Learning programs, learning methods, information processing devices, communication devices, and communication systems

Abstract

The present invention provides a learning program, a learning method, an information processing device, a communication device, and a communication system that improve the throughput of communication systems. [Solution] For each of the multiple received signals, a first channel estimate is calculated, which is the channel estimation result. A loss weight corresponding to the communication state is calculated for each first channel estimate. The first channel estimate is input into a machine learning model to obtain a second channel estimate, which is a correction applied to the first channel estimate. The mean squared error between the second channel estimate and the correct channel estimate corresponding to the first channel estimate is calculated for each first channel estimate. A weighted loss function is determined according to the mean squared error and the loss weight. The computer is then instructed to perform a process to update the machine learning model using the weighted loss function.

Description

Technical Field

[0001] The present invention relates to a learning program, a learning method, an information processing apparatus, a communication apparatus, and a communication system.

Background Art

[0002] In a wireless communication system, communication is performed by a transmitter transmitting a modulated data signal and a receiver demodulating the received data signal. Here, for example, the transmitter may be a base station and the receiver may be a UE (User Equipment), or conversely, the transmitter may be a UE and the receiver may be a base station. In this communication, the phase and amplitude of the data signal change while being transmitted from the transmitter to the receiver. Therefore, the receiver performs detection and demodulation on the data signal whose phase and amplitude change through the channel using channel information indicating the time and frequency assigned to the data signal.

[0003] In order to demodulate the downlink data signal by the receiver, the transmitter transmits a demodulation reference signal (DMRS), which is a known training reference signal, to the receiver in advance. The receiver estimates the channel using the DMRS received from the transmitter, and detects and decodes the data signal based on the estimation result.

[0004] The DMRS has a fixed allocation amount and arrangement pattern on the radio resource. Also, the DMRS is sparsely multiplexed in terms of time and frequency in the radio resource. Here, the radio resource is represented by a table called a resource grid having frequency and time as two axes. An individual radio resource indicated by a specific time and a specific frequency in the resource grid is called a resource element (RE).

[0005] To detect and demodulate the channel of the received data signal, the receiver utilizes channel information for all resource elements of the incoming data signal. Therefore, the receiver estimates the channel of the received data signal by interpolating the channels in the DRMS ​​resource elements according to noise suppression and the wireless environment. Specifically, if numerous reflected waves are generated due to obstacles between the transmitter and receiver, the channel may easily change in the frequency direction. Also, if the receiver is moving, the channel may easily change in the time direction due to the receiver's speed. The receiver performs channel interpolation considering these conditions.

[0006] For example, a receiver performs interpolation by linearly weighting the channel estimates for the DMRS resource elements to estimate the channel of the entire data signal. Known interpolation methods include linear interpolation and MMSE (Minimum Mean Square Error) estimation. Channel statistics are used in estimation using the MMSE method.

[0007] Furthermore, in recent years, the use of AI (Artificial Intelligence) in channel estimation processing has been explored. For example, there is a technique that inputs channel estimates for DMRS resource elements obtained using LS (Least Square) estimation into a neural network, and outputs the channel estimate for the entire data signal. In addition, AI-based channel estimation can sometimes yield good results without using the channel statistics information used in the MMSE method.

[0008] Furthermore, when performing a handover in mobile communications, a technology has been proposed that determines link quality by channel estimation, evaluates the wireless conditions, and optimizes data throughput based on the evaluation results. [Prior art documents] [Patent Documents]

[0009] [Patent Document 1] Japanese Patent Publication No. 2023-106527 [Overview of the project] [Problems that the invention aims to solve]

[0010] However, simply using MSE as the loss function in AI-based channel estimation may reduce estimation accuracy. This is because, while it is preferable to consider various conditions such as SNR, channel variation in the time domain, and channel variation in the frequency domain in channel estimation, simply using MSE as the loss function may not adequately consider these conditions. If channel estimation accuracy deteriorates, it may ultimately lead to a decrease in the throughput of the communication system.

[0011] Furthermore, in the technology used to determine link quality through channel estimation during mobile communication handover, the loss function used in AI-based channel estimation is not taken into consideration, making it difficult to improve the estimation accuracy of AI-based channel estimation. Therefore, even with this technology, it is difficult to improve the throughput of the communication system.

[0012] The disclosed technology was made in view of the above and aims to provide a learning program, learning method, information processing device, communication device, and communication system that improve the throughput of a communication system. [Means for solving the problem]

[0013] In one aspect of the learning program, learning method, information processing apparatus, communication apparatus, and communication system disclosed in the present application, for each first channel estimation value that is an estimation result of a channel for each of a plurality of received signals, a loss weight corresponding to a communication state is calculated, the first channel estimation value is input into a machine learning model, and a second channel estimation value obtained by correcting the first channel estimation value is determined for each of the first channel estimation values. An average squared error between the second channel estimation value and a correct channel estimation value that is a correct channel estimation value corresponding to the first channel estimation value is obtained for each of the first channel estimation values. Based on the average squared error and the loss weight, a weighted loss function is determined, and a process of updating the machine learning model using the weighted loss function is executed by a computer.

