Channel equalization model training method, equalization processing method and device for fading signal

Abstract

The embodiment of the specification discloses a channel equalization model training method, an equalization processing method and device of a fading signal, which comprises the following steps: generating sample data sets by simulating original signals; selecting an initial model based on a mixed attention mechanism; inputting training sample data in the sample data sets into the initial model, extracting all-around feature information of the fading signal by using a double-branch encoder, and outputting a reconstructed signal by using a decoder to upgrade the feature information; repeatedly iterating and adjusting a learning rate of the initial model based on a SmoothL1Loss loss value of the reconstructed signal and a noise-free signal in the training sample data to determine model parameters, and obtaining a trained channel equalization model. Therefore, the signal reconstruction is efficiently realized, the purpose of channel equalization is achieved, and the model training speed is improved, and the subsequent signal transmission efficiency is improved.

Description

Technical Field

[0001] The present application relates to the fields of communications and computer technology, and in particular to a channel equalization model training method, and a method and device for equalizing fading signals. Background Art

[0002] In wireless communication systems, transmitted signals are inevitably affected by fading channels, resulting in multipath effects and Doppler frequency deviation, which in turn lead to transmission signal deterioration, reduced signal modulation recognition accuracy at the receiving end, and a high demodulation bit error rate.

[0003] Global communication systems face numerous challenges, including the potential to obtain confidential information from inter-symbol interference (ISI). Channel equalization is typically used to compensate for the impairments caused by fading channels, fading channels, and to facilitate subsequent modulation recognition, signal demodulation, signal interception, or interference. Therefore, channel equalization has long been a research hotspot in wireless communications.

[0004] Traditional channel equalization schemes mainly rely on time domain or frequency domain channel equalization, which requires sending a large number of pilot signals to test the channel status and train the filter parameters. This not only requires a long training time but also reduces transmission efficiency. Summary of the Invention

[0005] The purpose of the embodiments of this specification is to provide a channel equalization model training method, a fading signal equalization processing method and device, so as to efficiently realize signal reconstruction, achieve the purpose of channel equalization, and increase the model training speed and improve the subsequent signal transmission efficiency.

[0006] To solve the above technical problems, the embodiments of this specification are implemented as follows:

[0007] First, a channel equalization model training method is proposed, including:

[0008] Determine an original signal, and simulate the original signal to generate a sample data set; wherein the sample data in the sample data set includes: a noise-free signal used as label data and a noise signal used as a training sample, wherein the noise signal is a modulated signal carrying the influence of Rayleigh fading;

[0009] An initial model based on a hybrid attention mechanism is selected; the initial model includes: a dual-branch encoder and a single-branch decoder, wherein the encoding block main trunk in the dual-branch encoder is used to extract detailed spatial features, and the time residual dilation convolution RTDC auxiliary branch is used to construct temporal feature dependencies; the initial model is an autoencoder; the autoencoder based on the hybrid attention mechanism includes: a dual-branch encoder, a multi-head self-attention mechanism MHSA fusion feature module, and a single-branch decoder; wherein the dual-branch encoder includes: an encoding block main trunk and an RTDC auxiliary branch, and the encoding block main trunk includes 4 encoding blocks; the single-branch decoder includes 3 stacked decoding blocks; the encoding block internally includes: a depthwise separable convolution DSC, and a 3-layer DSC cascade compression-excitation SE attention mechanism module;

[0010] Inputting the training sample data in the sample data set into the initial model, extracting all-round feature information of the noise signal using a dual-branch encoder, and performing dimension-upgrading on the feature information using a decoder to output a reconstructed signal;

[0011] Based on the SmoothL1Loss loss value of the reconstructed signal and the noise-free signal in the training sample data, the initial model is repeatedly iterated, the learning rate is adjusted to determine the model parameters, and a trained channel equalization model is obtained.

[0012] Optionally, determining an original signal and simulating the original signal to generate a sample data set specifically includes:

[0013] Generate a random binary bit stream and simulate the original signal to be transmitted;

[0014] Based on the original signal, simulate modulation according to a plurality of preset modulation modes to generate a plurality of corresponding noise-free signals, and perform preprocessing operations on the noise-free signals, wherein the preprocessing operations at least include: waveform adjustment and power normalization processing;

[0015] Simulating a Rayleigh fading channel, simulating a noise-free signal after preprocessing, a fading signal transmitted through the Rayleigh fading channel, and simulating a Gaussian white noise channel to add noise power to the fading signal to obtain a noise signal;

[0016] A variety of different noise-free signals and their corresponding noise signals are aggregated into a sample data set. Optionally, the encoding block backbone routing of the dual-branch encoder is composed of the first convolutional layer and four encoding blocks connected in sequence, and a residual connection of the CBAM attention mechanism is set between the first convolutional layer and the third encoding block, and residual connections are set in the first and third layers within the encoding block and after the last scaling operation; the single-branch decoder is composed of three decoding blocks and a transposed convolutional layer connected in series, a residual connection is set between the first decoding block and the second decoding block, and residual connections are set in the first and third layers within the decoding block and after the last scaling operation.

[0017] Optionally, each DSC layer is followed by a batch normalization layer and an activation function layer. The squeeze-excite SE attention mechanism module consists of a pooling layer, a first convolutional layer, a LeakyReLU activation function, a second convolutional layer, and a Sigmod activation function.

[0018] On the trunk side of the coding block, the dual-branch encoder is used to extract comprehensive feature information from the noise signal, specifically including:

[0019] The first layer of DSC is used to expand the channel dimension of the feature information of the noise signal input to the coding block to obtain the corresponding feature map;

[0020] Use the second layer of DSC to downsample the feature map by controlling the step size and padding;

[0021] The third layer DSC is used to further adjust the number of channels of the input feature map;

[0022] The input feature map is compressed using the pooling layer and the first convolutional layer to compress the feature dimension;

[0023] The activation function and the second convolutional layer are used to stimulate the compressed features, transform the feature dimensions, and assign weights to obtain the feature weight size.

[0024] Optionally, the RTDC auxiliary branch is composed of 3 layers of DSC cascaded with 4 layers of time dilation convolution and CBAM;

[0025] On the auxiliary branch side of the RTDC, the dual-branch encoder is used to extract comprehensive feature information from the noise signal, specifically including:

[0026] The feature information of the input noise signal is sequentially expanded, dimensionally reorganized, and the feature map channels and spatial dimensions are compressed using DSC with kernel sizes of 3×3, 1×2, and 3×3. The feature dimensions are then matched once using DSC with a kernel size of 2×4, and the residual operation is performed on the compressed output.

[0027] The time series features obtained by residual operation are convolved N times in sequence: the kernel size is 1×3 and the expansion rate is 2 N-1 The time-expanded convolution TCN performs padding convolution and downsampling, where N is a positive integer;

[0028] The CBAM mechanism is used to assign feature weights to the extracted time series feature information.

