data to the receiver, so that the receiver can train the receiver model using the received frequency domain symbol data.

[0060] Step 340: Receive the training stop request transmitted by the receiver and end the training of the receiver model.

[0061] Furthermore, based on the above-described embodiments, generating training frame data according to the training frame sequence includes:

[0062] The training frame sequence number is determined based on the current frame sequence number, the current time slot sequence number, and the step size; the training frame data is then determined within the training frame sequence according to the training frame sequence number.

[0063] The current frame sequence number can be the sequence number of the data frame to be sent by the transmitter, the current timing sequence number can be the sequence number of the time slot to be sent by the transmitter, and the step size can be the data frame interval set by the transmitter for sending information.

[0064] In this embodiment of the application, the sequence frame number can be determined by the current frame sequence number, the current time slot sequence number and the step size, and training frame data with the same sequence number can be determined in the training frame sequence according to the sequence frame sequence number.

[0065] In one exemplary implementation, the sequence frame sequence number = current frame sequence number * step size + current time slot sequence number.

[0066] Furthermore, based on the above-mentioned application embodiments, transmitting frequency domain symbol data to the receiver includes: transmitting one frame of frequency domain symbol data for every data frame interval of a threshold number, wherein the threshold number is adjusted according to the learning efficiency of the receiver model.

[0067] In this embodiment of the application, since the frequency domain symbol data used to train the receiver model is transmitted between the transmitter and the receiver during the training process of the receiver model, it occupies part of the traffic of the communication system. In order to ensure normal communication efficiency, one frame of frequency domain symbol data can be transmitted every data frame at intervals of a threshold number. The value of the threshold number of data frames can be adjusted according to the learning efficiency of the receiver model. For example, when improving the learning efficiency of the receiver model, the value of the threshold number can be reduced.

[0068] In some embodiments, the method further includes: sending symbol-level data to the receiver.

[0069] Symbol-level data can be modulated and coded information, and it can be transmitted from the transmitter to the receiver.

[0070] In this embodiment, the transmitter can transmit symbol-level data to the receiver, thereby realizing information transmission between the transmitter and the receiver. The symbol-level data transmitted by the transmitter can be generated by service data modulation and coding. The transmitter can send the symbol-level data in the form of a process. The transmission process of the symbol-level data can be carried out simultaneously with the training process of the receiver model, thereby reducing the impact of the receiver model on the normal service of the communication system.

[0071] Figure 5 is a flowchart of another receiver model training method provided in an embodiment of this application. This embodiment of the application is applicable to the scenario of fine-tuning the model parameters of the receiver model of a smart transceiver. This method can be executed by a receiver model training device. As shown in Figure 5, the method provided in this embodiment of the application specifically includes the following steps:

[0072] Step 410: Determine that the communication performance is lower than the preset target, and send a training start request to the transmitter.

[0073] Communication performance refers to the receiver's operating performance, which may include throughput, signal-to-noise ratio, bandwidth, and latency. Throughput can be defined as the amount of data successfully transmitted by the communication system per unit time. Pre-configured metrics are critical performance values ​​used for fine-tuning the receiver model's parameters. These metrics can be configured based on business scenarios or customer needs.

[0074] In this embodiment, the receiver can monitor the communication throughput. When it is determined that the communication throughput is lower than a preset target, it can generate a training start request and transmit the training start request to the transmitter, requesting the transmitter and receiver to jointly train the receiver model.

[0075] Step 420: Train the receiver model.

[0076] In the embodiments of this application, the receiver model can be trained, and the network parameters or model weights of the receiver model's neural network can be adjusted.

[0077] Step 430: Determine that the performance indicators of the receiver model have met the preset requirements, send a training stop request to the transmitter, and stop the training of the receiver model.

[0078] Among them, the preset requirements can be the pre-configured performance index values ​​of the receiver model. These performance index values ​​can indicate the minimum performance values ​​for the receiver model to work normally within the receiver. The performance indexes can be the indicator parameters of the receiver model for processing data, and these performance indexes can include performance index values ​​such as throughput, bit error rate, signal-to-noise ratio, latency, and bandwidth.

[0079] In this embodiment of the application, the performance indicators of the trained receiver model can be statistically analyzed. When the performance indicators of the receiver model reach the preset requirements, it can be determined that the receiver model has completed training, and a training stop request can be transmitted to the transmitter, thereby causing the transmitter to end the training of the receiver model.

[0080] In this embodiment, when the communication performance is determined to be below a threshold, a training start request is sent to the transmitter to train the receiver model. Once the receiver model's performance meets preset requirements, a training stop request is sent to the transmitter, ending the training of the receiver model. This embodiment allows for fine-tuning of the receiver model configured within the transceiver, improving the receiver model's adaptability, enhancing demodulation accuracy, and improving the system's communication quality.

[0081] Figure 6 is a flowchart of another receiver model training method provided in an embodiment of this application. This embodiment is a specific modification based on the above-mentioned embodiments. Referring to Figure 6, the method provided in this embodiment specifically includes the following steps:

[0082] Step 510: Determine that the communication performance is lower than the preset target, and send a training start request to the transmitter.

[0083] Step 520: Receive the training start consent instruction from the transmitter.

[0084] In this embodiment of the application, when the transmitter agrees to train the receiver model, it can generate a training start consent instruction and feed the training start consent instruction back to the receiver, which can receive the training start consent instruction.

[0085] Step 530: Start the online training subprocess and train the receiver model based on the online training subprocess.

[0086] In this embodiment of the application, after receiving the training start consent instruction from the transmitter, the receiver can start the online training subprocess inside the receiver and train the receiver model through the online training subprocess, thereby fine-tuning the model parameters of the receiver model.