Advantages of the Invention

[0014] In one aspect, the present invention can improve the throughput of a communication system.

Brief Description of the Drawings

[0015] [Figure 1] FIG. 1 is a block diagram of a machine learning system according to an embodiment. [Figure 2] FIG. 2 is a diagram showing an example of wireless resources. [Figure 3] FIG. 3 is a diagram showing an overview of channel estimation. [Figure 4] FIG. 4 is a diagram showing an overview of training of channel estimation AI. [Figure 5] FIG. 5 is a flowchart of a learning process of channel estimation AI by a machine learning apparatus according to an embodiment. [Figure 6] FIG. 6 is a diagram showing a comparison of channel estimation accuracies. [Figure 7] FIG. 7 is a hardware configuration diagram of an information processing apparatus. [Figure 8] FIG. 8 is a hardware configuration diagram of a receiver.

Modes for Carrying Out the Invention

[0016] Hereinafter, embodiments of the learning program, learning method, information processing apparatus, communication apparatus, and communication system disclosed in the present application will be described in detail based on the drawings. Note that the learning program, learning method, information processing apparatus, communication apparatus, and communication system disclosed in the present application are not limited by the following embodiments.

Embodiment

[0017] FIG. 1 is a block diagram of a machine learning system according to an embodiment. The communication system 100 includes a base station 2 and a machine learning apparatus 1. Here, an example in which the base station 2 is a receiver will be described.

[0018] The base station 2 communicates with a UE (not shown) accommodated in a cell. For example, the base station 2 operates as a receiver that receives a signal transmitted from the UE. The base station 2 includes a least square (LS) estimation unit 21, an inference unit 22, a demodulation unit 23, a decoding unit 24, and an output unit 25.

[0019] FIG. 2 is a diagram showing an example of radio resources. As shown in FIG. 2, the radio resources are represented by a resource grid 30 in which the vertical axis indicates frequency and the horizontal axis indicates time. In the radio resources, for example, 0.5 ms along the time axis is one slot, and a bandwidth of 360 kHz along the frequency axis is one resource block. Further, one slot includes 14 orthogonal frequency division multiplexing (OFDM) symbols. Also, one resource block includes 12 consecutive subcarriers. Each radio resource indicated by a subcarrier and an OFDM symbol is a resource element.

[0020] The radio resources include DMRS used for channel estimation, which is shown as a resource element 31 in the resource grid 30. Also, data is mainly stored in other resource elements in the resource grid 30.

[0021] Returning to Figure 1, let's continue the explanation. DMRS information (such as the location of DMRS resource elements and DMRS patterns) is shared with the UE in advance. The LS estimation unit 21 receives the signal transmitted from the UE via the antenna. The LS estimation unit 21 then obtains LS estimates for each DMRS resource element, using the DMRS in the received signal as a reference. The LS estimates are estimated channel values ​​obtained from the DMRS signal.

[0022] For example, let's explain using the SISO (Single Input Single Output) example. The received signal y(k,l) of the subcarrier #k and OFDM symbol #l on which DMRS is multiplexed is expressed by the following equation (1).

[0023]

number

[0024] Here, h(k,l), x DMRS (k,l) and n(k,l) represent the channel, DMRS, and noise in the radio resource of OFDM symbol #l for subcarrier #k, respectively.

[0025] DMRS is x DMRS Since (k,l) is a known signal in the LS estimation unit 21, the LS estimation unit 21 can calculate the LS estimate by dividing the received signal by the DMRS signal, as shown in the following equation (2).

[0026]

number

[0027] Subsequently, the LS estimation unit 21 outputs the LS estimate for each resource element to the inference unit 22. The channel estimation results for each resource block of the radio resource based on the DMRS obtained from the received signal, such as the LS estimate calculated by the LS estimation unit 21, are an example of the "first channel estimate" in the receiver. In other words, the LS estimation unit 21 calculates the first channel estimate based on the received signal.

[0028] The inference unit 22 has a channel estimation AI 26. The channel estimation AI 26 is a machine learning model that, for example, receives an LS estimate as input and outputs a channel estimate. The channel estimation AI 26 corrects the phase and amplitude of each channel in the LS estimate and calculates the channel estimate. The channel estimation AI 26 is a trained machine learning model that has been trained by the machine learning device 1 described later.