[0029] Optionally, the formula of the CBAM is:

[0030]

[0031] Among them, the calculation formula of the channel attention module is:

[0032]

[0033] The calculation formula of spatial attention is:

[0034]

[0035] Among them, W0 and W1 are the weight values ​​calculated in the multi-layer perception network MLP, σ is the Sigmoid function, and F C max and F C avg They represent the two-branch parallel processing operations in the channel attention operation process, namely the maximum pooling and average pooling operations; the output M of the channel attention module is C (F) is multiplied with the input feature map F to obtain the channel-optimized feature F′; the feature map F′ is used as the input feature map of the spatial attention module, and the spatial attention module performs spatial weight distribution of feature information to obtain the feature map F″; Represents the feature map obtained after the input feature map F performs the maximum pooling operation on each channel C; Represents the feature map obtained after performing the average pooling operation on each channel C; Represents the feature map obtained after the maximum pooling operation of feature F′ in the channel S dimension, represents the feature map obtained after the average pooling operation of feature F′ in the channel S dimension; f 7×7 Represents a convolution operation with a convolution kernel size of 7×7.

[0036] Optionally, the multi-head self-attention mechanism MHSA formula is as follows:

[0037] Q=FW q

[0038] K=FW k

[0039] V=FWν

[0040]

[0041] Among them, F represents the feature map input to MHSA, Q, K, and V represent the query matrix, key matrix, and value matrix obtained according to the weight matrix respectively; W q 、W k 、W v is the weight matrix of the corresponding matrix; Attn is the multi-head attention mechanism operation process, D K is the length of F; the input feature F first passes through the LayerNorm layer (LN), and then multiplied by different weight matrices to obtain the query (Q), key (K) and value (V), which are then fed into the MHSA mechanism. Subsequently, by measuring the correlation between Q and K, the self-attention matrix of the mapping Attn is calculated to encapsulate the global information. q 、W k 、W v D is the weight matrix of the corresponding matrices of Q, K and V. k is the length of F. The MHSA mechanism performs transformations on Q, K, and V through h different linear transformations to compute different outputs. It is important to note that the computations for each head are performed in parallel.

[0042] The multi-head attention mechanism includes two parallel attention heads; each attention head has an independent projection weight matrix;

[0043] After extracting features based on the coding block main road side and the RTDC auxiliary branch side, the following feature fusion operations are also included:

[0044] For each attention head, multiply the query, key, and numerical vector with the corresponding weight matrix to obtain the query, key, and numerical vector corresponding to each attention head;

[0045] Perform attention calculation on each attention head to obtain the attention weights and weighted numerical vectors of multiple attention heads;

[0046] The weighted numerical vectors of multiple attention heads are connected or spliced ​​to obtain the global feature information output by the multi-head attention.

[0047] Secondly, a method for equalizing fading signals is proposed, including:

[0048] Acquiring test data to be processed, wherein the test data is a modulated signal carrying Rayleigh fading;

[0049] Inputting the test data into a channel equalization processing model for processing to obtain an equalized prediction signal;

[0050] The channel equalization processing model is trained based on the channel equalization model training method described in the first aspect.

[0051] In a third aspect, a channel equalization model training device is proposed, comprising:

[0052] A simulation module, configured to determine an original signal and simulate the original signal to generate a sample data set; wherein the sample data in the sample data set includes: a noise-free signal used as label data and a noise signal used as a training sample, wherein the noise signal is a modulated signal carrying the influence of Rayleigh fading;

[0053] A selection module is used to select an initial model based on a hybrid attention mechanism; the initial model includes: a dual-branch encoder and a single-branch decoder, wherein the encoding block main trunk in the dual-branch encoder is used to extract detailed spatial features, and the time residual dilation convolution RTDC auxiliary branch is used to construct temporal feature dependencies; the initial model is an autoencoder; the autoencoder based on the hybrid attention mechanism includes: a dual-branch encoder, a multi-head self-attention mechanism MHSA fusion feature module, and a single-branch decoder; wherein the dual-branch encoder includes: an encoding block main trunk and an RTDC auxiliary branch, and the encoding block main trunk includes 4 encoding blocks; the single-branch decoder includes 3 stacked decoding blocks; the encoding block internally includes: a depthwise separable convolution DSC, and a 3-layer DSC cascade compression-excitation SE attention mechanism module;

[0054] a processing module, configured to input the training sample data in the sample data set into the initial model, extract all-round feature information of the noise signal using a dual-branch encoder, and perform dimension upgrading on the feature information using a decoder, and output a reconstructed signal;

[0055] The training module is used to repeatedly iterate the initial model and adjust the learning rate to determine the model parameters based on the SmoothL1Loss loss value of the reconstructed signal and the noise-free signal in the training sample data, so as to obtain a trained channel equalization model.

[0056] In a fourth aspect, a device for equalizing a fading signal is proposed, comprising:

[0057] An acquisition module, configured to acquire test data to be processed, wherein the test data is a modulated signal carrying Rayleigh fading;

[0058] A prediction module, configured to input the test data into a channel equalization processing model for processing to obtain an equalized prediction signal;

[0059] The channel equalization processing model is trained based on the channel equalization model training method described in the first aspect.

[0060] In a fifth aspect, an electronic device is provided, comprising:

[0061] processor; and

[0062] A memory arranged to store computer-executable instructions, which, when executed, cause the processor to perform the steps of the channel equalization model training method described in the first aspect, and / or the steps of the equalization processing method for fading signals described in the second aspect.

[0063] In the sixth aspect, a computer-readable storage medium is proposed, which stores one or more programs. When the one or more programs are executed by an electronic device including multiple applications, the electronic device executes the steps of the channel equalization model training method described in the first aspect, and / or executes the steps of the equalization processing method of the fading signal described in the second aspect.

[0064] It can be seen from the technical solutions provided in the above embodiments of this specification that depthwise separable convolution, causal time dilation convolution, transposed convolution, etc. are selected to build the equalization autoencoder DSTReA-HA as an equalization model. In DSTReA-HA, the encoder adopts a dual-branch parallel structure to perform all-round feature extraction of the fading signal. The backbone network consists of a coding block composed of four stacked depthwise separable convolution cascade SE mechanisms and an additional CBAM jump connection branch, which deeply extracts the spatial and channel features of the fading signal; the auxiliary branch is composed of three layers of DSC cascaded four layers of TCN and CBAM, which mainly extracts the dimensionality of the temporal dependency of the fading signal. Finally, the dual-branch features are fused through a multi-head self-attention mechanism to ensure the complete transmission of the modulated signal feature information. The decoder of DSTReA-HA is relatively simple, mainly consisting of a decoding block of three transposed convolution cascade SE modules, which completes the upsampling of the feature information from the MHSA mechanism, gradually restores the original signal dimension, and finally outputs it through a 3×3 convolution layer. The depthwise separable convolution introduced in this application significantly reduces model complexity while ensuring performance. The constructed encoding and decoding blocks fuse the SE mechanism to deeply extract channel attention features from fading signals. The cascaded RTDC branches compensate for the backbone network's lack of time series information extraction, and the MHSA facilitates the fusion and importance distribution of encoder-extracted features. Finally, during the training phase, the SmoothL1Loss loss function is used to calculate the difference between the output layer's predicted value and the true label value to adjust the network model parameters. This ensures the accuracy of the model's reconstructed signal, improves the model's training accuracy, and improves the accuracy of subsequent equalization processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the embodiments of this specification or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments recorded in this specification. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0066] Figure 1 This is a schematic diagram of the steps of a channel equalization model training method provided in Example 1 of the present application.

