Model training methods, communication device and storage medium

Abstract

The present application provides model training methods, a communication device and a storage medium. A model training method, applied to a first communication node, comprises: receiving raw data frames sent by a second communication node for training a neural network model; and using the raw data frames to perform online training on the neural network model.

Description

Model training method, communication device, and storage medium TECHNICAL FIELD

[0001] The present application relates to the technical field of communication, in particular to a model training method, a communication device, and a storage medium. BACKGROUND

[0002] At present, neural networks have been widely applied to various systems to replace traditional processing modules and have achieved good results. However, due to the black box characteristics of neural networks, their reliability and generalization ability have always been questioned. In the reasoning phase, if the neural network module encounters data that has not been seen in the training phase, its behavior cannot be estimated, and it may get completely illogical results. SUMMARY

[0003] Therefore, the embodiments of the present application provide a model training method, a communication device, and a storage medium to solve the technical problem of poor generalization ability and low reliability of a pre-trained model in the prior art.

[0004] The embodiments of the present application provide a model training method applied to a first communication node, comprising:

[0005] receiving an original data frame for neural network model training sent by a second communication node;

[0006] training the neural network model online using the original data frame.

[0007] The embodiments of the present application provide a model training method applied to a second communication node, comprising:

[0008] sending an original data frame for neural network model training to a first communication node, so that the first communication node trains the neural network model online using the original data frame.

[0009] The embodiments of the present application provide a model training device applied to a first communication node, comprising:

[0010] a receiver configured to receive an original data frame for neural network model training sent by a second communication node;

[0011] a training module configured to train the neural network model online using the original data frame.

[0012] The embodiments of the present application provide a model training method applied to a second communication node, comprising:

[0013] a transmitter configured to send an original data frame for neural network model training to a first communication node, so that the first communication node trains the neural network model online using the original data frame.

[0014] An embodiment of the present application provides a communication device, comprising a memory and one or more processors;

[0015] The memory is configured to store one or more programs;

[0016] When the one or more programs are executed by the one or more processors, the one or more processors implement the method in any of the above embodiments.

[0017] An embodiment of the present application provides a storage medium, which stores a computer program, and the computer program is executed by a processor to implement the method in any of the above embodiments. BRIEF DESCRIPTION OF DRAWINGS

[0018] FIG. 1 is a flowchart of a model training method according to an embodiment of the present application;

[0019] FIG. 2 is a schematic diagram of an implementation of a pilot-based OFDM system intelligent receiver model according to an embodiment of the present application;

[0020] FIG. 3 is a schematic diagram of an implementation of a pilot-free OFDM system intelligent receiver model according to an embodiment of the present application;

[0021] FIG. 4 is a schematic diagram of an implementation of a SIP-based OFDM system intelligent receiver model according to an embodiment of the present application;

[0022] FIG. 5 is a schematic diagram of an implementation of data cache updating according to an embodiment of the present application;

[0023] FIG. 6 is a schematic diagram of an implementation of data enhancement using a random phase method according to an embodiment of the present application;

[0024] FIG. 7 is a schematic diagram of an implementation of data enhancement using a constellation symmetry method according to an embodiment of the present application;

[0025] FIG. 8 is a schematic diagram of another implementation of data enhancement using a constellation symmetry method according to an embodiment of the present application;

[0026] FIG. 9 is a schematic diagram of an implementation of data enhancement using a resource symmetry method according to an embodiment of the present application;

[0027] FIG. 10 is a schematic diagram of another implementation of data enhancement using a resource symmetry method according to an embodiment of the present application;

[0028] FIG. 11 is a schematic diagram of an implementation of data enhancement using a bandwidth segmentation method according to an embodiment of the present application;

[0029] FIG. 12 is an implementation flowchart of a cross-validation training algorithm according to an embodiment of the present application;

[0030] FIG. 13 is an implementation flowchart of another cross-validation training algorithm according to an embodiment of the present application;

[0031] FIG. 14 is a flowchart of another model training method according to an embodiment of the present application;

[0032] FIG. 15 is a structural block diagram of a model training apparatus according to an embodiment of the present application;

[0033] FIG. 16 is a structural block diagram of another model training apparatus according to an embodiment of the present application;

[0034] FIG. 17 is a structural schematic diagram of a communication device according to an embodiment of the present application. DETAILED DESCRIPTION

[0035] Hereinafter, the embodiments of the present application will be described with reference to the accompanying drawings. The present application is described below by way of examples in conjunction with the accompanying drawings, which are presented for the purpose of illustration only and are not intended to limit the scope of the present application.

[0036] With the development and maturity of artificial intelligence technology, more and more technical fields have begun to realize the value of artificial intelligence technology and gradually applied it to the field to solve some problems that are difficult to solve by traditional means.

[0037] In the field of wireless communication, artificial intelligence technology has also achieved great success. In the 5G stage, device manufacturers and operators have tried to use artificial intelligence technology to improve link energy and optimize network deployment, and have achieved good results. At the beginning of the 6G discussion, many experts and scholars from operators and device manufacturers quickly reached a consensus that AI will play an important role in 6G standards and decided to work together to promote the application of artificial intelligence technology in 6G standards.

[0038] Although everyone from the academic community to the industry community agrees that artificial intelligence technology is a disruptive technology that will have a revolutionary impact on future communication systems, they have also expressed concerns about the reliability and generalization ability of the technology.

[0039] Traditional communication system design is based on solid theoretical foundation of mathematics, electromagnetism, etc. The processing of each module in the system has a logical explanation. As long as the basic theory is not broken, it can always give a reasonable solution. Artificial intelligence technology, more specifically, neural network technology has a well-known black box feature. It maps input data to output data, and the internal mapping process cannot be explained by basic theory. This unexplainability means that when the input of the neural network is a data that has never appeared in the training data set, the output of the neural network is completely unpredictable, and the performance may drop sharply.

[0040] One way to solve this problem is to use a large enough data set to train the model that can cover all possible use cases, so that the model can experience all possible inputs in the training stage. This method has two obvious shortcomings. First, in a wireless communication system, the biggest uncertainty comes from the channel, and the complexity of the channel environment is large. In addition to the well-known multipath fading characteristics, there are various types of completely unpredictable interference, radio frequency nonlinearity and other influencing factors. It is almost impossible to collect training data that can cover all cases. Second, based on common sense, a model that can adapt to various application scenarios must be a relatively large model. The number of parameters of the model and the amount of calculation of the inference process will exceed the model that adapts to one or several scenarios, which will bring great challenges to the cost control of the product. Considering the above factors, this solution is not feasible.

[0041] Another way to solve this problem is to introduce an online training / optimization mechanism, use a model with good generalization ability (called a pre-trained model), and continuously train the model during the interaction between the two communication parties, so that the model gradually matches the current channel environment. This method is based on a simple fact that although the channel environment is complex and variable, its statistical characteristics, such as time domain, frequency domain correlation, and signal-to-noise ratio (SNR) of the channel, are relatively stable within a certain period of time. Therefore, as long as the model converges fast enough in the online training process, the model can adapt to the change of the channel, and its generalization ability and reliability are guaranteed.

[0042] Current research on online training / optimization focuses on the Internet field and is mainly used to predict user click behavior, so as to make content recommendation and advertisement placement more accurate. There is currently less research on online training in the field of wireless communication. The present application provides an online training / optimization scheme in a wireless communication system, which can effectively solve the problems of poor generalization ability and low reliability of the pre-trained model.

[0043] In an embodiment, FIG. 1 is a flowchart of a model training method provided by embodiments of the present application. The embodiment is applied to the case of online training in a wireless communication system. The embodiment can be performed by a first communication node. In an example, the first communication node can be a terminal side, and the second communication node can be a base station side; in an example, the first communication node can be a base station side, and the second communication node can be a terminal side. As shown in FIG. 1, the embodiment includes S110-S120.

[0044] S110, receiving an original data frame for neural network model training sent by a second communication node.