[0029] The inference unit 22 obtains the channel estimate output by the channel estimation AI 26. The inference unit 22 then outputs the channel estimate for each resource element to the demodulation unit 23. This channel estimation AI 26 is an example of a "trained machine learning model". The channel estimation AI 26 also outputs a second channel estimate obtained by applying amplitude and phase corrections to the first channel estimate. The inference unit 22 inputs the first channel estimate calculated by the channel estimation AI 26, which is a trained machine learning model, and obtains the second estimate output by the channel estimation AI 26.

[0030] Figure 3 shows an overview of channel estimation. The LS estimation unit 21 shares DMRS information, indicated by resource block 203 of the radio resource 201, with the UE from the received signal received from the UE. The LS estimation unit 21 and the inference unit 22 then use the DMRS information to interpolate the signals between DMRS and correct the phase and amplitude, thereby estimating the channels of all resource blocks included in the radio resource 201. As a result, the LS estimation unit 21 and the inference unit 22 can obtain channel estimates for all resource blocks, as shown in the radio resource 202.

[0031] Returning to Figure 1, the explanation continues. The demodulation unit 23 receives channel estimates for each resource element as input from the inference unit 22. The demodulation unit 23 also receives the received signal transmitted via the antenna. The demodulation unit 23 then demodulates the received signal using the channel estimates. The demodulation unit 23 outputs the demodulated received signal to the decoding unit 24.

[0032] The decoding unit 24 receives the demodulated received signal as input from the demodulation unit 23. Next, the decoding unit 24 decodes the received signal and obtains data. Then, the decoding unit 24 outputs the obtained data to the output unit 25.

[0033] The output unit 25 receives data input from the decoding unit 24. The output unit 25 then outputs the data to a higher-level device or other base station.

[0034] As shown in Figure 1, the machine learning device 1 includes a control unit 10, a data storage unit 11, a channel estimation AI 12, an MSE calculation unit 13, a weighted loss function calculation unit 14, and a loss weight calculation unit 15.

[0035] The data storage unit 11 contains training data 17. The training data 17 includes, for example, multiple LS estimate samples with different SNRs (Signal to Noise Ratios). That is, the training data 17 includes LS estimate samples, each being an example of an LS estimate obtained from a DMRS signal transmitted under different SNRs.

[0036] Furthermore, the training data 17 includes ground truth channel samples, which are the correct labels for the channel estimates for each LS estimate sample. The ground truth channel samples are the training data corresponding to the LS estimate samples, and a pair of an LS estimate sample and a ground truth channel sample constitutes one set of data used for learning. Here, the LS estimate sample is an example of a "first channel estimate," and the ground truth channel sample is an example of a "ground truth channel estimate."

[0037] In addition, the data storage unit 11 may, for example, store information on the SNR used in generating each LS estimate sample of the training data 17.

[0038] The control unit 10 acquires a predetermined number of LS estimate samples from the training data 17 stored in the data storage unit 11. The control unit 10 then inputs the acquired predetermined number of LS estimate samples into the channel estimation AI 12 to perform channel estimation and outputs a channel estimate sample, which is the result of channel estimation for each LS estimate sample.

[0039] Furthermore, the control unit 10 obtains a weighted loss function from the weighted loss function calculation unit 14, to which weights are assigned to each LS estimated value sample. The weighted loss function will be explained in detail later. Then, the control unit 10 provides the weighted loss function to the channel estimation AI 12, causing the channel estimation AI 12 to perform backpropagation and update its parameters. After that, the control unit 10 transmits the updated channel estimation AI 12 to the base station 2, where it is used for communication as the channel estimation AI 26. In this way, the control unit 10 updates the channel estimation AI 12, which is a machine learning model, using the weighted loss function.

[0040] The control unit 10 controls the channel estimation AI 12 to perform mini-batch learning for each predetermined number of acquired LS estimation samples. The control unit 10 randomly changes a predetermined number of LS estimation sample selected from the training data 17 and repeats the mini-batch learning. The learning process using a specific mini-batch will be described below.

[0041] Channel Estimation AI12 is a machine learning model that outputs a channel estimate for an input LS estimate. Before training, Channel Estimation AI12 is the machine learning model used by base station 2 when it actually performs communication, before it is trained.

[0042] The channel estimation AI 12 receives LS estimated value samples as input from the control unit 10. The channel estimation AI 12 then performs channel estimation by forward propagating each input LS estimated value sample. The channel estimation AI 12 then outputs the channel estimation result for each LS estimated value sample as a channel estimated value sample to the MSE calculation unit 13.

[0043] A sample of channel estimates obtained by channel estimation by the channel estimation AI 12 is an example of a "second channel estimate." That is, the control unit 10 inputs the first channel estimates into the channel estimation AI 12, which is a machine learning model, and causes it to determine a second channel estimate for each of the first channel estimates by applying corrections to the first channel estimates. The control unit 10 also inputs the first channel estimates, which are the channel estimation results for each resource block of the DMRS-based wireless resource obtained from multiple received signals, into the channel estimation AI 12. Then, the control unit 10 causes the channel estimation AI 12 to output the second channel estimates, which are the first channel estimates with corrections for amplitude and phase applied to them.