[0087] Step 540: Determine that the performance indicators of the receiver model have met the preset requirements, send a training stop request to the transmitter, and stop the training of the receiver model.

[0088] Furthermore, based on the above-mentioned application embodiments, the method further includes: receiving symbol-level data transmitted by the transmitter; starting an online inference subprocess, and generating a receiver output result of symbol-level data by calling a second receiver model through the online inference subprocess, wherein the second receiver model has the same model weight as the receiver model before training; and obtaining a channel estimation result by decoding the receiver output result according to the decoder.

[0089] In this embodiment, the transmitter can also transmit symbol-level data to the receiver. The receiver can call a second receiver model through an online inference subprocess. The second receiver model processes the received symbol-level data and generates the corresponding bit information (LLR) as the receiver output. It is understood that the second receiver model can be used within the receiver via subprocesses. This second receiver model can be the same as the receiver model before training, and its model weights are the same as those of the receiver model before training. The receiver can call a decoder to decode the receiver output from the second receiver model and perform channel estimation using the decoding result to obtain the channel estimation result.

[0090] Based on the above application embodiments, once the receiver model training is completed, the receiver model is replaced in the online inference subprocess, and the second receiver model is replaced in the online training subprocess.

[0091] In this embodiment of the application, after the online training subprocess completes the training of the receiver model, the receiver model configured in the online training subprocess and the second receiver model in the online inference subprocess can be interchanged. The second receiver model is trained by the online training subprocess, and the data is processed by the receiver model by the online inference subprocess, thereby reducing the impact of model parameter fine-tuning on the communication system.

[0092] Figure 7 is a flowchart of another receiver model training method provided in an embodiment of this application. This embodiment is a specific modification based on the above embodiment. Referring to Figure 7, the method provided in this embodiment specifically includes the following steps:

[0093] Step 610: Determine that the communication performance is lower than the preset target, and send a training start request to the transmitter.

[0094] Step 620: Generate baseline training frame data based on the training frame sequence.

[0095] The training frame sequence can be a data sequence used to generate training frames. The training frame sequence can be pre-configured in the receiver, and the same training frame sequence can be configured in the transmitter. The benchmark training frame data can be teacher data used to train the receiver model. The benchmark training frame data can be directly generated from the training frame sequence.

[0096] In this embodiment of the application, the transmitter can acquire a training frame sequence and select one or more frame data within the training frame sequence as reference training frame data.

[0097] Step 630: Receive the frequency domain symbol data transmitted by the transmitter, and generate the model output result corresponding to the frequency domain symbol data according to the receiver module.

[0098] In this embodiment, the transmitter can send frequency domain symbol data to the receiver. The frequency domain symbol data can be generated by the transmitter based on the training frame data in its training frame sequence. The receiver can receive the frequency domain symbol data and input it into the receiver model. The receiver model processes the frequency domain symbol data to generate the bit information LLR corresponding to the frequency domain symbol data. The bit information LLR can be used as the model output result.

[0099] Step 640: Determine the binary cross-entropy of the baseline training frame data and the model output as the loss value of the receiver model.

[0100] Here, binary cross-entropy can be a loss value determined by the binary cross-entropy loss function, which measures the difference between the baseline training frame data and the model output. The formula for calculating binary cross-entropy is as follows:

[0101] Where N is the number of samples used to train the receiver model, y i This represents the baseline training frame data, while This indicates the output of the model.

[0102] In this embodiment of the application, the binary cross-entropy of the benchmark training frame data and the model output result can be calculated by a pre-configured loss value function, and the binary cross-entropy can be used as the loss value of the receiver model.

[0103] Step 650: Adjust the model weights of the receiver model based on the loss value.

[0104] The model weights can be parameters in the neural network or other machine learning models within the receiver model. The model weights determine how the input features of the receiver model affect the output results. The model weights can be used to connect neural network elements in different layers within the receiver model.

[0105] In this embodiment, the model weights of the receiver model can be adjusted in the direction of decreasing loss value, so that the receiver model after adjusting the model weights has a smaller loss value.

[0106] Step 660: Determine that the performance indicators of the receiver model have met the preset requirements, send a training stop request to the transmitter, and stop the training of the receiver model.

[0107] Based on the above-described embodiments, benchmark training frame data is generated from training frames, including:

[0108] The training frame sequence number is determined based on the current frame sequence number, the current time slot sequence number, and the step size; the training frame data is then determined within the training frame sequence according to the training frame sequence number and used as the baseline training frame data.

[0109] In this embodiment of the application, the sequence frame number can be determined by the current frame sequence number, the current time slot sequence number and the step size, and the training frame data with the same sequence number can be determined in the training frame sequence according to the sequence frame sequence number, and the training frame data can be used as the reference training frame data.

[0110] Figure 8 is a flowchart of another receiver model training method provided in an embodiment of this application. This embodiment is a specific modification based on the above-mentioned embodiments. Referring to Figure 8, the method provided in this embodiment specifically includes the following steps:

[0111] Step 710: Determine that the communication performance is lower than the preset target, and send a training start request to the transmitter.

[0112] Step 720: Determine the channel estimation result based on the verified symbol-level data transmitted by the transmitter.

[0113] In this embodiment of the application, the receiver can receive symbol-level data transmitted by the transmitter and verify the verification information of the symbol-level data. After determining that the verification information of the symbol-level data has passed the verification, the receiver can perform decoding, modulation and channel estimation on the symbol-level data to obtain the channel estimation result corresponding to the symbol-level data.

[0114] Step 730: Obtain the training label dataset, wherein the training label dataset includes at least one set of randomly generated simulated bit transmission data.