[0045] In an example, the first communication node is a terminal side, and the second communication node is a base station side. The second communication node can send the original data frame for neural network model training to multiple first communication nodes in its coverage range in a broadcast manner. Correspondingly, all first communication nodes in the coverage range of the second communication node receive the original data frame sent by the second communication node through their own antennas. In an example, the first communication node is a terminal side, and the second communication node is a base station side. The second communication node can send the original data frame for neural network model training to one first communication node in a unicast manner. In an example, in the process of transmitting the original data frame in a broadcast manner, the second communication node can actively issue the original data frame for neural network model training to the first communication node. In an example, in the process of transmitting the original data frame in a unicast manner, the first communication node sends a data frame request to the second communication node. After receiving the data frame request, the second communication node issues the original data frame for neural network model training to the first communication node corresponding to the data frame request.

[0046] S120, performing online training on the neural network model using the original data frame.

[0047] The second communication node performs online training on the neural network model using the original data frame, so that the trained neural network model can adapt to the current wireless communication environment, and solve the technical problems of low reliability and insufficient generalization ability of the neural network model.

[0048] In an embodiment, the receiving frequency of the original data frame comprises: periodic; aperiodic. In an example, the original data frame can be received periodically for the first communication node; the original data frame can be transmitted periodically for the second communication node. In an example, the transmission period and the receiving period of the original data frame can be characterized by the number of training frames. For example, the transmission period and the receiving period of the original data frame can be 5 training frames. In an example, the first communication node can periodically transmit the original data frame for neural network model training to the second communication node, or aperiodically transmit the original data frame for neural network model training to the second communication node.

[0049] In an embodiment, the application range of the original data frame comprises: specific terminal; specific cell. In an example, the original data frame can be UE-specific, or cell-specific.

[0050] In an embodiment, the transmission mode of the original data frame comprises: the second communication node actively transmitting to the first communication node; the first communication node requesting the second communication node to transmit. In an example, the second communication node can actively transmit the original data frame to the first communication node. In an example, the first communication node can transmit the original data frame to the second communication node after the second communication node transmits a request to the first communication node. In an example, in the case that the second communication node actively transmits the original data frame to the first communication node, the original data frame can be cell-specific, i.e. all terminals in the cell can receive the original data frame. In an example, in the case that the second communication node transmits the original data frame to the first communication node after the first communication node transmits a request to the second communication node, the original data frame can be UE-specific, i.e. only the first communication node that transmits the request can receive the original data frame.

[0051] In an embodiment, the bandwidth of the original data frame is greater than or equal to the processing bandwidth of the neural network model. In an example, the processing bandwidth of the neural network model refers to the frequency domain bandwidth required for the data that the neural network model can process at a time; the bandwidth of the original data frame refers to the frequency domain bandwidth required for the original data frame. In an example, the frequency domain bandwidth can be characterized by the number of subcarriers.

[0052] In an embodiment, the neural network model comprises one of: a pilot-based OFDM system intelligent receiver model; a pilot-free OFDM system intelligent receiver model; a SIP-based OFDM system intelligent receiver model.

[0053] In an example, the neural network model can be a pilot-based Orthogonal Frequency Division Multiplexing (OFDM) system intelligent receiver model. FIG. 2 is an implementation schematic diagram of a pilot-based OFDM system intelligent receiver model according to an embodiment of the present application. As shown in FIG. 2, in the pilot-based OFDM system intelligent receiver model, the OFDM demodulation signal and the reference signal are taken as the input of the intelligent receiver model, and the Log-Likelihood Ratio (LLR) information is taken as the output of the intelligent receiver model. The label during model training is bit stream, and the loss function is Binary Cross Entropy (BCE). The function is equivalent to the combination of the functions of the channel estimation, equalization and demodulation modules in the traditional link.

[0054] In an example, the neural network model can be a pilot-free OFDM system intelligent receiver model. FIG. 3 is an implementation schematic diagram of a pilot-free OFDM system intelligent receiver model according to an embodiment of the present application. As shown in FIG. 3, the pilot-free OFDM system modulates data using a special constellation (for example, the constellation with modulation order 4 in the figure). Specifically, the constellation has a rotation uniqueness, that is, unless the rotation angle is an integer multiple of 2π, the rotated constellation and the original constellation will not coincide. In the pilot-free OFDM system intelligent receiver model, the OFDM demodulation signal is taken as the input of the intelligent receiver model, and the LLR information is taken as the output of the intelligent receiver model. The label during model training is bit stream, and the loss function is BCE.

[0055] In an example, the neural network model can be a Session Initiation Protocol (SIP) based OFDM system intelligent receiver model. FIG. 4 is an implementation schematic diagram of a SIP based OFDM system intelligent receiver model according to an embodiment of the present application. As shown in FIG. 4, the SIP based OFDM system can send data on all resources, and also send reference signals on all resources, without distinguishing independent data resources and reference signal resources. The data resource network and the reference signal resource network are multiplied by different power weighting factors and then superimposed together, but the sum of the average powers of the data and the reference signals needs to be equal to 1. In the figure, A is the power weighting factor of the reference signal, which is a two-dimensional matrix and can be referred to as a weighting factor matrix. The dimension of the weighting factor matrix A is the same as the current resource network (subcarrier number x symbol number), and each element in A is a real number greater than 0 and less than 1, representing the power of the reference signal on the RE resource.

[0056] In the SIP-based OFDM system intelligent receiver model, the OFDM demodulation signal and the reference signal resource grid are taken as the input of the intelligent receiver model, the LLR information is taken as the output of the intelligent receiver model, the label during model training is the bit stream, and the loss function is the BCE. In an example, in the SIP-based OFDM system intelligent receiver model, a two-dimensional matrix composed of the power weighting factors of the reference signals can be taken as the input of the intelligent receiver model.

[0057] In an example, the intelligent receiver model can be referred to as an intelligent receiver network, and can also be referred to as an intelligent receiver for short.

[0058] In an embodiment, the neural network model is trained online by using the original data frames, including:

[0059] The original data frames are subjected to data cache updating and data enhancement to obtain target data frames;

[0060] The target data frames are subjected to data grouping according to a data grouping mode to obtain two groups of target data;

[0061] The neural network model is trained online based on a cross-validation training algorithm and the two groups of target data. In an example, the online training of the neural network model by using the original data frames includes the following four steps: data cache updating, data enhancement, data grouping, and training of the neural network model by using the cross-validation training algorithm.

[0062] In an embodiment, the implementation process of the data cache updating includes: pre-configuring a cache space with a target cache space size; storing the received new original data frames in the cache space, and deleting the original data frame with the earliest cache time in the cache space from the cache space. In an example, the target cache space size can be represented by the number of training frames. Illustratively, the target cache space size can be 3 training frames, that is, at most 3 training frames can be accommodated in the cache space with the target cache space size. In an example, the management of the cache space can be based on the first-in first-out (FIFO) strategy, that is, the first communication node adds the new original data frame to the cache space and deletes the original data frame with the earliest time in the cache space from the cache space each time a new original training frame is received.

[0063] Figure 5 is an implementation diagram of data cache updating according to an embodiment of the present application. The data cache updating refers to setting a target cache space size (denoted as save_length), when a new training frame (i.e. a new original data frame) arrives, the training frame data is added to the cache space, and the earliest data in the cache space is deleted. As shown in Figure 5, the target cache space size of the cache space is set to 3 training frames, and the training frames are periodically sent, and the sending period is 5 training frames.

[0064] Unlike the conventional offline training mode, the online training has the following characteristics: first, the training samples are limited; second, the training samples arrive one after another in time; third, the device for online training usually has limited storage space and computing resources; and fourth, the online training has certain real-time requirements compared with the offline training.