[0044] The MSE calculation unit 13 receives input for channel estimate samples for each LS estimate sample from the channel estimation AI 12. The MSE calculation unit 13 also obtains the correct channel sample corresponding to each LS estimate sample from the training data 17.

[0045] Next, the MSE calculation unit 13 calculates the MSE for each LS estimated value sample using the LS estimated value sample and the ground truth channel sample. For example, the MSE calculation unit 13 can calculate the MSE using the following formula (3). The MSE is the difference between the estimation result by the channel estimation AI 12 and the ground truth, and represents the loss in the estimation result. In other words, formula (3) can also be used as a loss function.

[0046]

number

[0047] Here, i is the sample number of the LS estimate sample in the minibatch. The sample numbers are assigned sequentially to the LS estimate samples. Also, N RE is the number of elements in the resource block. In this example, the AI ​​model estimates the channel value for one resource block, but it is also possible to use an AI model that estimates the channel value for multiple resource blocks. Also, h(k,l) is the ground truth channel sample for the resource element of subcarrier #k and OFDM symbol l. Also, h^(k,l) is the channel estimate obtained by the channel estimation AI12 for the resource element of subcarrier #k and OFDM symbol l (h^ indicates a symbol with a caret added above h).

[0048] Subsequently, the MSE calculation unit 13 outputs the MSE for each LS estimate sample to the weighted loss function calculation unit 14. In this way, by inputting the LS estimate, which is the first channel estimate, into the channel estimation AI 12, which is a machine learning model, the MSE calculation unit 13 obtains a second channel estimate, which is the first channel estimate obtained by the channel estimation AI 12 with corrections applied. Then, the MSE calculation unit 13 calculates the mean squared error between the channel estimate sample, which is the second channel estimate, and the ground truth channel estimate, which is the first channel estimate, corresponding to the LS estimate, for each of the first channel estimates.

[0049] The loss weight calculation unit 15 obtains information from the control unit 10 about each LS estimated value sample used for estimation by the channel estimation AI 12. Then, for each LS estimated value sample used for estimation by the channel estimation AI 12, the loss weight calculation unit 15 calculates a loss weight sample that represents a weight according to the communication state. A weight according to the communication state is, for example, a weight that is larger when the communication state is good and smaller when the communication state is poor. The loss weight 16 in Figure 1 shows a summary of the loss weight samples for each LS estimated value sample.

[0050] The loss weight calculation unit 15 can, for example, use SNR or MSE calculated using the MMSE method as loss weights, depending on the communication state. Alternatively, the loss weight calculation unit 15 can also perform loss weight calculations using functions corresponding to channel characteristics such as SNR, delay spread, and maximum Doppler frequency, depending on the communication state.

[0051] For example, the loss weight calculation unit 15 obtains SNR information for each LS estimated value sample used in the estimation from the data storage unit 11, based on the information of the LS estimated value samples. Then, the loss weight calculation unit 15 uses the SNR corresponding to each LS estimated value sample as the loss weight for each LS estimated value sample.

[0052] Furthermore, for example, the loss weight calculation unit 15 obtains MSE information obtained by the MMSE method for each LS estimated value sample used in the estimation from the MSE calculation unit 13, based on the information of the LS estimated value samples. The loss weight calculation unit 15 then uses the reciprocal of the MSE obtained by the MMSE method corresponding to each LS estimated value sample as the loss weight for each LS estimated value sample. The magnitude of the loss based on MSE is affected by the degree of difficulty in channel estimation according to the communication state. Specifically, under conditions where channel estimation is difficult, the MSE becomes large, and the loss based on MSE also becomes large. Conversely, under conditions where channel estimation is easy, the MSE becomes small, and the loss based on MSE also becomes small. Therefore, the reciprocal of the MSE obtained by the MMSE method can be considered as a weight corresponding to the communication state.

[0053] The loss weight calculation unit 15 calculates a loss weight according to the communication state for each LS estimated value sample, which is a first channel estimated value that is the channel estimation result for each of the multiple received signals. The loss weight calculation unit 15 may also calculate a weight based on the SNR for each first channel estimated value. Alternatively, the loss weight calculation unit 15 may calculate a weight based on the mean squared error for each first channel estimated value.

[0054] The weighted loss function calculation unit 14 receives the MSE input for each LS estimated value sample from the MSE calculation unit 13. The weighted loss function calculation unit 14 also obtains loss weights 16 that include loss weights corresponding to the communication state for each LS estimated value sample.

[0055] Next, the weighted loss function calculation unit 14 calculates a loss function that calculates the average between LS estimated samples after weighting the MSE for each LS estimated sample according to the communication state, based on the loss weight 16 for each LS estimated sample.