[0067] Figure 2a This is a schematic diagram of the steps of generating a sample data set by simulation according to an embodiment of this specification.

[0068] Figure 2b This is a flow chart of generating sample data by simulation according to an embodiment of this specification.

[0069] Figure 3 This is a schematic diagram of the steps of a method for equalizing a fading signal provided in an embodiment of the present application.

[0070] Figure 4 This is a schematic diagram of the equalization processing flow of the fading signal provided by this application.

[0071] Figure 5 This is a schematic diagram of the structure of the autoencoder based on the hybrid attention mechanism provided in this application.

[0072] Figure 6 This is a structural diagram of the CBAM attention layer provided in this application.

[0073] Figure 7 This is a schematic diagram of the structure of the residual time dilation convolution provided by this application.

[0074] Figure 8 This is a structural diagram of the multi-head self-attention mechanism MHSA provided in this application.

[0075] Figure 9 It is a structural diagram of a channel equalization model training device provided in an embodiment of the present application.

[0076] Figure 10 It is a structural diagram of an equalization processing device for fading signals provided in an embodiment of the present application.

[0077] Figure 11 This is a schematic diagram of the structure of an electronic device provided by an embodiment of this specification. DETAILED DESCRIPTION

[0078] To help those skilled in the art better understand the technical solutions in this specification, the following will provide a clear and complete description of the technical solutions in the embodiments of this specification, in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of this specification, not all of them. All other embodiments derived by those skilled in the art based on the embodiments in this specification without creative effort shall fall within the scope of protection of this specification.

[0079] With the rapid development of artificial intelligence in recent years, machine learning has become a hot topic, gaining widespread research and application. Deep learning, a branch of machine learning, has seen significant progress in image processing, natural language processing, and other fields, using convolutional and recurrent neural networks. Furthermore, deep learning has surpassed traditional methods in several unsupervised learning tasks, such as image restoration, style transfer, image enhancement, and super-resolution. Currently, researchers in the communications field are applying deep learning techniques to solve problems such as channel estimation, modulation recognition, and resource allocation, achieving promising results and performance. Therefore, intelligent and autonomous deep learning technology offers a new approach to channel equalization.

[0080] The core concept of channel equalization in wireless communication systems is to exploit the characteristic information of the modulated signal under severe Rayleigh fading to reconstruct the signal and restore its original waveform. Deep learning can combine low-level features to form more abstract high-level representations of attribute categories or features, thereby discovering distributed feature representations of data. Currently, equalization of complex fading channels requires multi-dimensional parameter influences or different networks to achieve different training results. Therefore, a time-residual autoencoder with a hybrid attention mechanism has been proposed to solve the equalization problem in Rayleigh fading channels.

[0081] The inventive concept of the present application lies in: an equalizing autoencoder based on a depthwise separable temporal residual equalizer with hybrid attention mechanisms (DSTREA-HA), which mainly designs a depthwise separable convolution (DSC) cascade compression-excitation (SE) encoding block and a transposed convolution cascade compression-excitation (SE) decoding block, supplemented by a temporal residual dilated convolution (RTDC) branch, and calls a multi-head self-attention mechanism (MHSA) to assist in feature fusion in the equalizing network model. The received IQ signal is fully extracted from the time domain information and frequency domain features, and then the original signal is reconstructed to achieve the purpose of equalization.

[0082] The application mainly includes two parts: 1, a channel equalization model training method; 2, a channel equalization processing method based on the trained channel equalization model.

[0083] The main content of the training method includes: determining an original signal, and simulating sample data set from the original signal; selecting an initial model based on a mixed attention mechanism; inputting training sample data in the sample data set into the initial model, using a double-branch encoder to extract all-around feature information of the noise signal, and using a decoder to upgrade the feature information, and outputting a reconstructed signal; based on the SmoothL1Loss loss value of the reconstructed signal and the noise-free signal in the training sample data, repeatedly iterating and adjusting the learning rate of the initial model to determine the model parameters, and obtaining the trained channel equalization model.

[0084] The main content of the equalization processing method includes: obtaining test data to be processed, wherein the test data is a modulation signal carrying Rayleigh fading; inputting the test data into a channel equalization processing model for processing to obtain an equalized predicted signal; wherein the signal equalization processing model is trained based on the channel equalization model training method described above.

[0085] Through the technical scheme of the application, signal reconstruction can be efficiently realized, the purpose of channel equalization can be achieved, the model training speed can be improved, and the subsequent signal transmission efficiency can be improved.

[0086] The scheme related to the application will be described in detail below through specific embodiments.

[0087] Embodiment one

[0088] Figure 1 As shown in the figure, it is a step schematic diagram of a channel equalization model training method provided by the embodiment one of the application. The execution subject of the method can be a software module or a hardware device, for example, a cloud host, a computer, a tablet computer, a mobile terminal device, a smart wearable device, etc.; the channel equalization model training method can include the following steps:

[0089] Step 102: determining an original signal, and simulating sample data set from the original signal; wherein the sample data in the sample data set includes noise-free signal as label data and noise signal as training sample, and the noise signal is a modulation signal carrying Rayleigh fading effect.

[0090] Optionally, referring to Figure 2a When determining an original signal, and simulating sample data set from the original signal, the application can specifically include the following steps:

[0091] Step 202: randomly generating a binary bit stream, and simulating an original signal to be transmitted.

[0092] Specifically, combined Figure 2b As shown, a binary bit stream is randomly generated to simulate the original data to be transmitted, and 8 different noise-free modulation signals are simulated according to 8 different modulation modes. It should be understood that the 8 types here are just for illustration. In different business scenarios, different types and numbers of modulation modes can be set according to different business needs, and this is not limited here. Among them, the simulation tool can be MATLAB2020b, and the 8 modulation modes can be: BPSK, QPSK, 8PSK, GMSK, CPFSK, OQPSK, 16QAM, and 64QAM. Call functions in MATLAB such as comm.BPSKModulator, comm.CPFSKModulator, etc. and set the oversampling rate to 8 and the sampling frequency to 1024kHz.

[0093] Step 204: Based on the original signal, simulate modulation according to a plurality of preset modulation modes to generate corresponding plurality of different noise-free signals, and perform preprocessing operations on the noise-free signals, wherein the preprocessing operations at least include: waveform adjustment and power normalization processing.

[0094] In specific implementation, a square root raised cosine filter can be used to adjust the waveform of the modulated signal and perform a "cut-off" operation on the generated edge data to ensure a smooth waveform of the sample. In order to facilitate the addition of fading and Gaussian white noise, the filtered signal needs to be power normalized to ensure consistent sample power.

[0095] The filter formula is as follows:

[0096]

[0097] Where a is the roll-off factor of the root raised cosine filter, t is the sampling time, and T is the symbol period. The modulated signal is processed by filter g, modulating the corresponding 0101... bit stream onto the corresponding waveform before it can be transmitted via radio frequency.