[0065] In an embodiment of the present application, a data cache updating mechanism is proposed, in which the save_length parameter is used to control the size of the cache space. The most ideal case is to use the training frame data to train the neural network model when each training frame arrives, and to discard it after use. In this way, the consumption of storage space and computing resources is the least, but the problem is that the value of the data is not fully utilized, and the less data used in training, the greater the risk of network overfitting. However, save_length cannot be set too large without limitation. On the one hand, the device for online training is limited by cost, and the storage space and computing resources are limited. On the other hand, in some environments, the channel statistical characteristics may change quickly, and the historical cached training data may not conform to the current channel characteristics, and may not help the training of the neural network model, or even have a negative effect. By controlling the size of save_length, old data that does not conform to the current channel characteristics can be discarded in time, and the effect of online training can be ensured.

[0066] The management of the cache space is based on the FIFO strategy, i.e. each time new training data arrives, the new data is added to the cache space, and the earliest data in time is deleted from the cache space.

[0067] In an embodiment, the method of data augmentation includes at least one of the following: random phase method; constellation symmetry method; resource symmetry method; bandwidth segmentation method. In an example, the number of samples for online training is much less than that for offline training, and the purpose of data augmentation is to derive more data from the original data frame (original training data) for model training. For example, in the field of computer vision (CV), the method of picture rotation is often used to generate new pictures to enhance the network's ability to identify pictures. In wireless communication, the key point of data augmentation is that the derived new data and the original data have the same statistical characteristics, i.e., the same time-frequency domain correlation, the same SNR, etc.

[0068] In an embodiment, the implementation of the random phase method for data augmentation includes: performing multiplication operation of random phase factor based on the original data frame and the signal transceiver relationship on each resource element (RE) to obtain a new signal transceiver relationship; obtaining the corresponding target data frame based on the new signal transceiver relationship.

[0069] In an example, the random phase method is a data augmentation method using knowledge in the field of communication. Assuming that the training frame data bit is b, the transmitted data after modulation is x, the channel is h, the noise is n, and the received signal is y, then for an OFDM system, the signal transceiver relationship on each RE can be: y = hx + n.

[0070] In the signal transceiver relationship, multiply the left and right sides by a random phase factor e jω , to obtain a new signal transceiver relationship: y jω = he jω ·x + ne jω , which can be equivalent to: y new = h new ·x + n new , wherein the random phase factor and the original channel h are combined to generate a new channel h new , h new and the original channel h only differ by a phase factor, and the statistical characteristics do not change at all. Similarly, n new and the original noise n only have a phase change, and the SNR does not change. The new signal transceiver relationship y new = h new ·x + n new can be used to generate a target data frame as a new training sample.

[0071] Two points need to be noted in the random phase method, one is that the label b of the training sample generated by the random phase method is consistent with the original data, because x in the equation does not change; the second is that the number of training samples generated by the random phase method is not the more the better, when the number of samples is more, the correlation between the samples will be higher, which does not have substantial help for the training of the model. In the embodiment of the present application, the phase space of 2π can be equally divided and uniformly sampled to generate training samples, and the value of N is determined according to the actual training effect. FIG. 6 is an implementation schematic diagram of data enhancement using the random phase method according to an embodiment of the present application. As shown in FIG. 6, when N = 8, the effect of the random phase method is as shown in FIG. 6.

[0072] In an example, the constellation symmetry method is a data enhancement method designed by utilizing the symmetry characteristics of the constellation, specifically, the symmetry of the constellation includes but is not limited to:

[0073] Rotational symmetry, that is, the original constellation is completely coincident in position with the original constellation after being rotated by a certain angle (0 < ω < 2π), which means that the data after rotation is still legal and interpretable data;

[0074] Axial symmetry, that is, the original constellation is flipped with a straight line as the axis, and the flipped constellation is completely coincident in position with the original constellation, which means that the data after flipping is still legal and interpretable data;

[0075] The two methods are described in detail. FIG. 7 is an implementation schematic diagram of data enhancement using the constellation symmetry method according to an embodiment of the present application. For rotational symmetry, taking a 16QAM constellation as an example, the constellation has π / 2, π, 3π / 2 rotational symmetry, that is, when the constellation is rotated by π / 2, π, 3π / 2, the positions of the constellation points are completely coincident with the original constellation, as shown in FIG. 7.

[0076] In an embodiment, the implementation manner of data enhancement using the constellation symmetry method includes: performing a multiplication operation of a rotational symmetry factor based on an original data frame and a signal transceiving relationship on each resource element, to obtain a new signal transceiving relationship; obtaining a corresponding target data frame based on the new signal transceiving relationship. Assuming that the training frame data bit is b, the transmitted data after modulation is x, the channel is h, the noise is n, and the received signal is y, then for an OFDM system, the signal transceiving relationship on each RE can be expressed as the following equation: y = hx + n; the signal transceiving relationship is multiplied by the rotational symmetry factor e jπ / 2 , e jπ and e j3π / 2 , to obtain a new signal transceiving relationship:

[0077] As shown above, the rotational symmetry factor and the original x can be combined to generate new data x. new Compared to the original signal transmission and reception formula, h remains unchanged, and x... new The data is still valid, n new Compared to the original n, only the phase changes, and the SNR remains unchanged. Now, using the new signal-to-receive relationship y... new =h·x new +n new This allows the generation of target data frames, which can then be used as new training samples.

[0078] Based on the pilotless OFDM system intelligent receiver model, pilotless transmission must use a constellation diagram with rotational uniqueness to modulate data. Therefore, the constellation diagram does not have rotational symmetry and cannot be used for data enhancement.

[0079] Figure 8 is a schematic diagram illustrating another implementation of data augmentation using constellation diagram symmetry provided in this application embodiment. Regarding axial symmetry, taking a 16QAM constellation as an example, the constellation diagram has symmetry about the x-axis, meaning that when the constellation diagram is flipped around the x-axis (i.e., the data is conjugate), the positions of the constellation points completely coincide with the original constellation diagram, as shown in Figure 8.

[0080] In one embodiment, the implementation of data augmentation using the constellation diagram symmetry method includes: taking conjugate values ​​based on the original data frame and the signal transmission / reception relationship on each resource unit to obtain a new signal transmission / reception relationship; and obtaining the corresponding target data frame based on the new signal transmission / reception relationship. Assuming the training frame data bits are b, the modulated transmitted data is x, the channel is h, the noise is n, and the received signal is y, then for an OFDM system, the signal transmission / reception relationship on each RE can be expressed as the following equation: y = hx + n; taking conjugate values ​​on both sides of this signal transmission / reception relationship simultaneously yields a new signal transmission / reception relationship: y * =h * x * +n * As shown above, a new signal transmission and reception relationship was generated by taking the conjugate. Compared with the original signal transmission and reception relationship, h * The statistical properties of x have not changed compared to the original h. * The data is still valid, n * Compared to the original n, the SNR has not changed. Now, using equation y... * =h * x * +n * This allows the generation of target data frames, which can then be used as new training samples.

[0081] In an embodiment, the implementation of data enhancement using the resource symmetry method includes: reversing the order of the original data frame on the orthogonal frequency division multiplexing resource grid to generate the corresponding target data frame. The resource symmetry method is a data enhancement algorithm designed using the symmetry of the OFDM resource. Compared with the random phase method and the constellation symmetry method for data enhancement, the resource symmetry method is a more macroscopic method for data enhancement. Specifically, the original data frame is reversed in order on the OFDM resource grid to generate a new data frame as the target data frame.

[0082] In an embodiment, the implementation of reversing the order includes: frequency domain mirroring; time domain mirroring; time-frequency domain mirroring. In an example, the frequency domain mirroring is to reverse the order of the subcarriers; the time domain mirroring is to reverse the order of the OFDM symbols; and the time-frequency domain mirroring is to reverse the order in the frequency domain and the time domain at the same time.