[0115] The training label dataset can be a dataset used to train the receiver model. The training label dataset may include at least one set of simulated bit transmission data, which can be randomly generated. The simulated bit transmission data can be randomly generated bit data used to simulate transmitter transmission.

[0116] In this embodiment of the application, a training label dataset can be obtained, which may include randomly generated simulated bit transmission data.

[0117] Step 740: Modulate the analog bit transmission data into analog frequency domain symbol-level data according to the modulation method of symbol-level data.

[0118] Specifically, the adjustment method for symbol-level data can be determined, and the analog bit transmission data can be modulated into analog frequency domain symbol-level data according to the modulation method. The modulation method can include, but is not limited to, amplitude modulation, frequency modulation, phase modulation, quadrature amplitude modulation, binary phase shift keying modulation, quaternary phase shift keying modulation, minimum frequency shift keying modulation, orthogonal frequency division multiplexing modulation, etc.

[0119] Step 750: Generate received frequency domain data based on the analog frequency domain symbol-level data and the channel estimation results.

[0120] In this embodiment of the application, channel estimation can be performed on the analog frequency domain symbol-level data based on the channel estimation results determined by the frequency domain symbol data of the transmitter, thereby generating received frequency domain data.

[0121] Step 760: Call the receiver model to process the received frequency domain data into the model output result.

[0122] Specifically, the receiver model can be invoked to process the received frequency domain data. By processing the received frequency domain data through the receiver model, the corresponding bit information LLR of the received frequency domain data can be generated, and this bit information LLR can be used as the output result of the model.

[0123] Step 770: Determine the binary cross-entropy of the baseline training frame data and the model output as the loss value of the receiver model.

[0124] In this embodiment of the application, the binary cross-entropy of the benchmark training frame data and the model output result can be calculated by a pre-configured loss value function, and the binary cross-entropy can be used as the loss value of the receiver model.

[0125] Step 780: Adjust the model weights of the receiver model based on the loss value.

[0126] In this embodiment, the model weights of the receiver model can be adjusted in the direction of decreasing loss value, so that the receiver model after adjusting the model weights has a smaller loss value.

[0127] Step 790: Determine that the performance indicators of the receiver model have met the preset requirements, send a training stop request to the transmitter, and stop the training of the receiver model.

[0128] Based on the above-described embodiments, determining that the performance indicators of the receiver model meet preset requirements includes:

[0129] The current loss value of the receiver model is lower than the initial loss value, where the initial loss value is the loss value of the receiver model after the previous training iteration.

[0130] In this embodiment of the application, the initial loss value of the receiver model can be obtained. This loss value can be the loss value when the receiver model completed training in the previous time. The loss value can be saved as the initial loss value after each training of the receiver model. The current loss value of the current training can be compared with the saved initial loss value. When the current loss value is less than the initial loss value, it is determined that the performance index of the receiver model has reached the preset requirements.

[0131] In one exemplary implementation, the online training process of the receiver model can be nested within the receiver workflow. Referring to Figure 9, an online training sub-process can be set within the workflow of the AI ​​receiver. This sub-process can run in parallel with the main process executing the workflow. When the performance index of the communication system falls below a preset lower limit, the online training process of the receiver model is initiated. If the performance of the trained model is better than the original model, iterative updates of the receiver model within the AI ​​receiver are performed; otherwise, the receiver model is not updated. The online training process stops when the performance of the receiver model meets the requirements. The specific signaling flow between the transmitter and receiver can be shown in Figure 10.

[0132] In this embodiment of the application, the online training process needs to construct a training dataset, that is, the online training process needs to obtain the transmitted and received signals of both the transmitting and receiving ends after being transmitted through the channel and aligned before and after.

[0133] In some embodiments, the transmitting and receiving ends share data. In this application embodiment, a transmission sequence c(n) can be created to generate the transmitter's bit sequence; this data is called the training frame data for the AI ​​receiver's online learning process. The training frame data consumes a portion of the communication system's bandwidth; therefore, it is not continuously transmitted but rather at intervals of a certain number of frames.

[0134] The AI ​​receiver receives symbol-level data, while the main process continues normal model inference and decoding. The online training subprocess, however, needs to generate bit data as label data based on the training frame sequence c(n), then synchronize this data with the symbol-level data received by the AI ​​receiver and perform normalization to ensure the data is aligned to a uniform size, facilitating feature learning by the AI ​​receiver's neural network model. The normalized data is then freqdata. norm The LLR of this training is obtained by inputting it into the neural network model of the AI ​​receiver. The LLR and the label are used to calculate the loss, which is used to update the model parameters and save the model weights.

[0135] In the above process, the transmitter can perform the following procedures:

[0136] After receiving a request signaling from the receiver to start online training, the transmitter replies with a signaling message indicating its agreement to start online training and starts the online training subprocess.

[0137] The transmitter generates training frame data Bit based on the training frame sequence c(n). label As shown in equation (1):

[0138]

[0139] Where seed is the seed for the training frame sequence, frame is the current frame number, slot is the current time slot number, and seed is a hyperparameter for generating the training frame sequence. If the number of RBs currently scheduled is N... RB The number of scheduled REs is N RE The number of symbols to be scheduled is N. Symbol The current number of scheduled flows is N. l The modulation method is Q m The length len of the generated sequence is:

[0140] len = N RE ·N Symbol ·N l ·Q m

[0141] Bit of training frame data label The modulated symbol-level frequency domain data is mapped onto the scheduled stream and symbols, serving as the frequency domain data for the transmitted training frames. This training frame frequency domain data is then transmitted. The frames and slots of the training frame data are synchronized with the receiver, and the training frame sequence c(n) and hyperparameters are also kept consistent with the receiver to ensure that, given identical scheduling information, the receiver can accurately reconstruct the transmitter's training frame data bits. label Meanwhile, since training frames consume downlink bandwidth, an interval of num is chosen to ensure communication throughput.frame A training frame is sent, num frame Adjustments can be made based on the actual environment or the efficiency of online learning.