[0056] For example, the weighted loss function calculation unit 14 can calculate a loss function using SNR weighting with the following formula (4).

[0057]

number

[0058] Here, N B is the number of LS estimate samples in the minibatch. Also, h^(k,l) is the channel estimate sample for the subcarrier #k and OFDM symbol l resource elements estimated by channel estimation AI12. Also, SNR i This is the SNR of the LS estimate sample.

[0059] When MSE is used as the loss function, channel estimation accuracy decreases and MSE increases for LS estimate samples with low SNR. Conversely, channel estimation accuracy increases and MSE decreases for LS estimate samples with high SNR. In other words, when MSE is used as the loss function, the low SNR region where MSE is large is emphasized during learning, making it difficult to perform machine learning with high estimation accuracy over a wide SNR range. In contrast, a loss function that uses SNR weighting is weighted in a way that cancels out the difference in MSE size due to high and low SNR. Therefore, the loss is normalized with respect to SNR, making it possible to generate a channel estimation AI12 with high estimation accuracy over a wide SNR range.

[0060] Furthermore, the weighted loss function calculation unit 14 can calculate a loss function using weighting by the reciprocal of the MSE in the MMSE method using the following formula (5).

[0061]

number

[0062] Here, MSE MMSE,i This is the MSE obtained by the MMSE method of the LS estimate sample i. In this way, the weighted loss function calculation unit 14 determines the weighted loss function based on the mean squared error and the loss weights. More specifically, the weighted loss function calculation unit 14 weights the mean squared error based on the loss weight for each first channel estimate, and determines the weighted loss function as the average of the weighted mean squared errors among the first channel estimates.

[0063] The accuracy of channel estimation depends not only on the SNR, but also on the delay spread, which represents the multipath state, and the Doppler frequency, which is determined by the movement speed of the UE. Therefore, the weighted loss function calculation unit 14 can use a loss function that is weighted not only on the SNR, but also on the frequency correlation determined by the delay spread and the time correlation determined by the Doppler frequency.

[0064] In signal processing channel estimation, for example, the MMSE method was used to perform channel estimation using adaptive weights corresponding to SNR and time and frequency correlation. In contrast, the weighted loss function calculation unit 14 can calculate a loss function using the MSE obtained by the MMSE method and weighting by the reciprocal of the MSE.

[0065] Under conditions where channel estimation is difficult, the MSE represents the MSE. i The value becomes larger, and the loss shown by the MSE also increases. On the other hand, under conditions where channel estimation is easy, the MSE represents the MSE. iThe value becomes smaller, and the loss shown by the MSE also becomes smaller. Therefore, the weighted loss function calculation unit 14 can use a loss function weighted by the reciprocal of the MSE obtained by the MMSE method in order to cancel out the differences in the magnitude of the loss shown by the MSE of the samples depending on the conditions. The loss shown by this loss function can also be interpreted as the gain on the channel estimation accuracy that can be achieved by the MMSE method.

[0066] The control unit 10, channel estimation AI 12, MSE calculation unit 13, weighted loss function calculation unit 14, and loss weight calculation unit 15 repeat the above process, for example, until the estimation results converge for a specific mini-batch or until a predetermined number of iterations are completed. Furthermore, the control unit 10, channel estimation AI 12, MSE calculation unit 13, weighted loss function calculation unit 14, and loss weight calculation unit 15 repeat the parameter update by performing the next mini-batch using, for example, another predetermined number of randomly selected LS estimate samples. The control unit 10, channel estimation AI 12, MSE calculation unit 13, weighted loss function calculation unit 14, and loss weight calculation unit 15 repeat the parameter update until, for example, a predetermined number of mini-batches have been executed. The control of these iterations is performed, for example, under the leadership of the control unit 10.

[0067] Figure 4 shows an overview of the training of the channel estimation AI. Next, we will refer to Figure 4 to summarize the training of the channel estimation AI 12. Figure 4 shows the processing in one mini-batch.

[0068] The loss weight calculation unit 15 receives input from a group of LS estimated value samples 101, which includes a predetermined number of LS estimated value samples such as LS estimated value samples 111 and 112 used in minibatches included in the training data 17. The loss weight calculation unit 15 then performs a loss weight calculation process for each of the LS estimated value samples 111 and 112, etc. (step S1). As a result, the loss weight calculation unit 15 calculates a group of loss weight samples 104, which includes loss weight samples 141 and 142 used in minibatches.

[0069] The channel estimation AI 12 receives the LS estimate sample group 101 as input. The channel estimation AI 12 then performs channel estimation on each of the LS estimate samples 111 and 112, etc., to obtain a channel estimate sample group 103 to be used in a minibatch including channel estimate samples 131 and 132, etc. (Step S2).