[0083] In an embodiment, for a pilot-based orthogonal frequency division multiplexing system, the original data frame is reversed in order on the orthogonal frequency division multiplexing resource grid to generate the corresponding target data frame, including one of the following:

[0084] If the pilot pattern has time domain symmetry, the original data frame is reversed in order on the orthogonal frequency division multiplexing resource grid using time domain mirroring to generate the corresponding target data frame;

[0085] If the pilot pattern has frequency domain symmetry, the original data frame is reversed in order on the orthogonal frequency division multiplexing resource grid using frequency domain mirroring to generate the corresponding target data frame.

[0086] FIG. 9 is an implementation diagram of data enhancement using the resource symmetry method according to an embodiment of the present application. For a pilot-based OFDM system, the principle is shown in FIG. 9. p1, p2, p3 and p4 in the original resource grid are frequency domain resources, and s1, s2, …, s31 and s32 in the original resource grid are time domain resources. A new resource grid can be used to generate a new training sample. It should be noted that the resource symmetry method requires that the resource grid has certain symmetry characteristics, and more specifically, the pilot pattern has symmetry. In the case where the pilot pattern has time domain symmetry, time domain mirroring can be used to generate a target data frame as a new training sample. In the case where the pilot pattern has frequency domain symmetry, frequency domain mirroring can be used to generate a target data frame as a new training sample.

[0087] In an embodiment, for a session initiation protocol-based orthogonal frequency division multiplexing system, the original data frame is reversed in order on the orthogonal frequency division multiplexing resource grid to generate the corresponding target data frame, including one of the following:

[0088] If the weighting factor matrix has time domain symmetry, the original data frame is reversed in order on the orthogonal frequency division multiplexing resource grid by time domain mirroring to generate a corresponding target data frame.

[0089] If the weighting factor matrix has frequency domain symmetry, the original data frame is reversed in order on the orthogonal frequency division multiplexing resource grid by frequency domain mirroring to generate a corresponding target data frame.

[0090] FIG. 10 is another implementation schematic diagram of the resource symmetry method for data enhancement provided by the embodiment of the present application. For the SIP-based OFDM system, the weighting factor matrix (for example, the A matrix) is required to have certain symmetry, as shown in FIG. 10. In the case that the A matrix has time domain symmetry, the target data frame can be generated by using time domain mirroring as a new training sample; in the case that the A matrix has frequency domain symmetry, the target data frame can be generated by using frequency domain mirroring as a new training sample.

[0091] For the OFDM system without pilot, the resource grid naturally has symmetry in time domain and frequency domain, and the method can be directly used to generate the mirrored resource grid to generate the target data frame as a new training sample.

[0092] In an embodiment, the implementation manner of the bandwidth segmentation method for data enhancement includes: performing resource sliding segmentation based on the transmission bandwidth of the original data frame and the processing bandwidth of the neural network model to obtain a corresponding resource grid; and generating a corresponding target data frame based on each resource grid. In an example, the bandwidth segmentation method is a data enhancement method for the case that the training frame bandwidth is greater than the AI receiver bandwidth, and is also a macroscopic data enhancement method. Assuming that the training frame bandwidth is M resource blocks (RBs), the AI receiver bandwidth is B (the receiver bandwidth refers to the bandwidth size of the data sent into the receiver at a time), and M > B, the resource is sliding segmented on the bandwidth M in units of B, and the target data frame is generated by using the segmented resource grid as a new training sample. FIG. 11 is an implementation schematic diagram of the bandwidth segmentation method for data enhancement provided by the embodiment of the present application. As shown in FIG. 11, the training frame bandwidth is 8 RBs, and the AI receiver bandwidth is 3 RBs, and the resource is sliding segmented on RB0-RB7 to obtain 6 resource grids, which are block0-block5. In FIG. 11, M = 8 and B = 3, and a total of 6 resource grids conforming to the AI receiver bandwidth are segmented, which can be used to generate training samples.

[0093] In an example, the random phase method, the constellation symmetry method, the resource symmetry method and the bandwidth segmentation method can be independent of each other, or can be combined with each other to generate more training samples.

[0094] In an embodiment, the data grouping manner includes one of the following: randomly grouping into two groups; grouping into two groups according to the receiving order of the original data frames; and grouping by using different data enhancement methods. The data grouping is to divide the enhanced training samples into two groups according to a certain method, and the principle of grouping is that the lower the correlation of the two groups of data, the better. The grouping method includes: randomly grouping into two groups; grouping into two groups according to the order of arrival of the training frames; and grouping into two groups according to whether a certain data enhancement algorithm is used, for example, data obtained by using an axis symmetry enhancement algorithm can be taken as one group, and data obtained without using the axis symmetry enhancement algorithm can be taken as another group.

[0095] In an embodiment, the online training of the neural network model based on the cross-validation training algorithm and the two groups of target data includes: dividing a first group of target data in the two groups of target data into a training group, and dividing a second group of target data into a validation group; verifying the neural network model using the target data in the validation group to obtain a corresponding initial model performance; iteratively training the neural network model using the target data in the training group, and verifying the current model performance of the neural network model updated once by iteration using the target data in the validation group; and performing online training based on a comparison result between the initial model performance and the current model performance.

[0096] In an embodiment, the online training based on the comparison result between the initial model performance and the current model performance includes:

[0097] If the current model performance is greater than the initial model performance, the online training of the neural network model is stopped.

[0098] If the current model performance is less than or equal to the initial model performance, the online training of the neural network model is continued until the current iteration number reaches a preset iteration number threshold.

[0099] In an embodiment, the online training based on the comparison result between the initial model performance and the current model performance includes:

[0100] If the current model performance is greater than the initial model performance, the current learning rate and the penalty factor of the neural network model are multiplied, and the next iteration training is performed until the current learning rate is less than a preset learning rate threshold.

[0101] In an example, the implementation process of the cross-training algorithm includes:

[0102] Step 1, defining two groups of target data as training group A and validation group B respectively, wherein the training group is used to perform training of the model and update the parameters, and the validation group is used to verify the performance of the model;

[0103] Step 2, calculating the initial model performance using the validation group data, denoted as p;

[0104] Step 3, assuming the maximum number of iterations is K, the number of iterations is increased by 1 each time the entire data of the training set is used to train, and the current model performance is verified using the validation set data, denoted as p new ;

[0105] Step 4, determine the size relationship between p and p new , if p new is greater than p, it means that the model may have been over-fitted, and the model training is stopped; if p new is less than p, it means that the performance of the model is normally improved, p = p new , and the next iteration is continued.

[0106] Step 5, for the processing of d, another parallel solution is to determine the size relationship between p and p new , if p new is greater than p, it means that the model may have been over-fitted, but at this time the training is not stopped, only the current learning rate is multiplied by a penalty factor a less than 1, and the next iteration is continued. When the learning rate is less than a predetermined threshold (i.e., a predetermined learning rate threshold), the training is stopped.

[0107] Step 6, if the current number of iterations has reached a predetermined maximum value, the model training is stopped.

[0108] Step 7, optionally, the training set data and the validation set data can be role-exchanged, i.e., the original validation set data is used as the training set, and the original validation set is used as the training set to perform the training again.

[0109] FIG. 12 is an implementation flowchart of a cross-validation training algorithm according to an embodiment of the present application. As shown in FIG. 12, the implementation process of the cross-validation training algorithm includes:

[0110] Step 1, initialization: set the maximum number of iterations K, the training iteration number k = 0, and calculate the initial model performance p of the validation set.

[0111] Step 2, use the training set data to perform model training and parameter updating.

[0112] Step 3, calculate the performance p new of the validation set data.

[0113] Step 4, determine whether p is less than p new , if yes, execute step 5; if no, execute step 7.

[0114] Step 5, determine whether k is less than K, if yes, execute step 6, if no, execute step 7.

[0115] Step 6, k = k + 1, p = p new .

[0116] Step 7, training stops.