[0142] If the performance of the AI ​​receiver's neural network after online training meets the target, the receiver will send a request signal to the transmitter to stop the online training. Upon receiving the signal, the transmitter will stop the online training process and issue a signal agreeing to the stop.

[0143] The receiver can perform the following processing in the above process:

[0144] If the receiver determines that the communication throughput is below the preset lower limit, it sends a signal to request the start of online training. After receiving the transmitter's approval reply, it starts the online training subprocess.

[0145] The receiver maintains two sub-processes running in parallel. Process A handles the AI ​​receiver workflow, using the AI ​​neural network model for inference and decoding. Process B is used for online learning and training, updating the AI ​​receiver neural network model. During receiver initialization, each process preheats an AI receiver model on the device side. A With model B At this time, model B As a neural network model that needs to be updated during online training, it does not participate in the AI ​​receiver process; it serves as a preparatory model. A The neural network model, running as the main process, directly participates in the AI ​​receiver process. If the received frequency domain symbol-level data is freqdata, it is normalized to obtain freqdata. norm As the input to the neural network model of the AI ​​receiver, the output is LLR, as shown in equation (2). Meanwhile, freqdata... norm Synchronize to process B to obtain LLR for training. train As in equation (3):

[0146] LLR = model A (freqdata norm (2)

[0147] LLR train =model B (freqdata norm (3)

[0148] Process B uses the same training frame sequence as the transmitter, synchronously using the transmitter's training frame sequence c(n) and hyperparameters in equation (1) to generate tag data Bit. label .

[0149] Label data generated from training frame sequences (Bit) label With online training model B The inference result LLR train The loss value of the online training model is obtained by calculating the binary cross-entropy.

[0150] Loss = BCE(LLR) train Bit label )

[0151] Update the online-trained AI receiver model based on the calculated Loss value. B Network parameters. Iterate N times and save the updated model weights. new Where N is the maximum number of iterations for online training, set manually. If the model converges early, the iteration stops prematurely. If the trained model... B-new Performance verification shows it outperforms the original model. B If the algorithm is updated, the neural network model of the AI ​​receiver will be updated; otherwise, it will not be updated.

[0152] Use the updated model weights new Reload the online-trained AI receiver model that is already located on the device. B The updated model is obtained. B-new During the interval between inference execution and waiting for the next inference, the models in processes A and B are replaced. On the device side, only the model pointer needs to be switched, and the model update is completed at a cost of microseconds.

[0153] model A =model B-new

[0154] model B =model B-new

[0155] This method of updating and switching models in the AI ​​receiver process does not affect the original receiver process and allows for online training and iteration of the model in the background, saving a significant amount of time. However, it increases the receiver's computing power overhead. If the updated AI receiver meets the performance requirements, it uploads a request to stop online training and stops the online training process upon receiving a consent signal from the transmitter.

[0156] For example, suppose a single-input single-output (OFDM SISO) communication system has a single data stream and the number of RBs currently scheduled is N. RB =24, the number of scheduled REs is N RE =288, the number of symbols scheduled is N Symbol=14, modulation mode is Q m =4, the transceiver process of this application embodiment is as follows:

[0157] After receiving a request signaling from the receiver to start online training, the transmitter replies with a signaling message indicating its agreement to start online training and starts the online training subprocess.

[0158] The transmitter generates training frame data Bit based on the training frame sequence c(n). label As shown in equation (4):

[0159]

[0160] Where seed is the seed for the training frame sequence, frame is the current frame number, slot is the current slot number, seed is the hyperparameter for generating the training frame sequence, and step = 20.

[0161] The sequence length len generated from the training frame sequence is:

[0162] len = N RE ·N Symbol ·N l ·Q m =16128

[0163] Bit sequence label The modulated symbol-level frequency domain data is mapped onto the scheduled stream and symbols, and transmitted as the frequency domain data of the training frames. The frames and slots of the training frame data are synchronized with the receiver, and the training frame sequence c(n) and hyperparameters are also kept consistent with the receiver to ensure that, given the same scheduling information, the receiver can accurately reconstruct the original bit sequence of the transmitter's training frame data. label Meanwhile, since training frames consume downlink bandwidth, an interval of num is selected to ensure communication throughput. frame One training frame is sent per frame. Considering the actual environment and online learning efficiency, num frame Set to 512, which means sending one training frame every 5.12 seconds.

[0164] If the performance of the AI ​​receiver's neural network after online training meets the target, the receiver will send a request signal to the transmitter to stop the online training. Upon receiving the signal, the transmitter will stop the online training process and issue a signal agreeing to the stop.

[0165] The receiver determines that the communication throughput is less than a minimum threshold, i.e., Throughout. <T Thoughout This means sending a signal to request the initiation of online training, and starting the online training subprocess after receiving a confirmation reply from the transmitter.

[0166] The receiver maintains two sub-processes running in parallel. Process A handles the AI ​​receiver workflow, using the AI ​​neural network model for inference and decoding. Process B is used for online learning and training, updating the AI ​​receiver neural network model. During receiver initialization, each process preheats an AI receiver model on the device side. A With model B At this time, model B As a neural network model that needs to be updated during online training, it does not participate in the AI ​​receiver process; it serves as a preparatory model. A The neural network model, running as the main process, directly participates in the AI ​​receiver process. If the received frequency domain symbol-level data is freqdata, it is normalized to obtain freqdata. norm As the input to the neural network model of the AI ​​receiver, the output is LLR, as shown in equation (5). Meanwhile, freqdata... norm Synchronize to process B to obtain LLR for training. train As in equation (6):

[0167] LLR = model A (freqdata norm (5)

[0168] LLR train =model B (freqdata norm (6)

[0169] Process B uses the same training frame sequence as the transmitter, synchronously using the transmitter's c(n) and hyperparameters in equation (4) to generate the tag data Bit. label .