[0070] The MSE calculation unit 13 acquires a group of ground truth channel samples 102 and a group of channel estimate samples 103 to be used in minibatches, each containing ground truth channel samples 121 and 122, etc., corresponding to LS estimate samples 111 and 112, etc. Then, the MSE calculation unit 13 uses the group of ground truth channel samples 102 and the group of channel estimate samples 103 to perform an MSE calculation process for each LS estimate sample and calculates the MSE (step S3).

[0071] The weighted loss function calculation unit 14 calculates the weighted loss function by performing a weighted loss function calculation process using the MSE for each LS estimated value sample and the loss weight sample group 104 (step S4).

[0072] Subsequently, the weighted loss function calculation unit 14 provides the weighted loss function to the channel estimation AI 12, causing it to update the parameters through backpropagation (step S5).

[0073] Figure 5 is a flowchart of the learning process of the channel estimation AI using the machine learning device according to the embodiment. Next, referring to Figure 5, the flow of the learning process of the channel estimation AI 12 using the machine learning device 1 according to the embodiment will be explained.

[0074] The control unit 10 selects a predetermined number of LS estimate samples to be used in the mini-batch included in the training data 17 (step S101).

[0075] Next, the loss weight calculation unit 15 calculates a loss weight sample by performing a loss weight calculation process for each LS estimate sample selected by the control unit 10 (step S102).

[0076] The channel estimation AI 12 receives input from the LS estimated value samples selected by the control unit 10. The channel estimation AI 12 then forward propagates the LS estimated value samples, performs channel estimation for each LS estimated value sample, and outputs the channel estimated value sample for each LS estimated value sample (step S103).

[0077] The MSE calculation unit 13 performs MSE calculation processing using the channel estimate sample and ground truth channel sample output from the channel estimation AI 12 to calculate the MSE for each LS estimate sample (step S104).

[0078] The weighted loss function calculation unit 14 calculates the weighted loss function using the MSE and loss weight samples for each LS estimated value sample (step S105).

[0079] The control unit 10 provides the weighted loss function calculated by the weighted loss function calculation unit 14 to the channel estimation AI 12 and causes it to perform backpropagation (step S106).

[0080] The control unit 10 updates the parameters of the channel estimation AI 12 according to the results of backpropagation (step S107).

[0081] Subsequently, the control unit 10 determines whether to terminate the mini-batch based on whether the estimation results have converged or whether a predetermined number of parameter updates have been completed (step S108). If the mini-batch has not been terminated (step S108: negative), the learning process of the channel estimation AI 12 returns to step S103.

[0082] In response to this, if the mini-batch is completed (step S108: affirmative), the control unit 10 determines whether the execution of a specified number of mini-batches has been completed (step S109).

[0083] If the execution of the specified number of minibatches has not been completed (step S109: negative), the learning process of the channel estimation AI 12 returns to step S101. Conversely, if the execution of the specified number of minibatches has been completed (step S109: positive), the control unit 10 terminates the learning process of the channel estimation AI 12.

[0084] In this embodiment, the receiver is described using base station 2 as an example, but the same applies if the receiver is a UE. Also, in this embodiment, the machine learning device 1 for training the channel estimation AI 12 is provided separately from base station 2, but the receiver such as base station 2 may also have the function to train the channel estimation AI 12 in machine learning device 1. If the machine learning device 1 for training the channel estimation AI 12 is provided separately from the receiver, offline training is also possible. Furthermore, if the function to train the channel estimation AI 12 is stored within the receiver, the receiver can perform proactive training. For example, the receiver can train the channel estimation AI 12 and update its parameters once every hour.

[0085] As described above, the machine learning device 1 according to this embodiment determines weightings for each LS estimate sample according to the communication conditions such as SNR and channel characteristics, and uses these weightings to calculate a weighted loss function that represents the average among the LS estimate samples. Then, the machine learning device 1 causes the channel estimation AI 12 to perform backpropagation using the calculated weighted loss function to update its parameters, thereby causing the channel estimation AI 12 to learn.

[0086] This allows for learning using a loss function that expresses losses considering communication conditions such as SNR and MSE, compared to learning using MSE as the loss function, thereby reducing the decrease in estimation accuracy depending on the communication conditions. For example, when MSE is used as the loss function, the magnitude of MSE is affected by the degree of difficulty in channel estimation depending on the communication conditions. In contrast, machine learning device 1 uses a weighted loss function that uses the reciprocal of MSE as a weight, thereby reducing the influence of communication conditions on MSE and improving estimation accuracy. By improving channel estimation accuracy in this way, communication between the transmitter and receiver can be performed appropriately, and the throughput of the communication system 100 can be improved.

[0087] Figure 6 shows a comparison of channel estimation accuracy. In Figure 6, the vertical axis represents MSE and the horizontal axis represents SNR. Here, the MSE on the vertical axis represents the mean squared error between the channel estimate obtained by each method and the ground truth channel estimate for the received signal. In other words, the lower the MSE value, the better the estimation accuracy.