[0117] FIG. 13 is a flowchart of another implementation of the cross-validation training algorithm according to an embodiment of the present application. As shown in FIG. 13, the implementation of the cross-validation training algorithm includes the following steps:

[0118] Step 1, initialization: set the maximum number of iterations K, the number of training iterations k = 0, calculate the initial model performance p of the validation group data, the learning rate μ, the learning rate penalty factor β, and the iteration stop learning rate threshold μ_t.

[0119] Step 2, perform model training and parameter update using the training group data.

[0120] Step 3, calculate the performance p of the validation group data new .

[0121] Step 4, determine whether p is less than p new , if yes, execute step 5; if no, execute step 6.

[0122] Step 5, determine whether k is less than K, if yes, execute step 8, if no, execute step 9.

[0123] Step 6, μ = μ * β.

[0124] Step 7, determine whether μ is greater than μ_t, if yes, execute step 5; if no, execute step 9.

[0125] Step 8, k = k + 1, p = p new .

[0126] Step 9, training stops.

[0127] One of the main features of online training is that the number of training samples is very limited. The number of samples in offline training is usually in the order of hundreds of thousands or millions, and the accuracy of the model can be focused on during the training process, but there is a risk of overfitting. However, the number of samples in offline training is very small, and even if various enhancement algorithms are used to expand the samples, the sample size is still very limited compared to offline training, and overfitting is likely to occur during training.

[0128] In actual offline training, too few iterations during model training can result in underfitting, and too many iterations can result in overfitting. How to set a proper number of iterations is a difficult problem, and the best number of iterations for different scenarios is not necessarily the same. It is best to have an adaptive method to control the number of iterations. In the training of neural networks, an early stopping iteration idea can avoid overfitting. The target data can be divided into a training group and a validation group. If the performance of the validation group is found to be declining, it means that there is a high risk of overfitting, and training needs to be stopped immediately or the current learning rate needs to be reduced, which effectively reduces the risk of overfitting while ensuring the effectiveness of model training, especially in the case of a large number of training group and validation group samples.

[0129] In an embodiment, the model training method applied to the first communication node further includes: dividing the first group of target data into a validation group and the second group of target data into a training group. The data of the training group and the data of the validation group can be exchanged, and the original validation group data is used as the training group, and the original validation group is used as the training group to perform training again. In this way, the value of the original validation group data is fully exploited, and online training can converge faster.

[0130] In an embodiment, FIG. 14 is a flowchart of another model training method provided by the embodiment of the application. The embodiment is applied to the case of online training in a wireless communication system. The embodiment can be performed by the second communication node. In an example, the first communication node can be the terminal side, and the corresponding second communication node can be the base station side; in an example, the first communication node can be the base station side, and the corresponding second communication node can be the terminal side. As shown in FIG. 14, the embodiment includes S210.

[0131] S210, sending an original data frame for neural network model training to the first communication node, so that the first communication node performs online training on the neural network model using the original data frame.

[0132] In an embodiment, the receiving frequency of the original data frame includes: periodic; aperiodic.

[0133] In an embodiment, the application range of the original data frame includes: a specific terminal; a specific cell.

[0134] In an embodiment, the sending mode of the original data frame includes: the second communication node actively sends to the first communication node; the first communication node requests the second communication node to send.

[0135] In an embodiment, the bandwidth of the original data frame is greater than or equal to the processing bandwidth of the neural network model.

[0136] In an embodiment, the neural network model comprises one of: a pilot-based orthogonal frequency division multiplexing system intelligent receiver model; a pilot-free orthogonal frequency division multiplexing system intelligent receiver model; and a session initiation protocol-based orthogonal frequency division multiplexing system intelligent receiver model.

[0137] In an embodiment, the first communication node performs online training of the neural network model using the original data frames, comprising:

[0138] The first communication node performs data cache updating and data enhancement on the original data frames to obtain target data frames;

[0139] The first communication node performs data grouping on the target data frames according to a data grouping manner to obtain two groups of target data.

[0140] The first communication node performs online training of the neural network model based on a cross-validation training algorithm and the two groups of target data.

[0141] In an embodiment, the implementation process of data cache updating comprises:

[0142] A cache space with a target cache space size is pre-configured;

[0143] The received new original data frames are stored in the cache space, and the original data frame with the earliest cache time in the cache space is deleted from the cache space.

[0144] In an embodiment, the data enhancement method comprises at least one of: a random phase method; a constellation symmetry method; a resource symmetry method; and a bandwidth segmentation method.

[0145] In an embodiment, the implementation of the random phase method for data enhancement comprises:

[0146] A random phase factor is multiplied based on the original data frame and the signal transmission relationship on each resource unit to obtain a new signal transmission relationship;

[0147] A corresponding target data frame is obtained based on the new signal transmission relationship.

[0148] In an embodiment, the implementation of the constellation symmetry method for data enhancement comprises:

[0149] A rotation symmetry factor is multiplied based on the original data frame and the signal transmission relationship on each resource unit to obtain a new signal transmission relationship;

[0150] A corresponding target data frame is obtained based on the new signal transmission relationship.

[0151] In an embodiment, the implementation of the constellation symmetry method for data enhancement comprises:

[0152] Conjugate values are obtained based on the original data frame and the signal-transmitting relationship on each resource unit, and a new signal-transmitting relationship is obtained.

[0153] A corresponding target data frame is obtained based on the new signal-transmitting relationship.

[0154] In an embodiment, the implementation manner of data enhancement by using the resource symmetry method includes:

[0155] The original data frame is reversed in order on the orthogonal frequency division multiplexing resource grid to generate a corresponding target data frame.

[0156] In an embodiment, the implementation manner of reversing the order includes: frequency domain mirroring; time domain mirroring; time-frequency domain mirroring.

[0157] In an embodiment, for a pilot-based orthogonal frequency division multiplexing system, the original data frame is reversed in order on the orthogonal frequency division multiplexing resource grid to generate a corresponding target data frame, including one of the following:

[0158] If the pilot pattern has time domain symmetry, the original data frame is reversed in order on the orthogonal frequency division multiplexing resource grid by using time domain mirroring to generate a corresponding target data frame;

[0159] If the pilot pattern has frequency domain symmetry, the original data frame is reversed in order on the orthogonal frequency division multiplexing resource grid by using frequency domain mirroring to generate a corresponding target data frame.

[0160] In an embodiment, for a session initiation protocol-based orthogonal frequency division multiplexing system, the original data frame is reversed in order on the orthogonal frequency division multiplexing resource grid to generate a corresponding target data frame, including one of the following:

[0161] If the weighting factor matrix has time domain symmetry, the original data frame is reversed in order on the orthogonal frequency division multiplexing resource grid by using time domain mirroring to generate a corresponding target data frame;

[0162] If the weighting factor matrix has frequency domain symmetry, the original data frame is reversed in order on the orthogonal frequency division multiplexing resource grid by using frequency domain mirroring to generate a corresponding target data frame.

[0163] In an embodiment, the implementation manner of data enhancement by using the bandwidth segmentation method includes:

[0164] The resource grid is obtained by performing resource sliding segmentation based on the transmission bandwidth of the original data frame and the processing bandwidth of the neural network model.

[0165] A corresponding target data frame is generated based on each resource grid.

[0166] In an embodiment, the data grouping manner comprises one of the following: grouping randomly into two groups; grouping into two groups according to the receiving order of the original data frames; and grouping by using different data enhancement methods.

[0167] In an embodiment, the first communication node performs online training on the neural network model based on the cross-validation training algorithm and the two groups of target data, comprising:

[0168] The first communication node divides a first group of target data in the two groups of target data into a training group, and divides a second group of target data into a validation group;

[0169] The first communication node uses the target data in the validation group to verify the neural network model, and obtains a corresponding initial model performance;

[0170] The first communication node uses the target data in the training group to perform iterative training on the neural network model, and uses the target data in the validation group to verify the current model performance of the neural network model updated once in each iteration;

[0171] The first communication node performs online training based on a comparison result between the initial model performance and the current model performance.