[0170] Label data generated from training frame sequences (Bit) label With online training model B The inference result LLR train The LOSS value of the online training model is obtained by calculating the binary cross-entropy.

[0171] Loss = BCE(LLR) train (label)

[0172] Update the online-trained AI receiver model based on the calculated LOSS value. B Network parameters. Iterate N times and save the updated model weights. newWhere N is the maximum number of iterations for online training, set manually. If the model converges early, the iteration stops early. If the loss value after training is lower than the original model, the neural network model of the AI ​​receiver is updated; otherwise, it is not updated.

[0173] Use the updated model weights new Reload the online-trained AI receiver model that is already located on the device. B The updated model is obtained. B-new During the interval between inference execution and waiting for the next inference, the models in processes A and B are replaced. On the device side, only the model pointer needs to be switched, and the model update is completed at a cost of microseconds.

[0174] model A =model B-new

[0175] model B =model B-new

[0176] This method of updating and switching models in the AI ​​receiver process does not affect the original receiver process and allows for online training and iteration of the model in the background, saving a significant amount of time, but increasing the receiver's computing power overhead. If the updated AI receiver's performance meets the requirements, it uploads a request to stop online training and stops the online training process upon receiving a consent signal from the transmitter. This embodiment of the application shares a set of training frame data between the transmitter and receiver, using the training frame sequence to generate training frame bit data, which serves as both the transmitter's transmission signal and the receiver's online training tag data. This solves the problem of difficulty in synchronizing training data between the transmitter and receiver; the receiver only needs timestamps to reconstruct the transmitter's original training frame data.

[0177] In other embodiments, the receiver's online training dataset for the AI ​​receiver is reconstructed from the channel data of the frame decrypted using CRC decoding. The transmitter no longer needs to use the training frames for data alignment between the transmitting and receiving ends, reducing throughput overhead. The receiver performs modulation mapping based on the decoded bit data to reconstruct the transmitted frequency domain data and performs channel estimation with the received frequency domain data, saving the channel estimation result H of the CRC-verified frame. CRCOK Multiple sets of transmitted bit data are randomly generated at the receiver side and used as labels for this training set. The bit data is then mapped to Freqdata, the frequency domain symbol-level data generated by the modulation scheme of the frame with correct CRC check. label Simulate passing through channel H CRCOK Freqdata (Further frequency domain data) recvThe input is fed into the AI ​​receiver's neural network model to obtain the inference result LLR. The LLR and label are used to calculate the loss, which is used to update the model parameters and save the model weights, as shown in Figure 11. The updated model weights from online learning are loaded into the device model of the sub-process, completing the update of the device-side model. This sub-process model is then replaced with the main process model, completing one online learning operation.

[0178] The transmitter can execute existing transmitter procedures without requiring additional online training processes. The receiver, however, initiates an online training subprocess when it determines that the communication throughput is below a preset lower limit.

[0179] The receiver maintains two sub-processes running in parallel. Process A handles the AI ​​receiver workflow, using the AI ​​neural network model for inference and decoding. Process B is used for online learning and training, updating the AI ​​receiver neural network model. During receiver initialization, each process preheats an AI receiver model on the device side. A With model B At this time, model B As a neural network model that needs to be updated during online training, it does not participate in the AI ​​receiver process; it serves as a preparatory model. A The neural network model, which runs as the main process, directly participates in the AI ​​receiver process.

[0180] The receiver collects data frames that have passed N Cyclic Redundancy Check (CRC) tests. Based on the decoded bit data, it performs modulation mapping to reconstruct the transmitted frequency domain data and performs channel estimation with the received frequency domain data. The channel estimation result H of the CRC-checked frames is then extracted and saved. CRCOK During the online training subprocess, multiple sets of random simulated bit data will be generated. label This serves as the labeled dataset for online training. Bit... label Frequency domain symbol-level data Freqdata is obtained by modulation mapping according to the modulation scheme scheduled by the frame with correct CRC check. label And simulate passing through channel H CRCOK The simulated received frequency domain data Freqdata is obtained. recv As shown in equation (7):

[0181] Freqdata recv =H CRCOK ·Freqdata label (7)

[0182] The simulated received frequency domain data Freqdata recvAI receiver model fed into online training B To obtain the LLR used for training train As in equation (8):

[0183] LLR train =model B (Freqdata recv (8)

[0184] Label data generated from training frame sequences (Bit) label With online training model B The inference result LLR train The loss value of the online training model is obtained by calculating the binary cross-entropy.

[0185] Loss = BCE(LLR) train Bit label )

[0186] Update the online-trained AI receiver model based on the calculated Loss value. B Network parameters. Iterate N times and save the updated model weights. new , where N is the maximum number of iterations for online training set manually. If the model converges early, the iteration stops early. If the loss value after training is lower than the original model, the neural network model of the AI ​​receiver is updated; otherwise, it is not updated.

[0187] Use the updated model weights new Reload the online-trained AI receiver model that is already located on the device. B The updated model is obtained. B-new During the interval between inference execution and waiting for the next inference, the models in processes A and B are replaced. On the device side, only the model pointer needs to be switched, and the model update is completed at a cost of microseconds.