[0098] The power normalization formula is as follows:

[0099]

[0100] Wherein, N represents the length of the signal; x(t) represents the input modulated signal that needs to be power normalized; and x'(t) represents the modulated signal after power normalization.

[0101] Step 206: Simulate a Rayleigh fading channel, simulate a noise-free signal after preprocessing, a fading signal transmitted through the Rayleigh fading channel, and simulate a Gaussian white noise channel to add noise power to the fading signal to obtain a noise signal.

[0102] In the implementation case of this application, a fading channel is first created. According to the comm.RayleighChannel tool provided by MATLAB, a six-path fading channel can be simulated and generated, where the path delay of each fading channel is [0200800120023003700]*1e-8 (seconds) and the average fading path gain is [0-0.9-4.9-8-7.8-23.9] (dB).

[0103] At the same time, using the MATLAB environment, a Gaussian white noise signal is simulated, and noise power with a signal-to-noise ratio of -12dB to 8dB is added to the fading signal with an interval of 2dB. Thus, the required noise signal is obtained.

[0104] Step 208: Aggregate the multiple different noise-free signals and their corresponding noise signals into a sample data set.

[0105] Finally, the real and imaginary parts of the generated IQ dataset are separated to create a sample dataset of size 2×1024. This sample dataset is then fed into the subsequent initial model composed of the autoencoder for training.

[0106] It should be understood that in this application, in addition to using Figure 2a In addition to generating sample data by the method steps shown, sample data that meets the above requirements can also be generated by other existing methods, which will not be described in detail in this application.

[0107] Step 104: Select an initial model based on a hybrid attention mechanism; the initial model includes: a dual-branch encoder and a single-branch decoder, wherein the encoding block main branch in the dual-branch encoder is used to extract detailed spatial features, and the time residual dilation convolution RTDC auxiliary branch is used to construct temporal feature dependencies.

[0108] Optionally, the initial model is an autoencoder; the autoencoder based on the hybrid attention mechanism includes: a dual-branch encoder, a multi-head self-attention mechanism MHSA fusion feature module, and a single-branch decoder; wherein the dual-branch encoder includes: a coding block main path and an RTDC auxiliary branch, and the coding block main path includes 4 coding blocks; the single-branch decoder includes 3 stacked decoding blocks; the dual-branch encoder is composed of a first convolution layer and 4 coding blocks connected in sequence, and a residual connection of the CBAM attention mechanism is set between the first convolution layer and the third coding block, and a residual connection is set in the first layer and the third layer within the coding block and after the last scaling operation; the single-branch decoder is composed of 3 decoding blocks and a transposed convolution layer in series, a residual connection is set between the first decoding block and the second decoding block, and a residual connection is set in the first layer and the third layer within the decoding block and after the last scaling operation.

[0109] Step 106: Input the training sample data in the sample data set into the initial model, use a dual-branch encoder to extract all-round feature information of the noise signal, and use a decoder to increase the dimension of the feature information to output a reconstructed signal.

[0110] Optionally, the encoding block includes: a depthwise separable convolution (DSC) and a three-layer DSC cascade compression-excitation (SE) attention mechanism module, wherein each layer of DSC is followed by a batch normalization layer and an activation function layer, and the SE module is composed of a pooling layer, a first convolution layer, a LeakyReLU activation function, a second convolution layer, and a Sigmod activation function; on the trunk side of the encoding block, when a dual-branch encoder is used to extract comprehensive feature information of the noise signal, specifically including:

[0111] The first layer of DSC is used to expand the channel dimension of the feature information of the noise signal input to the coding block to obtain the corresponding feature map; the second layer of DSC is used to downsample the feature map by controlling the step size and padding; the third layer of DSC is used to further adjust the number of channels of the input feature map; the pooling layer and the first convolution layer are used to compress the input feature map to compress the feature dimension; the activation function and the second convolution layer are used to excite the compressed features, transform the feature dimension, and assign weights to obtain the feature weight size.

[0112] In this application, the ordinary convolution within the encoding block is replaced by the depth-wise separable convolution (DSC). The decoding block consists of a three-layer DSC cascade SE attention mechanism connected to it. Each DSC layer is followed by a batch normalization layer (BatchNormalization) and a LeakyReLU activation function layer. The SE module consists of pooling, the first convolution layer, the LeakyReLU activation function, the second convolution layer, and the Sigmund activation function. The specific formula of the depth-wise separable convolution function DSC is as follows:

[0113]

[0114] Among them, K represents the size of D k ×D k ×m depth convolution kernel, F is the feature map input to Dsc, when i and j∈(0, D k ), the corresponding convolution sum operation is performed on the feature map F. In this configuration, each filter mth in K is applied to the corresponding channel mth in the input feature map F, thereby generating the mth channel in the filtered output feature map DSC.

[0115] The first layer DSC is used to expand the channel dimension of the feature information input to the coding block; the second layer DSC is used to downsample the input feature map by controlling the step size and padding; the third layer DSC is used to further adjust the number of channels of the input feature map; the pooling layer and the third convolution layer are used to compress the input feature map to compress the feature dimension to D k ×1×1; use the activation function and the fourth convolution layer to stimulate the compressed features and transform the dimension to D k / r×1×1 in the weighted value to obtain the feature weight size D k ×1×1.

[0116] Optionally, the RTDC auxiliary branch is composed of three layers of DSC cascaded with four layers of time dilation convolution and CBAM. On the RTDC auxiliary branch side, when a dual-branch encoder is used to extract comprehensive feature information of the noise signal, the following steps are specifically included:

[0117] The feature information of the input noise signal is sequentially expanded, dimensionally reorganized, and the feature map channels and spatial dimensions are compressed using DSC with kernel sizes of 3×3, 1×2, and 3×3. The feature dimensions are matched once based on 2×4 DSC and the residual operation is performed on the compressed output. The temporal features obtained by the residual operation are sequentially convolved N times: using a kernel size of 1×3 and an expansion rate of 2 N-1 The time dilation convolution TCN is padded and downsampled, where N is a positive integer; the CBAM mechanism is used to assign feature weights to the extracted temporal feature information.

[0118] In this application solution, the RTDC branch introduced by the coding block auxiliary branch performs time series feature extraction. The formula is as follows:

[0119]

[0120] In the formula, the input sequence of TCN is represented by xi, and the convolution kernel of TCN is represented by k. Where d represents the expansion rate, fd Represents the convolution kernel of the dilated causal convolution. The item x in the input sequence t-di represents the value of the modulating signal at time t-di, illustrating that the dilated convolution operation includes earlier time instances of the modulating signal samples.