[0172] In an embodiment, the first communication node performs online training based on a comparison result between the initial model performance and the current model performance, comprising:

[0173] If the current model performance is greater than the initial model performance, the online training of the neural network model is stopped;

[0174] If the current model performance is less than or equal to the initial model performance, the online training of the neural network model is continued until the current iteration number reaches a preset iteration number threshold.

[0175] In an embodiment, the first communication node performs online training based on a comparison result between the initial model performance and the current model performance, comprising:

[0176] If the current model performance is greater than the initial model performance, the current learning rate and the penalty factor of the neural network model are multiplied, and the next iteration training is performed until the current learning rate is less than a preset learning rate threshold.

[0177] In an embodiment, FIG. 15 is a structural block diagram of a model training device provided by an embodiment of the present application. The present embodiment is applied to a first communication node. As shown in FIG. 15, the model training device in the present embodiment comprises a receiver 310 and a training module 320.

[0178] The receiver 310 is configured to receive an original data frame sent by a second communication node for neural network model training;

[0179] The training module 320 is configured to perform online training of the neural network model by using the original data frame.

[0180] In an embodiment, the receiving frequency of the original data frame includes: periodicity; non-periodicity.

[0181] In an embodiment, the application range of the original data frame includes: a specific terminal; a specific cell.

[0182] In an embodiment, the sending mode of the original data frame includes: the second communication node actively sends to the first communication node; the first communication node requests the second communication node to send.

[0183] In an embodiment, the bandwidth of the original data frame is greater than or equal to the processing bandwidth of the neural network model.

[0184] In an embodiment, the neural network model includes one of: a pilot-based orthogonal frequency division multiplexing system intelligent receiver model; a pilot-free orthogonal frequency division multiplexing system intelligent receiver model; a session initiation protocol-based orthogonal frequency division multiplexing system intelligent receiver model.

[0185] In an embodiment, the online training of the neural network model by using the original data frame includes:

[0186] Data cache updating and data enhancement are performed on the original data frame to obtain a target data frame;

[0187] The target data frame is grouped according to a data grouping manner to obtain two groups of target data.

[0188] The neural network model is trained online based on a cross-validation training algorithm and the two groups of target data.

[0189] In an embodiment, the implementation process of data cache updating includes:

[0190] A cache space with a target cache space size is pre-configured;

[0191] The received new original data frame is stored in the cache space, and the original data frame with the earliest cache time in the cache space is deleted from the cache space.

[0192] In an embodiment, the data enhancement method includes at least one of: a random phase method; a constellation diagram symmetry method; a resource symmetry method; a bandwidth segmentation method.

[0193] In an embodiment, the implementation of the random phase method for data enhancement includes:

[0194] A random phase factor multiplication operation is performed based on the original data frame and the signal transmission relationship on each resource unit to obtain a new signal transmission relationship.

[0195] Based on the new signal transceiver relationship formula, the corresponding target data frame is obtained.

[0196] In an embodiment, the implementation mode of data enhancement by constellation symmetry method includes:

[0197] Based on the original data frame and the signal transceiver relationship formula on each resource unit, a multiplication operation of the rotational symmetry factor is performed to obtain a new signal transceiver relationship formula;

[0198] Based on the new signal transceiver relationship formula, the corresponding target data frame is obtained.

[0199] In an embodiment, the implementation mode of data enhancement by constellation symmetry method includes:

[0200] Based on the original data frame and the signal transceiver relationship formula on each resource unit, a conjugate value is taken to obtain a new signal transceiver relationship formula;

[0201] Based on the new signal transceiver relationship formula, the corresponding target data frame is obtained.

[0202] In an embodiment, the implementation mode of data enhancement by resource symmetry method includes:

[0203] The original data frame is reversed in order on the orthogonal frequency division multiplexing resource grid to generate the corresponding target data frame.

[0204] In an embodiment, the implementation mode of reversing the order includes: frequency domain mirroring; time domain mirroring; time-frequency domain mirroring.

[0205] In an embodiment, for a pilot-based orthogonal frequency division multiplexing system, the original data frame is reversed in order on the orthogonal frequency division multiplexing resource grid to generate the corresponding target data frame, including one of the following:

[0206] If the pilot pattern has time domain symmetry, time domain mirroring is used to reverse the order of the original data frame on the orthogonal frequency division multiplexing resource grid to generate the corresponding target data frame;

[0207] If the pilot pattern has frequency domain symmetry, frequency domain mirroring is used to reverse the order of the original data frame on the orthogonal frequency division multiplexing resource grid to generate the corresponding target data frame.

[0208] In an embodiment, for a session initiation protocol-based orthogonal frequency division multiplexing system, the original data frame is reversed in order on the orthogonal frequency division multiplexing resource grid to generate the corresponding target data frame, including one of the following:

[0209] If the weighting factor matrix has time domain symmetry, time domain mirroring is used to reverse the order of the original data frame on the orthogonal frequency division multiplexing resource grid to generate the corresponding target data frame;

[0210] If the weighting factor matrix has frequency domain symmetry, the original data frame is reversed in order on the orthogonal frequency division multiplexing resource grid using frequency domain mirroring to generate the corresponding target data frame.

[0211] In an embodiment, the implementation mode of data enhancement using bandwidth segmentation method includes:

[0212] Based on the transmission bandwidth of the original data frame and the processing bandwidth of the neural network model, resource sliding segmentation is performed to obtain the corresponding resource grid;

[0213] Based on each resource grid, the corresponding target data frame is generated.

[0214] In an embodiment, the data grouping mode includes one of the following: randomly divided into two groups; divided into odd and even groups according to the receiving order of the original data frame; and grouped using different data enhancement methods.

[0215] In an embodiment, the neural network model is trained online based on the cross-validation training algorithm and the two groups of target data, including:

[0216] The first group of target data in the two groups of target data is divided into a training group, and the second group of target data is divided into a validation group;

[0217] The target data in the validation group is used to verify the neural network model to obtain the corresponding initial model performance;

[0218] The target data in the training group is used to iteratively train the neural network model, and the target data in the validation group is used to verify the current model performance of the neural network model updated once by iteration;

[0219] Based on the comparison result between the initial model performance and the current model performance, online training is performed.

[0220] In an embodiment, the online training based on the comparison result between the initial model performance and the current model performance includes:

[0221] If the current model performance is greater than the initial model performance, the online training of the neural network model is stopped;

[0222] If the current model performance is less than or equal to the initial model performance, the online training of the neural network model is continued until the current iteration number reaches the preset iteration number threshold.

[0223] In an embodiment, the online training based on the comparison result between the initial model performance and the current model performance includes:

[0224] If the current model performance is greater than the initial model performance, the current learning rate and the penalty factor of the neural network model are multiplied, and the next iteration training is performed until the current learning rate is less than the preset learning rate threshold.

[0225] In an embodiment, the model training apparatus applied to the first communication node further includes:

[0226] The dividing module is configured to divide the first group of target data into a verification group and divide the second group of target data into a training group.

[0227] The model training apparatus provided in the embodiment is configured to implement the model training method applied to the first communication node in the embodiment shown in FIG. 1. The model training apparatus provided in the embodiment has similar implementation principles and technical effects, and thus will not be described here.

[0228] In an embodiment, FIG. 16 is a structural block diagram of another model training apparatus provided in an embodiment of the present application. The embodiment is applied to a second communication node. As shown in FIG. 16, the model training apparatus in the embodiment includes a transmitter 410.

[0229] The transmitter 410 is configured to send an original data frame used for neural network model training to the first communication node, so that the first communication node performs online training on the neural network model by using the original data frame.

[0230] In an embodiment, the receiving frequency of the original data frame includes periodicity and aperiodicity.

[0231] In an embodiment, the application range of the original data frame includes a specific terminal and a specific cell.

[0232] In an embodiment, the sending mode of the original data frame includes that the second communication node actively sends to the first communication node and that the first communication node requests the second communication node to send.