[0188] model A =model B-new

[0189] model B =model B-new

[0190] This method of updating and switching models in the AI ​​receiver process does not affect the original receiver process and allows for online training and iteration of the model in the background, saving a significant amount of time. However, it increases the receiver's computing power overhead. If the updated AI receiver meets the performance requirements, the online training process is stopped.

[0191] In one exemplary implementation, assume that the data in an OFDM MIMO communication system consists of two streams, and the number of RBs currently scheduled is N. RB =6, the number of scheduled REs is N RE =72, the number of symbols scheduled is N Symbol =14, modulation mode is Q m =6, the receiver determines that the communication bit error rate is higher than the preset bit error rate threshold, i.e., Bler>T Bler This means starting the online training subprocess.

[0192] The receiver maintains two sub-processes running in parallel. Process A handles the AI ​​receiver workflow, using the AI ​​neural network model for inference and decoding. Process B is used for online learning and training, updating the AI ​​receiver neural network model. During receiver initialization, each process preheats an AI receiver model on the device side. A With model B At this time, model B As a neural network model that needs to be updated during online training, it does not participate in the AI ​​receiver process; it serves as a preparatory model. A The neural network model, which runs as the main process, directly participates in the AI ​​receiver process.

[0193] If the receiver's CRC check is correct, the transmitted frequency domain data is reconstructed by modulation mapping based on the decoded bit data, and channel estimation is performed between the reconstructed data and the received frequency domain data. The channel estimation result H of the frame with correct CRC check is then saved. CRCOK The process collects N frames of channel information. During the online training subprocess, multiple sets of random simulated transmission bit data are generated. label This serves as the labeled dataset for online training. Bit... label According to the modulation scheme Q of the frame with correct CRC check. m =6 modulation mapping yields frequency domain symbol-level data Freqdata label And simulate passing through channel H CRCOK The simulated received frequency domain data Freqdata is obtained. recv As shown in equation (9):

[0194] Freqdata recv =H CRCOK ·Freqdata label (9)

[0195] The simulated received frequency domain data Freqdata recv AI receiver model fed into online training B To obtain the LLR used for training train As in equation (10):

[0196] LLR train =model B (Freqdata recv (10)

[0197] Label data generated from training frame sequences (Bit) label With online training model B The inference result LLR train The loss value of the online training model is obtained by calculating the binary cross-entropy.

[0198] Loss = BCE(LLR) train Bit label )

[0199] Update the online-trained AI receiver model based on the calculated Loss value. B Network parameters. Iterate N times and save the updated model weights. new , where N is the maximum number of iterations for online training set manually. If the model converges early, the iteration stops early. If the loss value after training is lower than the original model, the neural network model of the AI ​​receiver is updated; otherwise, it is not updated.

[0200] Use the updated model weights new Reload the online-trained AI receiver model that is already located on the device. B The updated model is obtained. B-new During the interval between inference execution and waiting for the next inference, the models in processes A and B are replaced. On the device side, only the model pointer needs to be switched, and the model update is completed at a cost of microseconds.

[0201] model A =model B-new

[0202] model B =model B-new

[0203] This method of updating and switching models in the AI ​​receiver process does not affect the original receiver process and allows for online training and iteration of the model in the background, saving a significant amount of time, but increasing the receiver's computing power overhead. If the performance of the updated AI receiver meets the requirements, the online training process stops. In this embodiment, multiple sets of bit data are simulated as training tags by verifying the channel data of the correct frames using CRC. The simulated transmission and reception process passes through this channel data to obtain simulated received data. The simulated received data and training tags are used together for online training, which can eliminate throughput loss. The entire online training process can be completed only on the receiver side, without the need for a transmitter.

[0204] The method provided in this application embodiment can fine-tune the model parameters of the AI ​​receiver model without affecting the receiver signal reception and processing flow, thereby improving the adaptability of the receiver model to the environment.

[0205] Figure 12 is a schematic diagram of a receiver model training device provided in an embodiment of this application. This device can execute the receiver model training method provided in any embodiment of this application, and possesses the corresponding functional modules and beneficial effects of the method. This device can be implemented by software and / or hardware. The device provided in this embodiment specifically includes:

[0206] The training module 810 is used to obtain the training start request transmitted by the receiver and to train the receiver model.

[0207] The stop module 820 is used to receive the training stop request transmitted by the receiver and end the training of the receiver model.

[0208] Based on the above-described embodiments, the training module 810 includes:

[0209] The instruction feedback unit is used to respond to the training start request transmitted by the receiver and to provide a training start consent instruction to the receiver.

[0210] The process training unit is used to start an online training subprocess and train the receiver model based on the online training subprocess.

[0211] Based on the above-described embodiments, the training module 810 includes:

[0212] The data generation unit is used to generate training frame data based on the training frame sequence.

[0213] The data modulation unit is used to modulate the training frame data into frequency domain symbol data;

[0214] A data transmission unit is used to transmit the frequency domain symbol data to the receiver so that the receiver can train the receiver model.

[0215] Based on the above application embodiments, the data generation unit is specifically used to: determine the training frame sequence number according to the current frame sequence number, the current time slot sequence number and the step size; and determine the training frame data in the training frame sequence according to the training frame sequence number.

[0216] Based on the above application embodiments, the sequence length of the training frame data is determined by at least one of the following: the number of scheduling resource elements, the number of scheduling symbols, the number of scheduling streams, and the adjustment method.

[0217] Based on the above-described embodiments, the data generation unit determines the training frame sequence number according to the current frame sequence number, the current time slot sequence number, and the step size, including:

[0218] Determine the product of the current frame sequence number and the step size, and use the sum of the product and the current time slot sequence number as the training frame sequence.