[0121] During the specific extraction process, 3×3 depth-wise separable convolution can be used to expand the input feature channels; 1×2 depth-wise separable convolution can be used to reorganize the length and width dimensions of the feature map; 3×3 depth-wise separable convolution can be used to further compress the feature map channels and spatial dimensions; 2×4 depth-wise separable convolution can be used to match the input feature dimensions and perform residual operations with the output of the previous step; 1×3 time-expanded convolution with a dilation rate of 1 and a TCN noise reduction convolution interval of 1 are used; 1×3 time-expanded convolution with a dilation rate of 2 and a TCN noise reduction convolution interval of 2 are used, which is doubled compared to the previous step expansion dimension, and the receptive field of the model is also doubled, which can extract time features of different dimensions. The dilation rates of 4 and 8 have the same effect as above; the CBAM mechanism is used to assign important feature weights to the extracted temporal feature information.

[0122] The above multi-level dense residual connections solve the problem of network degradation and ensure the smooth transmission of original information.

[0123] Optionally, the formula of the CBAM is:

[0124]

[0125] Among them, the calculation formula of the channel attention module is:

[0126]

[0127] The calculation formula of spatial attention is:

[0128]

[0129] Among them, W0 and W1 are the weight values ​​calculated in MLP, σ is the Sigmoid function, and F C max and F C avg They represent the two-branch parallel processing operations in the channel attention operation process, namely the maximum pooling and average pooling operations. First, we take the output M of the channel attention module C (F) is multiplied by the input feature map F to obtain the channel-optimized feature F′. Represents the feature map obtained after the input feature map F performs the maximum pooling operation on each channel C; Represents the feature map obtained after performing the average pooling operation on each channel C; Represents the feature map obtained after the maximum pooling operation of feature F′ in the channel S dimension, represents the feature map obtained after the average pooling operation of feature F′ in the channel S dimension; secondly, the feature map F′ is used as the input feature map of the spatial attention module, and the feature information spatial weight distribution is performed by the spatial attention module to obtain the feature map F″, which ignores the interference of the fading channel in the spatial dimension and emphasizes the original information of the modulated signal; f 7 ×7 Represents a convolution operation with a convolution kernel size of 7×7.

[0130] Optionally, the multi-head self-attention mechanism MHSA formula is as follows:

[0131] Q=FW q

[0132] K=FW k

[0133] V=FW ν

[0134]

[0135] The input feature F first passes through the LayerNorm layer (LN), and then multiplied by different weight matrices to obtain the query (Q), key (K) and value (V), which are then fed into the MHSA mechanism. Subsequently, by measuring the correlation between Q and K, the self-attention matrix of the mapping Attn is calculated to encapsulate the global information. q 、W k 、W v D is the weight matrix of the corresponding matrices of Q, K and V. k is the length of F. The MHSA mechanism performs transformations on Q, K, and V through h different linear transformations to compute different outputs. It is important to note that the computations for each head are performed in parallel.

[0136] The multi-head attention mechanism includes two parallel attention heads; each attention head has an independent projection weight matrix;

[0137] After extracting features based on the coding block main road side and the RTDC auxiliary branch side respectively, the following feature fusion operations are also included: for each attention head, the query, key, and numerical vector are multiplied with the corresponding weight matrix to obtain the query, key, and numerical vector corresponding to each attention head; attention calculation is performed on each attention head to obtain the attention weights and weighted numerical vectors of multiple attention heads; the weighted numerical vectors of multiple attention heads are connected or spliced ​​to obtain the global feature information output by the multi-head attention.

[0138] Step 108: Based on the SmoothL1Loss loss value of the reconstructed signal and the noise-free signal in the training sample data, the initial model is repeatedly iterated, the learning rate is adjusted to determine the model parameters, and a trained channel equalization model is obtained.

[0139] In the present application, an autoencoder method based on a hybrid attention mechanism is constructed, and a training set is input for iteration, and the learning rate is adjusted for network training. The autoencoder consists of an encoder and a decoder, and the convolution layer is first connected for signal input and output. The encoder is used to comprehensively extract the characteristics of the fading signal, and the decoder is used to upgrade the dimension of the feature information. The signal reconstruction is completed and finally output by the output layer. The SmoothL1Loss function is used to determine the model parameters by outputting the loss value of the non-fading signal corresponding to the data set.

[0140] In this case, we first need to select a balanced autoencoder as the training model. Using the dataset samples generated in step 1, we package the training samples with fading noise into a fixed-size (batchsize×1×2×1024) tensor. The tensor size is expanded through the first convolutional layer (input layer) of the autoencoder, and finally restored through the last convolutional layer (output layer). The output tensor data is judged by the loss function SmoothL1Loss to determine the difference between it and the labeled data. The Adam optimizer is then used to determine the direction of the next round of model training. This step is repeated until the loss value drops to a certain range, that is, the predicted value approaches the label value, achieving a balanced effect. The SmoothL1Loss function formula SLL() is as follows:

[0141]

[0142] Where N represents the length of the input signal, and the variable z i The definition of is as follows:

[0143]

[0144] Where S i is the true value, S i ″ is the predicted value of the fading signal obtained by the equalizer.

[0145] Reference Figure 3 The figure shows a schematic diagram of the steps of a method for equalizing a fading signal provided by an embodiment of the present application. The execution subject of the method can be a software module or a hardware device, such as a cloud host, a computer, a tablet computer, a mobile terminal device, a smart wearable device, etc. The channel equalization model training method may include the following steps:

[0146] Step 302: Acquire test data to be processed, wherein the test data is a modulated signal carrying Rayleigh fading.

[0147] The acquisition of the test data may refer to step 102 and will not be described in detail here.

[0148] Step 304: Input the test data into the channel equalization processing model for processing to obtain an equalized prediction signal; wherein the signal equalization processing model is trained based on the channel equalization model training method described in steps 102 to 108.

[0149] When implementing it, you can load Figure 1 The model training method shown uses the saved model parameters and then feeds the modulated signal to be equalized from the test set (test data) into the trained equalization autoencoder. The autoencoder automatically outputs a predicted value of the original signal after equalization, filtering out the effects of Rayleigh fading and achieving equalization. After determining the optimal model parameter set path using training and validation samples, the test sample is directly operated on the data using this determined path set to directly obtain the signal after network equalization.

[0150] Figure 4 As shown, the equalization processing flow of the fading signal provided by this application includes: a model training process and a prediction process.

[0151] The model training process may include:

[0152] -Simulation to generate sample data sets

[0153] A dataset of modulated signals with Rayleigh fading was generated using MATLAB simulation. The dataset included the original noise-free signal used as label data and the signal affected by fading used as training samples. In addition to the Rayleigh fading effect, an additive white Gaussian noise (AWGN) channel was added. The signal-to-noise ratio ranged from -12dB to 8dB, with a 2dB interval. Each type of modulated signal had 2,000 data points under each SNR condition. The data was divided into training, validation, and test sets in an 8:1:1 ratio. The complex signal data was separated by real and imaginary components and constructed as a 2×1024 input tensor.

[0154] -The training set is input into the autoencoder for training

[0155] Reference Figure 5 As shown in the figure, the hybrid attention-based autoencoder consists of four encoding blocks in the backbone network, an auxiliary branch (Residual Time Dilated Convolution (RTDC)), a feature-fusion module (Multi-Head Self-Attention (MHSA)), and a decoder consisting of three stacked decoding blocks. During the training phase, a signal with fading noise is input into the model, and through iterations, model parameters are obtained that approximate the label. During the batch data balancing phase, these parameters are loaded into the model to calculate the balancing results.