[0088] In Figure 6, Graph 211 shows the estimation results when using the channel estimation AI 12 trained with the loss function (equation (5)) by the machine learning device 1 according to the embodiment. Graph 212 shows the estimation results when the channel estimation AI 12 is trained using the conventional MSE as the loss function. Graph 213 shows the estimation results when the channel estimation AI 12 is trained using the loss function (equation (4)) in which the MSE is weighted by the SNR.

[0089] As shown in Figure 6, the estimation results when using the channel estimation AI 12 trained with the machine learning device 1 shown in Graph 211 using the loss function (equation (5)) show a significant improvement in estimation accuracy in the high SNR region compared to the conventional estimation results when MSE is used as the loss function, as shown in Graph 212. Furthermore, the estimation results when using the channel estimation AI 12 trained with the machine learning device 1 shown in Graph 211 also show better estimation accuracy across the entire SNR region compared to Graph 213, which shows the result when a loss function (equation (4)) weighted by SNR for MSE is calculated. In other words, by using the channel estimation AI 12 trained with the machine learning device 1 according to this embodiment, the estimation accuracy of the channel estimation AI 12 is improved compared to when MSE is used as the loss function or when the loss function is calculated using MSE and SNR directly.

[0090] (Hardware configuration) Figure 7 is a hardware configuration diagram of the information processing device. The machine learning device 1 can be realized, for example, by the information processing device 80 shown in Figure 7.

[0091] As shown in Figure 7, the information processing device 80 includes, for example, a processor 81, a wireless communication line 82, a storage device 83, and memory 84. The processor 81 is connected to the wireless communication line 82, the storage device 83, and memory 84.

[0092] The wireless communication link 82 is an interface for wireless communication between the information processing device 80 and an external device. The wireless communication link 82 is used, for example, to transmit the learned channel estimation AI 12 to the base station 2.

[0093] The storage device 83 is an auxiliary storage device such as a hard disk or an SSD (Solid State Drive). The storage device 83 implements the functions of the data storage unit 11 as illustrated in Figure 1. The storage device 83 may also store the channel estimation AI 12 and the loss weights 16. Furthermore, the storage device 83 stores various programs, including programs for implementing the functions of the control unit 10, MSE calculation unit 13, weighted loss function calculation unit 14, and loss weight calculation unit 15 as illustrated in Figure 1.

[0094] Memory 84 is the main memory. Memory 84 can be, for example, DRAM (Dynamic Random Access Memory).

[0095] The processor 81 reads various programs from the storage device 83, loads them into memory 84, and executes them. In this way, the processor 81 realizes the functions of the control unit 10, MSE calculation unit 13, weighted loss function calculation unit 14, and loss weight calculation unit 15 as illustrated in Figure 1.

[0096] Figure 8 is a hardware configuration diagram of the receiver. Base station 2 can be implemented, for example, by the receiver 90 shown in Figure 8.

[0097] As shown in Figure 8, the receiver 90 includes, for example, a communication interface 91, a processor 92, a wireless communication line 93, a storage device 94, and memory 95. The processor 92 is connected to the communication interface 91, the wireless communication line 93, the storage device 94, and memory 95.

[0098] The wireless communication link 93 is an interface for wireless communication between the receiver 90 and external devices. The wireless communication link 82 is an interface for communication between, for example, the base station 2 and the UE. In addition, the wireless communication link 82 is used to receive trained channel estimation AI 12 from the machine learning device 1. The communication interface 91 is an interface for communication other than wireless communication between the receiver 90 and external devices. The communication interface 91 implements the functions of the output unit 25 illustrated in Figure 1. For example, the communication interface 91 is used for communication between the base station 2 and a higher-level device, or between the base station 2 and other base stations.

[0099] The storage device 94 is an auxiliary storage device such as a hard disk or SSD. The storage device 94 may store the channel estimation AI 26 illustrated in Figure 1. Furthermore, the storage device 94 stores various programs, including programs for realizing the functions of the LS estimation unit 21, inference unit 22, demodulation unit 23, and decoding unit 24 illustrated in Figure 1.

[0100] Memory 95 is the main memory. Memory 95 can be, for example, DRAM.