[0219] Based on the above-mentioned application embodiments, the data transmission unit is specifically used to: transmit one frame of the frequency domain symbol data for every data frame interval of a threshold number, wherein the threshold number is adjusted according to the learning efficiency of the receiver model.

[0220] Based on the above-described embodiments, the apparatus further includes: a symbol transmission module, used to send symbol-level data to the receiver.

[0221] Figure 13 is a schematic diagram of another receiver model training device provided in an embodiment of this application. This device can execute the receiver model training method provided in any embodiment of this application, and has the corresponding functional modules and beneficial effects of the method execution. This device can be implemented by software and / or hardware. The device provided in this embodiment specifically includes:

[0222] The training start module 910 is used to determine that the communication performance is lower than the preset index and send a training start request to the transmitter.

[0223] Model training module 920 is used to train the receiver model.

[0224] The training termination module 930 is used to determine that the performance indicators of the receiver model have reached the preset requirements, send a training stop request to the transmitter, and stop the training of the receiver model.

[0225] Based on the above-described embodiments, the model training module 920 includes:

[0226] The instruction receiving unit is used to receive the training start consent instruction fed back by the transmitter.

[0227] The training execution unit is used to start an online training subprocess and train the receiver model based on the online training subprocess.

[0228] Based on the above-described embodiments, the model training module 920 is specifically configured to: generate baseline training frame data according to the training frame sequence; receive frequency domain symbol data transmitted by the transmitter, and generate model output results corresponding to the frequency domain symbol data according to the receiver module; determine the binary cross-entropy of the baseline training frame data and the model output results as the loss value of the receiver model; and adjust the model weights of the receiver model according to the loss value.

[0229] In the above-mentioned application embodiment, the model training module 920 generates benchmark training frame data based on training frames, including: determining the training frame sequence number according to the current frame sequence number, the current time slot sequence number, and the step size; and determining training frame data as the benchmark training frame data within the training frame sequence according to the training frame sequence number.

[0230] Based on the above-described embodiments, the model training module 920 is further specifically configured to: determine the channel estimation result based on the verified symbol-level data transmitted by the transmitter; acquire a training label dataset, wherein the training label dataset includes at least one set of randomly generated simulated bit transmission data; modulate the simulated bit transmission data into simulated frequency domain symbol-level data according to the modulation method of the symbol-level data; generate received frequency domain data based on the simulated frequency domain symbol-level data and the channel estimation result; call the receiver model to process the received frequency domain data into a model output result; determine the binary cross-entropy of the baseline training frame data and the model output result as the loss value of the receiver model; and adjust the model weights of the receiver model according to the loss value.

[0231] Based on the above-mentioned embodiments, the apparatus further includes: a channel estimation module, used to receive symbol-level data transmitted by the transmitter; to start an online inference subprocess, and to generate a receiver output result of the symbol-level data by calling a second receiver model through the online inference subprocess, wherein the second receiver model has the same model weights as the receiver model before training; and to obtain a channel estimation result by decoding the receiver output result according to the decoder.

[0232] Based on the above-mentioned embodiments, the apparatus further includes: a model replacement module, configured to determine that the receiver model training is complete, replace the receiver model in the online inference subprocess, and replace the second receiver model in the online training subprocess.

[0233] Based on the above application embodiments, the training end module 930 determines that the performance index of the receiver model has reached the preset requirements, including: the current loss value of the receiver model is lower than the initial loss value, wherein the initial loss value is the loss value of the receiver model in the previous training.

[0234] Figure 14 is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device includes a processor 10, a memory 11, an input device 12, and an output device 13. The number of processors 10 in the electronic device can be one or more. Figure 14 shows one processor 10 as an example. The processor 10, memory 11, input device 12, and output device 13 in the electronic device can be connected by a bus or other means. Figure 14 shows a connection via a bus as an example.

[0235] The memory 11, as a computer-readable storage medium, can be used to store software programs, computer-executable programs, and modules, such as the modules corresponding to the device in the embodiments of this application (training module 810 and stop module 820, or training start module 910, model training module 920, and training end module 930). The processor 10 executes various functional applications and data processing of the electronic device by running the software programs, instructions, and modules stored in the memory 11, that is, it implements the above-described receiver model training method.

[0236] The memory 11 may primarily include a program storage area and a data storage area. The program storage area may store the operating system and at least one application program required for a given function; the data storage area may store data created based on the use of the electronic device. Furthermore, the memory 11 may include high-speed random access memory and non-volatile memory, such as at least one disk storage device, flash memory, or other non-volatile solid-state storage device. In some instances, the memory 11 may further include memory remotely located relative to the processor 10, which can be connected to the electronic device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0237] Input device 12 can be used to receive input digital or character information, and to generate key signal inputs related to user settings and function control of the electronic device. Output device 13 may include display devices such as a display screen.

[0238] This application embodiment also provides a storage medium containing computer-executable instructions, which, when executed by a computer processor, are used to perform a receiver model training method, the method comprising:

[0239] Obtain the training start request from the receiver and train the receiver model;

[0240] The training stop request transmitted by the receiver is received, and the training of the receiver model is terminated.

[0241] And / or, the method includes:

[0242] If the communication performance is determined to be below the preset target, a training start request is sent to the transmitter.

[0243] Train the receiver model;

[0244] Once the performance metrics of the receiver model are determined to meet the preset requirements, a training stop request is sent to the transmitter, and the training of the receiver model is stopped.

[0245] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0246] It is worth noting that in the embodiments of the above-mentioned device, the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be realized; in addition, the specific names of each functional unit are only for easy differentiation and are not used to limit the scope of protection of this application.