[0233] In an embodiment, the bandwidth of the original data frame is greater than or equal to the processing bandwidth of the neural network model.

[0234] In an embodiment, the neural network model includes one of the following: a pilot-based orthogonal frequency division multiplexing system intelligent receiver model, a pilot-free orthogonal frequency division multiplexing system intelligent receiver model, and a session initiation protocol-based orthogonal frequency division multiplexing system intelligent receiver model.

[0235] In an embodiment, the first communication node performs online training on the neural network model by using the original data frame, including:

[0236] The first communication node performs data cache updating and data enhancement on the original data frame to obtain a target data frame.

[0237] The first communication node data packets the target data frame to obtain two groups of target data according to a data packet mode;

[0238] The first communication node trains the neural network model based on the cross-validation training algorithm and the two groups of target data.

[0239] In an embodiment, the implementation process of the data cache update comprises:

[0240] A cache space with a target cache space size is pre-configured;

[0241] The received new original data frame is stored in the cache space, and the original data frame with the earliest cache time in the cache space is deleted from the cache space.

[0242] In an embodiment, the data enhancement method comprises at least one of the following: a random phase method, a constellation symmetry method, a resource symmetry method, and a bandwidth segmentation method.

[0243] In an embodiment, the implementation of the random phase method for data enhancement comprises:

[0244] Based on the original data frame and the signal transceiving relationship on each resource unit, a multiplication operation of a random phase factor is performed to obtain a new signal transceiving relationship;

[0245] The corresponding target data frame is obtained based on the new signal transceiving relationship.

[0246] In an embodiment, the implementation of the constellation symmetry method for data enhancement comprises:

[0247] Based on the original data frame and the signal transceiving relationship on each resource unit, a multiplication operation of a rotational symmetry factor is performed to obtain a new signal transceiving relationship;

[0248] The corresponding target data frame is obtained based on the new signal transceiving relationship.

[0249] In an embodiment, the implementation of the constellation symmetry method for data enhancement comprises:

[0250] Based on the original data frame and the signal transceiving relationship on each resource unit, a conjugate value is obtained to obtain a new signal transceiving relationship;

[0251] The corresponding target data frame is obtained based on the new signal transceiving relationship.

[0252] In an embodiment, the implementation of the resource symmetry method for data enhancement comprises:

[0253] The original data frame is reversed in order on an orthogonal frequency division multiplexing resource grid to generate a corresponding target data frame.

[0254] In an embodiment, the implementation of reversing order includes: frequency domain mirroring; time domain mirroring; time-frequency domain mirroring.

[0255] In an embodiment, for a pilot-based OFDM system, the original data frame is reversed on the OFDM resource grid to generate the corresponding target data frame, including one of the following:

[0256] If the pilot pattern has time domain symmetry, the original data frame is reversed on the OFDM resource grid using time domain mirroring to generate the corresponding target data frame;

[0257] If the pilot pattern has frequency domain symmetry, the original data frame is reversed on the OFDM resource grid using frequency domain mirroring to generate the corresponding target data frame.

[0258] In an embodiment, for a pilot-based OFDM system, the original data frame is reversed on the OFDM resource grid to generate the corresponding target data frame, including one of the following:

[0259] If the pilot pattern has time domain symmetry, the original data frame is reversed on the OFDM resource grid using time domain mirroring to generate the corresponding target data frame;

[0260] If the pilot pattern has frequency domain symmetry, the original data frame is reversed on the OFDM resource grid using frequency domain mirroring to generate the corresponding target data frame.

[0261] In an embodiment, the implementation of data enhancement using bandwidth segmentation includes:

[0262] Based on the transmission bandwidth of the original data frame and the processing bandwidth of the neural network model, resource sliding segmentation is performed to obtain the corresponding resource grid;

[0263] Based on each resource grid, the corresponding target data frame is generated.

[0264] In an embodiment, the data grouping method includes one of the following: randomly divided into two groups; divided into odd and even groups according to the receiving order of the original data frame; and grouped using different data enhancement methods.

[0265] In an embodiment, the first communication node trains the neural network model based on cross-validation training algorithm and two groups of target data, including:

[0266] The first communication node divides the first group of target data in the two groups of target data into a training group, and the second group of target data into a validation group;

[0267] The first communication node verifies the neural network model by using the target data in the verification set to obtain an initial model performance;

[0268] The first communication node iteratively trains the neural network model by using the target data in the training set, and verifies a current model performance of the neural network model updated once by using the target data in the verification set;

[0269] The first communication node performs online training based on a comparison result between the initial model performance and the current model performance.

[0270] In an embodiment, the first communication node performs online training based on a comparison result between the initial model performance and the current model performance, including:

[0271] If the current model performance is greater than the initial model performance, the online training of the neural network model is stopped;

[0272] If the current model performance is less than or equal to the initial model performance, the online training of the neural network model is continued until a preset iteration threshold is reached.

[0273] In an embodiment, the first communication node performs online training based on a comparison result between the initial model performance and the current model performance, including:

[0274] If the current model performance is greater than the initial model performance, the current learning rate and the penalty factor of the neural network model are multiplied, and the next iteration training is performed until the current learning rate is less than a preset learning rate threshold.

[0275] The model training apparatus provided in the embodiment is configured to implement the model training method applied to the second communication node in the embodiment shown in FIG. 14, and the model training apparatus provided in the embodiment has similar implementation principles and technical effects, which will not be described herein again.

[0276] In an embodiment, FIG. 17 is a structural schematic diagram of a communication device provided in an embodiment of the present application. As shown in FIG. 17, the device provided in the present application includes a processor 510, a memory 520, and a communication module 530. The number of processors 510 in the device can be one or more, and one processor 510 is taken as an example in FIG. 17. The number of memories 520 in the device can be one or more, and one memory 520 is taken as an example in FIG. 17. The processor 510, the memory 520, and the communication module 530 of the device can be connected through a bus or other manners, and the connection through the bus is taken as an example in FIG. 17. In the embodiment, the device can be the first communication node or the second communication node.

[0277] The memory 520, as a computer readable storage medium, can be configured to store software programs, computer executable programs and modules, such as program instructions / modules of the device according to any embodiment of the present application (for example, the receiver 310 and the training module 320 in the model training apparatus applied to the first communication node). The memory 520 can include a program storage area and a data storage area, where the program storage area can store an operating system and at least one application required by a function; and the data storage area can store data created during use of the device, etc. In addition, the memory 520 can include a high-speed random access memory, and can also include a nonvolatile memory, such as at least one magnetic disk storage device, a flash memory device, or other nonvolatile solid-state memory device. In some examples, the memory 520 can further include a memory disposed remotely from the processor 510, which can be connected to the device through a network. Examples of the network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0278] In the case where the communication device is the first communication node, the device provided above can be configured to perform the model training method applied to the first communication node provided by any embodiment above, and has the corresponding functions and effects.

[0279] In the case where the communication device is the second communication node, the device provided above can be configured to perform the model training method applied to the second communication node provided by any embodiment above, and has the corresponding functions and effects.

[0280] The embodiment of the present application further provides a storage medium containing computer executable instructions, which, when executed by a computer processor, are used to perform a model training method applied to a first communication node, the method comprising: receiving original data frames for neural network model training sent by a second communication node; and performing online training on the neural network model by using the original data frames.

[0281] The embodiment of the present application further provides a storage medium containing computer executable instructions, which, when executed by a computer processor, are used to perform a model training method applied to a second communication node, the method comprising: sending original data frames for neural network model training to a first communication node, so that the first communication node performs online training on the neural network model by using the original data frames.

[0282] Those skilled in the art will appreciate that the term user equipment encompasses any suitable type of wireless user equipment, such as a mobile phone, a portable data processing apparatus, a portable web browser, or a vehicle-mounted mobile station.