[0156] The main branch of the encoding block consists of four stacked encoding blocks. A residual connection with the CBAM attention mechanism is added between the first and third encoding blocks. The encoding block consists of three layers of depthwise separable convolution (DSC), which is roughly one-ninth the complexity of a standard convolutional CNN. The kernel size of the first and third DSC layers is 3×3, while the kernel size of the second DSC layer is adjusted as needed to achieve dimensionality reduction of the feature map. Figure 5 For space reasons, the batch norm layer BatchNorm and the LeakyReLU activation function included in each DSC are omitted from the diagram. This is cascaded with the compressed excitation mechanism (SE). The output of the third DSC layer is fed into the SE module to weight the spatial features. This is then weighted with the output of the third DSC layer using a scaling function, providing some resistance to the effects of fading channels. Finally, a residual operation is performed with the initial input to ensure the integrity of the original information and avoid network degradation.

[0157] The formula of the coding block can be expressed as follows:

[0158]

[0159] Among them, k1, k2, and k3 represent the convolution kernel sizes of the first three depth-wise separable convolution (DSC) layers in the coding block, Add() is a point-by-point addition function, and Scale() represents the scaling point convolution operation. C Represents the output feature map after three layers of DSC convolution; X C is the signal feature map input to the encoding block. C V C The weight matrix with channel features after the residual compression-excitation attention mechanism (SE) is expressed as follows:

[0160] s c =F ex (g c , W) = σ(W2δ(W1g c ))

[0161] Scale(v c , s c )=s c V c

[0162] g c represents the feature map after global average pooling, with dimensions of number of channels × 1 × 1. The LeakyReLU function with slope parameter α = 0.001 is denoted as δ, and the sigmoid activation function is denoted as σ. W1 and W2 are the weight values ​​after excitation and compression by the SE module.

[0163] In this example, the 32 input layers use 3×3 convolution kernels to transform the input tensor to batchsize×32×2×1024. The first three encoding blocks then transform the feature maps to obtain feature maps with batchsize×128×2×512 dimensions. To ensure dimensionality matching for residual operations, the feature maps obtained from the input layer undergo a layer of DSC processing for dimensionality reduction and are then passed to CBAM for channel and spatial attention extraction. The fourth encoding block reduces the feature dimensions to batchsize×128×1×256.

[0164] Reference Figure 6 As shown in Figure 1, the CBAM attention layer consists of a channel attention module and a spatial attention module in series. The input data is F, which is a tensor of shape 128×2×512, where 128 is the number of channels, 2 is the height, and 512 is the width. The entire attention process can be summarized as:

[0165]

[0166] CBAM obtains a one-dimensional channel attention feature map F' and a two-dimensional spatial attention feature map F" through the following steps.

[0167] First, the channel attention module passes the input data F through a global average pooling layer, compressing the feature map of each channel into a scalar value. The pooled feature map is then fed into two fully connected layers, which learn the channel weights and biases, respectively. For each channel, its corresponding weight and bias are applied to the original feature map to obtain a weighted feature map. Finally, the weighted feature map is passed through a sigmoid activation function to limit its value between 0 and 1. This feature map is called the one-dimensional channel attention feature map F'.

[0168] Next is the spatial attention module: the one-dimensional channel attention feature map F' is passed through a global max pooling layer to compress the feature map of each channel into a scalar value. The pooled feature map is fed into two convolutional layers, which are used to learn the weights and biases of the spatial attention, respectively. The weights and biases of each channel are applied to the original feature map to obtain a weighted feature map. Finally, the weighted feature map is passed through a sigmoid activation function to limit its value between 0 and 1. This feature map is called the two-dimensional spatial attention feature map F".

[0169] Reference Figure 7 As shown in the figure, the residual time dilation convolution consists of three layers of DSC cascaded with four layers of time dilation convolution and CBAM, which is used to extract temporal dependencies in fading signals. The input features are sequentially compressed by three layers of DSC with kernel sizes of 3×3, 1×2, and 3×3. A 2×4 DSC residual connection is also used to prevent gradient vanishing and gradient exploding.

[0170] The kernel size of the four-layer time dilation convolution TCN is 1×3, and the dilation rates are [1, 2, 4, 8]. The input data is a time series represented by X = [x1, x2, x3, ..., x n ], where x i is the input of time step i. For each dilation rate, TCN uses a convolution layer to process the input data. Assuming the current dilation rate is d, the input data is padded according to the current dilation rate d. The purpose of padding is to keep the output length after convolution consistent with the input length. Padding can use zero padding or periodic padding. Convolution operation is performed on the padded input data. The activation function LeakyReLU is applied to the result of the convolution operation. After that, the pooling operation and stride are used to reduce the length of the convolution output to reduce the computational cost and increase the receptive field of the model. Repeat the above steps, use different dilation rates to convolve and downsample the data, and use multiple convolution layers to increase the depth and complexity of the model. The features after the output of the 4-layer TCN are sent to CBAM for further feature extraction.

[0171] The specific formula of TCN is as follows:

[0172]

[0173] Where d represents the expansion rate, f d Represents the convolution kernel of the dilated causal convolution. The item x in the input sequence t-di Represents the value of the modulated signal at time t-di.

[0174] Reference Figure 8 As shown in , in this example, the features extracted by the two parallel branches need to be further fused, so the multi-head self-attention mechanism MHSA is introduced to complete this task. The structure of MHSA is as follows Figure 8 shown.

[0175] The input sequence X is of size batchsize×128×1×256, which contains the Rayleigh fading feature signal. The input sequence X is linearly transformed and mapped into the query (Q), key (K) and value (V) spaces of multiple attention heads. This can be achieved by combining X with the weight matrix W q 、W k 、W v The dimensions of these weight matrices are consistent with X. Then, the mapped query (Q), key (K), and value (V) are divided into multiple attention heads. In this example, considering performance and complexity, the design is h = 2 attention heads. Then for each attention head i (i ranges from 1 to 4), the query (Q i ), key (K i) and value (V i ). Scaled dot-product attention operation: For each attention head i, a scaled dot-product attention operation is performed.

[0176] The steps MHSA takes include:

[0177] ① Calculate attention weight: By taking the query (Q i ) and key (K i ) performs a dot product operation, scales the result, and then applies the softmax function to obtain the attention weight matrix (attention matrix) A i .

[0178] ② Weighted sum: The attention weight matrix A i and value (V i ) performs matrix multiplication to obtain the output matrix O of attention head i i .

[0179] ③Multi-head series connection: The output matrix O of each attention head i Connect them in series to obtain the final output matrix O of the multi-head self-attention mechanism.

[0180] ④ Linear transformation: Perform a linear transformation on the output matrix O of the multi-head self-attention mechanism and map it back to the original feature space. This can be done by comparing it with the weight matrix W O Multiply to achieve.

[0181] ⑤Output: The output matrix after linear transformation is obtained, which represents the signal features after being processed by the multi-head self-attention mechanism.