[0247] Those skilled in the art will understand that all or some of the steps, apparatuses, or functional modules / units in the methods disclosed above can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0248] In hardware implementations, the division between functional modules / units mentioned in the above description does not necessarily correspond to the division of physical components; for example, a physical component may have multiple functions, or a function or step may be performed collaboratively by several physical components. Some or all physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. The corresponding software may be distributed on a computer-readable medium, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, program modules, or other data). Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disc (DVD) or other optical disc storage, magnetic cartridges, magnetic tape, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer. Furthermore, as is known to those skilled in the art, communication media typically contain computer-readable instructions, data structures, program modules, or other data in modulated data signals such as carrier waves or other transmission mechanisms, and may include any information delivery medium.

[0249] The above description, with reference to the accompanying drawings, illustrates preferred embodiments of the present invention, but does not limit the scope of the invention. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and spirit of the present invention should be considered within the scope of the present invention.

Claims

1. A receiver model training method, characterized in that, Applied to a transmitter, the method includes: acquiring a training start request transmitted by a receiver and training a receiver model; receiving a training stop request transmitted by the receiver and ending the training of the receiver model.

2. The method according to claim 1, characterized in that, The step of obtaining the training start request transmitted by the receiver and training the receiver model includes: responding to the training start request transmitted by the receiver by sending a training start consent instruction to the receiver; starting an online training subprocess and training the receiver model based on the online training subprocess.

3. The method according to claim 1, characterized in that, The training of the receiver model includes: generating training frame data based on the training frame sequence; modulating the training frame data into frequency domain symbol data; and transmitting the frequency domain symbol data to the receiver so that the receiver can train the receiver model.

4. The method according to claim 3, characterized in that, The step of generating training frame data based on the training frame sequence includes: determining the training frame sequence number based on the current frame sequence number, the current time slot sequence number, and the step size; and determining the training frame data within the training frame sequence according to the training frame sequence number.

5. The method according to claim 3, characterized in that, The sequence length of the training frame data is determined by at least one of the following: the number of scheduling resource elements, the number of scheduling symbols, the number of scheduling streams, and the adjustment method.

6. The method according to claim 4, characterized in that, The step of determining the training frame sequence number based on the current frame sequence number, the current time slot sequence number, and the step size includes: determining the product of the current frame sequence number and the step size, and using the sum of the product and the current time slot sequence number as the training frame sequence.

7. The method according to claim 3, characterized in that, The step of transmitting the frequency domain symbol data to the receiver includes: transmitting one frame of the frequency domain symbol data at intervals of a threshold number of data frames, wherein the threshold number is adjusted according to the learning efficiency of the receiver model.

8. The method according to claim 1, characterized in that, Also includes: Symbol-level data is sent to the receiver.

9. A receiver model training method, characterized in that, The method, applied to a receiver, includes: determining that the communication performance is lower than a preset indicator, sending a training start request to the transmitter; training the receiver model; determining that the performance indicator of the receiver model has reached a preset requirement, sending a training stop request to the transmitter, and stopping the training of the receiver model.

10. The method according to claim 9, characterized in that, The process of training the receiver model includes: receiving a training start consent instruction from the transmitter; starting an online training subprocess; and training the receiver model based on the online training subprocess.

11. The method according to claim 9, characterized in that, The training of the receiver model includes: generating baseline training frame data based on the training frame sequence; receiving frequency domain symbol data transmitted by the transmitter and generating model output results corresponding to the frequency domain symbol data based on the receiver module; determining the binary cross-entropy of the baseline training frame data and the model output results as the loss value of the receiver model; and adjusting the model weights of the receiver model based on the loss value.

12. The method according to claim 11, characterized in that, The step of generating baseline training frame data based on training frames includes: determining the training frame sequence number based on the current frame sequence number, the current time slot sequence number, and the step size; and determining training frame data within the training frame sequence according to the training frame sequence number as the baseline training frame data.

13. The method according to claim 9, characterized in that, The training of the receiver model includes: determining the channel estimation result based on the verified symbol-level data transmitted by the transmitter; acquiring a training label dataset, wherein the training label dataset includes at least one set of randomly generated simulated bit transmission data; modulating the simulated bit transmission data into simulated frequency domain symbol-level data according to the modulation method of the symbol-level data; generating received frequency domain data based on the simulated frequency domain symbol-level data and the channel estimation result; calling the receiver model to process the received frequency domain data into a model output result; determining the binary cross-entropy of the baseline training frame data and the model output result as the loss value of the receiver model; and adjusting the model weights of the receiver model according to the loss value.

14. The method according to any one of claims 9, 11, or 13, characterized in that, Also includes: Receive symbol-level data transmitted by the transmitter; Start the online inference subprocess, and use the online inference subprocess to call the second receiver model to generate the receiver output result of the symbol-level data, wherein the second receiver model has the same model weight as the receiver model before training; The channel estimation result is obtained by decoding the receiver output based on the decoder.

15. The method according to claim 14, characterized in that, Also includes: Once the receiver model training is complete, the receiver model is replaced in the online inference subprocess, and the second receiver model is replaced in the online training subprocess.

16. The method according to claim 9, characterized in that, The step of determining that the performance index of the receiver model meets the preset requirements includes: the current loss value of the receiver model is lower than the initial loss value, wherein the initial loss value is the loss value of the receiver model in the previous training.

17. An electronic device, characterized in that, The electronic device includes: one or more processors; a memory for storing one or more programs; and when the one or more programs are executed by the one or more processors, the one or more processors implement the receiver model training method as described in any one of claims 1-16.

18. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, which are executed by one or more processors to implement the receiver model training method as described in any one of claims 1-16.

Images