[0283] In general, the various embodiments of the application can be implemented in hardware or special-purpose circuits, software, logic or any combination thereof. For example, some aspects can be implemented in hardware, while other aspects can be implemented in

[0284] Embodiments of the application can be implemented by a data processor of a mobile device executing computer program instructions, for example in a processor entity, or by hardware, or by a combination of software and hardware. Computer program instructions can be in assemblies, Instruction Set Architecture (ISA), machine, machine-related, microcode, firmware, state-setting data, or source code or object code written in any combination of one or more programming languages to be executed by or to configure the data processor.

[0285] Any flowchart of the figures herein can represent program steps, or can represent interconnected logic circuits, modules, and functions, or can represent a combination of program steps and logic circuits, modules, and functions. The computer program can be stored on a memory. The memory can be of any type suitable to the local technical environment and can be implemented using any suitable data storage technology, such as, but not limited to, random access memory (RAM), read-only memory (ROM), optical storage devices, and systems such as digital video disc (DVD) or compact disc (CD), and the like. The computer readable media can include non-transitory storage media. The data processor can be of any type suitable to the local technical environment, and can include, for example, without limitation, a general purpose computer, a special purpose computer, a microprocessor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), and processors based on a multi-core processor architecture, as examples.

[0286] The embodiments of the application further provide a computer program product, comprising a computer program which, when executed by a processor, can implement the model training method provided by any of the embodiments of the application.

[0287] Computer program product, in the implementation of the process, can be written in one or more programming language or its combination for executing computer program code of the operation of the present application, programming language includes object-oriented programming language, such as Java, Smalltalk, C++, also includes conventional process type programming language, such as "C" language or similar programming language. Program code can be completely executed on user computer, partially executed on user computer, executed as an independent software package, partially executed on user computer and partially on remote computer, or completely executed on remote computer or server. In the case of remote computer, remote computer can be connected to user computer through any kind of network, including Local Area Network (LAN) or Wide Area Network (WAN), or can be connected to external computer (for example, using Internet service provider to connect through Internet).

[0288] The above only is the preferred embodiment of the present application, and is not used to limit the present application, for the person skilled in the art, the present application can have various changes and changes. Any modification, equivalent replacement, improvement, etc. within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A model training method applied to a first communication node, comprising: receiving an original data frame sent by a second communication node for neural network model training; training the neural network model online using the original data frame.

2. The method of claim 1, wherein, The receiving frequency of the original data frame includes: periodicity; non-periodicity.

3. The method of claim 1, wherein, The application range of the original data frame includes: a specific terminal; a specific cell.

4. The method of claim 1, wherein, The sending mode of the original data frame includes: the second communication node actively sending to the first communication node; the first communication node requesting the second communication node to send.

5. The method of claim 1, wherein, The bandwidth of the original data frame is greater than or equal to the processing bandwidth of the neural network model.

6. The method of claim 1, wherein, The neural network model includes one of the following: a pilot-based orthogonal frequency division multiplexing system intelligent receiver model; a pilot-free orthogonal frequency division multiplexing system intelligent receiver model; a session initiation protocol-based orthogonal frequency division multiplexing system intelligent receiver model.

7. The method of claim 1, wherein, The training of the neural network model online using the original data frame includes: updating and enhancing the data of the original data frame to obtain a target data frame; grouping the target data frame according to a data grouping method to obtain two groups of target data; training the neural network model online based on a cross-validation training algorithm and the two groups of target data.

8. The method of claim 7, wherein, The implementation process of the data cache update includes: pre-configuring a cache space with a target cache space size; storing the received new original data frame in the cache space, and deleting the original data frame cached for the earliest time in the cache space from the cache space.

9. The method of claim 7, wherein, The data enhancement method includes at least one of the following: a random phase method; a constellation symmetry method; a resource symmetry method; a bandwidth segmentation method.

10. The method of claim 9, wherein, The implementation of the random phase method for data enhancement includes: multiplying a random phase factor based on the original data frame and the signal transmission relationship on each resource unit to obtain a new signal transmission relationship; obtaining the corresponding target data frame based on the new signal transmission relationship.

11. The method of claim 9, wherein, The implementation of the constellation symmetry method for data enhancement includes: multiplying a rotational symmetry factor based on the original data frame and the signal transmission relationship on each resource unit to obtain a new signal transmission relationship; obtaining the corresponding target data frame based on the new signal transmission relationship.

12. The method of claim 9, wherein, The implementation of the constellation symmetry method for data enhancement includes: taking a conjugate value based on the original data frame and the signal transmission relationship on each resource unit to obtain a new signal transmission relationship; obtaining the corresponding target data frame based on the new signal transmission relationship.

13. The method of claim 9, wherein, The implementation of the resource symmetry method for data enhancement includes: reversing the order of the original data frame on the orthogonal frequency division multiplexing resource grid to generate the corresponding target data frame.

14. The method of claim 13, wherein, The implementation of the reversed order includes: frequency domain mirroring; time domain mirroring; time-frequency domain mirroring.

15. The method of claim 13, wherein, For a pilot-based orthogonal frequency division multiplexing system, the reversing the order of the original data frame on the orthogonal frequency division multiplexing resource grid to generate the corresponding target data frame includes one of the following: in response to the pilot pattern having time domain symmetry, performing time domain mirroring on the original data frame on the OFDM resource grid to generate a corresponding target data frame; in response to the pilot pattern having frequency domain symmetry, performing frequency domain mirroring on the original data frame on the OFDM resource grid to generate a corresponding target data frame.

16. The method of claim 13, wherein, for a session initiation protocol based OFDM system, the reversing the order of the original data frame on the OFDM resource grid to generate a corresponding target data frame comprises one of: in response to the weighting factor matrix having time domain symmetry, performing time domain mirroring on the original data frame on the OFDM resource grid to generate a corresponding target data frame; in response to the weighting factor matrix having frequency domain symmetry, performing frequency domain mirroring on the original data frame on the OFDM resource grid to generate a corresponding target data frame.

17. The method of claim 9, wherein, The implementation of the bandwidth segmentation method for data enhancement includes: segmenting the resource based on the transmission bandwidth of the original data frame and the processing bandwidth of the neural network model to obtain a corresponding resource grid; generating a corresponding target data frame based on each of the resource grids.

18. The method of claim 7, wherein, The data grouping method includes one of the following: randomly divided into two groups; divided into odd and even groups according to the receiving order of the original data frame; and grouped using different data enhancement methods.

19. The method of claim 7, wherein, The online training of the neural network model based on the cross-validation training algorithm and the two groups of target data includes: dividing a first group of target data in the two groups of target data into a training group and a second group of target data into a validation group; verifying the neural network model using target data in the validation group to obtain a corresponding initial model performance; iteratively training the neural network model using target data in the training group and verifying the current model performance of the neural network model updated once by iteration using target data in the validation group; performing online training based on the comparison result between the initial model performance and the current model performance.

20. The method of claim 19, wherein, The online training based on the comparison result between the initial model performance and the current model performance includes: in response to the current model performance being greater than the initial model performance, stopping the online training of the neural network model; in response to the current model performance being less than or equal to the initial model performance, continuing the online training of the neural network model until the current iteration number reaches a preset iteration number threshold.

21. The method of claim 19, wherein, The online training based on the comparison result between the initial model performance and the current model performance includes: in response to the current model performance being greater than the initial model performance, multiplying the current learning rate and the penalty factor of the neural network model and performing next iteration training until the current learning rate is less than a preset learning rate threshold.

22. The method of any one of claims 19-21, the method further comprising: The first group of target data in the two groups of target data is divided into a verification group, and the second group of target data is divided into a training group.

23. A model training method applied to a second communication node, comprising: sending, to a first communication node, an original data frame for neural network model training, so that the first communication node adopts the original data frame to perform online training on the neural network model.

24. A communication device comprising: a memory, and one or more processors; the memory is configured to store one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the method in any one of claims 1-22 or 23.

25. A storage medium, the storage medium storing a computer program, the computer program being executed by a processor to implement the method in any one of claims 1-22 or 23.

Images