[0182] Through the above process, MHSA can simultaneously focus on different positions and features in the input sequence and use attention weights to capture the global feature information of the Rayleigh fading signal. Each attention head can learn different attention patterns, thereby extracting useful feature representations and performing signal equalization to improve the received signal quality. Its formula is as follows:

[0183] Q=FW q

[0184] K=FW k

[0185] V=FW ν

[0186]

[0187] The input feature F first passes through the LayerNorm layer (LN), and then multiplied by different weight matrices to obtain the query (Q), key (K) and value (V), which are then fed into the MHSA mechanism. Subsequently, by measuring the correlation between Q and K, the self-attention matrix of the mapping Attn is calculated to encapsulate the global information. q 、W k 、W v D is the weight matrix of the corresponding matrices of Q, K and V. k is the length of F. The MHSA mechanism transforms Q, K, and V through h different linear transformations to compute different outputs. It is important to note that the computation of each head is performed in parallel. The multi-head attention mechanism contains two parallel attention heads; each attention head has an independent projection weight matrix.

[0188] Finally, the fading signal feature information after MHSA fusion is passed through three decoding blocks to reconstruct the signal and restore the waveform. Figure 5 Each decoding block is sequentially connected by the third convolutional layer, transposed convolution, and the fourth convolutional layer. The post-cascade SE mechanism extracts spatial feature weights, and then performs residual operations on the residual connection branches formed between the third and fourth convolutions and with the fifth convolution after the scaling operation. Transposed convolution is introduced in the decoding block to complete the upsampling process. Transposed convolution can be regarded as the inverse operation of the convolution layer. It can effectively filter out the influence of fading noise by increasing the dimension through the extracted feature information.

[0189] ---Training to obtain a balanced autoencoder

[0190] The forecasting process includes:

[0191] --Determine the test data set

[0192] The test dataset can be obtained from the pre-assigned test set in the sample dataset above. After using the training and validation samples to determine the optimal model parameter set path, for the test sample, the data is directly operated on the determined path set to directly obtain the signal after network equalization.

[0193] -- Input balanced autoencoder for testing

[0194] Load the model parameters saved in the previous step, and input the modulated signal to be equalized in the test set into the trained equalization autoencoder. The autoencoder automatically outputs the predicted value of the original signal after equalization of the modulated signal to filter out the influence of Rayleigh fading and achieve the purpose of equalization.

[0195] --Input reconstruction signal

[0196] The above-mentioned technical solution demonstrates advantages over current mainstream approaches in the hybrid attention-based autoencoder approach. This approach demonstrates strong generalization capabilities. Furthermore, the channel equalization performance of the residual temporal autoencoder based on the hybrid attention mechanism for eight types of modulation signals outperforms other noise reduction and equalization models in terms of improved signal-to-noise ratio (SNR), error vector magnitude (EVM) reduction, and demodulation bit error rate (BER). This hybrid attention-based autoencoder approach can directly extract feature maps from one-dimensional IQ modulated signal data without requiring secondary transformations. Compared to image-based data, it exhibits lower complexity and fewer floating-point operations (FLOPs), resulting in better generalization and easier deployment in practical applications. Furthermore, leveraging the CNN's expertise in extracting detailed feature information and the RNN's ability to establish temporal dependencies, an autoencoder was designed that can fully extract both spatial and temporal features. Building on this, the conventional CNN was replaced with a depthwise separable convolution (DSC), significantly reducing model parameters and improving training speed while maintaining performance. Furthermore, the residual time dilation convolution branch uses causal time dilation convolution with dilation rates of [1, 2, 4, 8]. This expands the model's receptive field while reducing computational effort, enabling it to capture the temporal characteristics of modulated signals under the influence of fading channels, thereby improving the model's robustness against fading. The CBAM and MHSA attention mechanisms are introduced into the encoder to weight important features and mitigate fading channel interference. Furthermore, multi-head parallel computation improves feature extraction efficiency.

[0197] Example 2

[0198] Figure 9This is a schematic diagram of the structure of a channel equalization model training device provided in an embodiment of the present application. The device 900 includes: a simulation module 901, a selection module 902, a processing module 903, and a training module 904. Among them, the simulation module 901 is used to determine the original signal and simulate the original signal to generate a sample data set; wherein the sample data in the sample data set includes: a noise-free signal used as label data and a noise signal used as a training sample, and the noise signal is a modulated signal carrying the influence of Rayleigh fading; the selection module 902 is used to select an initial model based on the hybrid attention mechanism; the initial model includes: a dual-branch encoder and a single-branch decoder, wherein the encoding block main path in the dual-branch encoder is used to extract detailed spatial features, and the time residual dilation convolution RTDC auxiliary branch is used to construct temporal feature dependencies; the processing module 903 is used to input the training sample data in the sample data set into the initial model, use the dual-branch encoder to extract all-round feature information of the noise signal, and use the decoder to upgrade the feature information, and output a reconstructed signal; the training module 904 is used to repeatedly iterate the initial model based on the SmoothL1Loss loss value of the reconstructed signal and the noise-free signal in the training sample data, adjust the learning rate to determine the model parameters, and obtain a trained channel equalization model.

[0199] Optionally, as an embodiment, when the simulation module 901 determines the original signal and simulates the original signal to generate a sample data set, it is specifically used to randomly generate a binary bit stream and simulate the original signal to be transmitted; based on the original signal, according to a plurality of preset different modulation modes, simulate modulation to generate corresponding multiple different noise-free signals, and perform preprocessing operations on the noise-free signal, wherein the preprocessing operation at least includes: waveform adjustment and power normalization processing; simulating a Rayleigh fading channel, and simulating the noise-free signal after the preprocessing operation, the fading signal transmitted through the Rayleigh fading channel, and simulating a Gaussian white noise channel to add noise power to the fading signal to obtain a noise signal; summarizing the multiple different noise-free signals and their corresponding noise signals into a sample data set.

[0200] The types of signals in the sample dataset are as follows:

[0201] (1) BPSK (Binary Phase Shift Keying): BPSK is a basic digital modulation technique used to modulate a digital bit stream into a sinusoidal signal. It uses two phase states to represent two binary values ​​(usually 0 and 1), with the phase of one state being 180 degrees out of phase with the other.

[0202] (2) QPSK (Quadrature Phase Shift Keying): QPSK is a common digital modulation technique used to simultaneously modulate two bits into a sinusoidal signal. It uses four different phase states (0 degrees, 90 degrees, 180 degrees, and 270 degrees) to represent each bit combination.

[0203] (3) 8-Phase Shift Keying (8PSK): 8PSK is a digital modulation technique used to simultaneously modulate three bits into a sinusoidal signal. It uses eight different phase states to represent each bit combination.

[0204] (4) OQPSK (Offset Quadrature Phase Shift Keying) is a digital modulation technique that is an improvement on QPSK. In OQPSK, the phase difference between adjacent symbols is 90 degrees. Unlike QPSK, the phase change occurs at the symbol edge rather than the symbol center.