[0101] The processor 92 reads various programs from the storage device 94, loads them into memory 95, and executes them. This allows the processor 92 to implement the functions of the LS estimation unit 21, inference unit 22, demodulation unit 23, and decoding unit 24, as illustrated in Figure 1. [Explanation of Symbols]

[0102] 1. Machine Learning Device 2 base station 10 Control Unit 11 Data storage unit 12-channel estimation AI 13 MSE calculation section 14. Weighted Loss Function Calculation Unit 15 Loss weight calculation section 16 Loss weights 17 Training data 21 LS estimation part 22 Reasoning part 23 Demodulation Unit 24 Decoding section 25 Output section

Claims

1. For each of the multiple received signals, a first channel estimate is calculated, which is the channel estimation result for each of the multiple received signals. A loss weight is then calculated according to the communication state. The first channel estimate is input into a machine learning model, and a second channel estimate is determined for each of the first channel estimates by applying a correction to the first channel estimate. The mean squared error between the second channel estimate and the ground truth channel estimate, which is the correct channel estimate corresponding to the first channel estimate, is calculated for each of the first channel estimates. A weighted loss function is determined based on the mean squared error and the loss weights. The machine learning model is updated using the weighted loss function. A learning program characterized by having a computer perform the processing.

2. The process for determining the second channel estimate is as follows: The first channel estimate, which is the channel estimation result for each resource block of the wireless resource based on DRMS ​​obtained from the plurality of received signals, is input to the machine learning model. The machine learning model outputs a second channel estimate obtained by applying amplitude and phase corrections to the first channel estimate. The learning program according to claim 1, characterized by including processing.

3. The learning program according to claim 1, characterized in that the process for calculating the loss weights includes a process for calculating weights based on the SNR for each of the first channel estimates.

4. The learning program according to claim 1, characterized in that the process for calculating the loss weights includes a process for calculating weights based on the mean squared error for each of the first channel estimates.

5. The learning program according to claim 1, characterized in that the process of determining the weighted loss function includes a process of weighting the mean squared error based on the loss weight for each of the first channel estimates, and determining the average of the weighted mean squared errors among the first channel estimates as the weighted loss function.

6. Information processing device, For each of the multiple received signals, a first channel estimate is calculated, which is the channel estimation result for each of the multiple received signals. A loss weight is then calculated according to the communication state. The first channel estimate is input into a machine learning model, and a second channel estimate is determined for each of the first channel estimates by applying a correction to the first channel estimate. The mean squared error between the second channel estimate and the ground truth channel estimate, which is the correct channel estimate corresponding to the first channel estimate, is calculated for each of the first channel estimates. A weighted loss function is determined based on the mean squared error and the loss weights. The machine learning model is updated using the weighted loss function. A learning method characterized by performing a process.

7. A loss weight calculation unit calculates a loss weight according to the communication state for each of the first channel estimates, which are the channel estimation results for each of the multiple received signals. The MSE calculation unit inputs the first channel estimates into a machine learning model to determine a second channel estimate obtained by applying a correction to the first channel estimates output from the machine learning model, and calculates the mean squared error between the second channel estimates and the correct channel estimates corresponding to the first channel estimates for each of the first channel estimates. A weighted loss function calculation unit that determines a weighted loss function according to the mean squared error and the loss weights, A control unit that updates the machine learning model using the weighted loss function, An information processing device characterized by having the following features.

8. A loss weight calculation unit calculates a loss weight according to the communication state for each of the first channel estimates, which are the channel estimation results for each of the multiple received signals. The MSE calculation unit inputs the first channel estimates into a machine learning model to determine a second channel estimate obtained by applying a correction to the first channel estimates output from the machine learning model, and calculates the mean squared error between the second channel estimates and the correct channel estimates corresponding to the first channel estimates for each of the first channel estimates. A weighted loss function calculation unit that determines a weighted loss function according to the mean squared error and the loss weights, A control unit that updates and trains the machine learning model using the weighted loss function, An LS estimation unit that determines the first channel estimate value based on other received signals, An inference unit that inputs the determined first channel estimate into the trained machine learning model and obtains the second channel estimate output from the trained machine learning model, Based on the second channel estimate, a demodulation unit processes the other received signals. A communication device characterized by being equipped with the following features.

9. A communication system having a transmitter, a receiver and an information processing device, The aforementioned information processing device is A loss weight calculation unit calculates a loss weight according to the communication state for each of the first channel estimates, which are the channel estimation results for each of the multiple received signals. The MSE calculation unit inputs the first channel estimates into a machine learning model to determine a second channel estimate obtained by applying a correction to the first channel estimates output from the machine learning model, and calculates the mean squared error between the second channel estimates and the correct channel estimates corresponding to the first channel estimates for each of the first channel estimates. A weighted loss function calculation unit that determines a weighted loss function according to the mean squared error and the loss weights, The system includes a control unit that updates and trains the machine learning model using the weighted loss function, The aforementioned receiver is An LS estimation unit that determines the first channel estimate value based on other received signals transmitted from the aforementioned transmitter, An inference unit that inputs the determined first channel estimate into the trained machine learning model and obtains the second channel estimate output from the trained machine learning model, The system comprises a demodulation unit that processes the other received signals based on the second channel estimate. A communication system characterized by the following features.

Images