[0205] (5) GMSK (Gaussian Minimum Shift Keying): GMSK is a frequency modulation technique used to convert digital signals into continuous sinusoidal signals. It uses a Gaussian filter to achieve continuous frequency variation to reduce spectrum spread and interference.

[0206] (6) CPFSK (Continuous Phase Frequency Shift Keying): CPFSK is a phase-continuous frequency modulation technique used to convert digital signals into analog signals. It represents different digital bits by keying at different frequencies while maintaining phase continuity.

[0207] (7) 16QAM (16-Quadrature Amplitude Modulation): 16QAM is an amplitude modulation technique used to simultaneously modulate four bits into a sinusoidal signal. It uses four different amplitude levels and four different phase states to represent each bit combination.

[0208] (8) 64QAM (64-Quadrature Amplitude Modulation): 64QAM is an amplitude modulation technique used to simultaneously modulate six bits into a sinusoidal signal. It uses six different amplitude levels and four different phase states to represent each bit combination.

[0209] The number of samples of each type of modulation signal at each signal-to-noise ratio is 2000, and they are divided into training set, validation set and test set in the ratio of 8:1:1.

[0210] In a specific implementation of the embodiment of this specification, the initial model is an autoencoder; the autoencoder based on the hybrid attention mechanism includes: a dual-branch encoder, a multi-head self-attention mechanism MHSA fusion feature module, and a single-branch decoder; wherein the dual-branch encoder includes: a coding block main path and an RTDC auxiliary branch, and the coding block main path includes 4 coding blocks; the single-branch decoder includes 3 stacked decoding blocks; the dual-branch encoder is composed of a first convolutional layer and 4 coding blocks connected in sequence, and a residual connection of the CBAM attention mechanism is set between the first convolutional layer and the third coding block, and a residual connection is set in the first layer and the third layer within the coding block and after the last scaling operation; the single-branch decoder is composed of 3 decoding blocks and a transposed convolutional layer in series, a residual connection is set between the first decoding block and the second decoding block, and a residual connection is set in the first layer and the third layer within the decoding block and after the last scaling operation.

[0211] In another specific implementation of the embodiments of this specification, the coding block includes: depthwise separable convolution DSC, and a 3-layer DSC cascade SE attention mechanism, wherein each layer of DSC is followed by a batch normalization layer and an activation function layer, and the compression-excitation SE module is composed of a pooling layer, a first convolution layer, a LeakyReLU activation function, a second convolution layer and a Sigmod activation function; on the main road side of the coding block, when the processing module 903 uses a dual-branch encoder to extract all-round feature information of the noise signal, it is specifically used to use the first layer of DSC to expand the channel dimension of the feature information of the noise signal input to the coding block to obtain the corresponding feature map; use the second layer of DSC to downsample the feature map by controlling the step size and padding; use the third layer of DSC to further adjust the number of channels of the input feature map; use the pooling layer and the first convolution layer to compress the input feature map to compress the feature dimension; use the activation function and the second convolution layer to excite the compressed features, transform the feature dimension, and assign weights to obtain the feature weight size.

[0212] In another specific implementation of the embodiment of this specification, the RTDC auxiliary branch is composed of 3 layers of DSC cascaded 4 layers of time dilation convolution and CBAM; on the RTDC auxiliary branch side, when the processing module 903 uses a dual-branch encoder to extract all-round feature information of the noise signal, it is specifically used to input the feature information of the noise signal, and sequentially perform channel expansion, dimension reorganization, and compression of the feature map channel and spatial dimension on the feature information through DSCs with kernel sizes of 3×3, 1×2, and 3×3, and perform residual operation on the feature dimension based on 2×4 DSC and the compressed output; the temporal features obtained by the residual operation are sequentially convolved N times: using a kernel size of 1×3 and a dilation rate of 2 N-1 The time dilation convolution TCN is padded and downsampled, where N is a positive integer; the CBAM mechanism is used to assign feature weights to the extracted temporal feature information.

[0213] In another specific implementation of the embodiment of this specification, the formula of the CBAM is:

[0214]

[0215] Among them, the calculation formula of the channel attention module is:

[0216]

[0217] The calculation formula of spatial attention is:

[0218]

[0219] Among them, W0 and W1 are the weight values ​​calculated in MLP, σ is the Sigmoid function, and the output is multiplied by the input feature map F to obtain the channel-optimized feature F' as the input feature map of the spatial attention module; f 7×7 Represents a convolution operation with a convolution kernel size of 7×7.

[0220] In another specific implementation of the embodiment of this specification, the multi-head self-attention mechanism MHSA formula is as follows:

[0221] Q=FW q

[0222] K=FW k

[0223] V=FW ν

[0224]

[0225] The multi-head attention mechanism includes two parallel attention heads; each attention head has an independent projection weight matrix;

[0226] After extracting features based on the main road side of the coding block and the auxiliary branch side of the RTDC, the processing module 903 is also used to perform the following feature fusion operations: for each attention head, the query, key, and numerical vector are multiplied with the corresponding weight matrix to obtain the query, key, and numerical vector corresponding to each attention head; attention calculation is performed on each attention head to obtain the attention weights and weighted numerical vectors of multiple attention heads; the weighted numerical vectors of multiple attention heads are connected or spliced ​​to obtain the global feature information output by the multi-head attention.

[0227] It should be understood that the channel equalization model training device of the embodiment of this specification can also perform Figure 1 The method shown in FIG. 1 and the channel equalization model training device are implemented in Figure 1 The functions of the illustrated embodiment will not be described in detail here.

[0228] Figure 10 This is a schematic diagram of the structure of a device for equalizing fading signals provided in an embodiment of the present application. The device 1000 includes an acquisition module 1001 and a prediction module 1002. The acquisition module 1001 is configured to acquire test data to be processed, wherein the test data is a modulated signal carrying Rayleigh fading. The prediction module 1002 is configured to input the test data into a channel equalization processing model for processing to obtain a predicted signal after equalization. The channel equalization processing model is trained based on the channel equalization model training method described above.

[0229] Example 3

[0230] An embodiment of the present application provides a computer-readable storage medium, which stores at least one instruction, and the at least one instruction is loaded and executed by a processor to implement the channel equalization model training method described above, and / or the equalization processing method of the fading signal described above.

[0231] An embodiment of the present application provides an electronic device, which includes a processor and a memory, wherein the memory stores at least one instruction, and the instruction is loaded and executed by the processor to implement the channel equalization model training method described above, and / or the equalization processing method of the fading signal described above.

[0232] In the technical solution of this application, the collection, storage, use, processing, transmission, provision and disclosure of user personal information involved comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0233] Figure 11A schematic block diagram of an example electronic device 1100 that can be used to implement embodiments of the present application is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present application described and / or claimed herein.

[0234] like Figure 11 As shown, the electronic device 1100 includes a computing unit 1101, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1102 or a computer program loaded from a storage unit 1108 into a random access memory (RAM) 1103. Various programs and data required for the operation of the electronic device 1100 can also be stored in the RAM 1103. The computing unit 1101, the ROM 1102, and the RAM 1103 are connected to each other via a bus 1104. An input / output (I / O) interface 1105 is also connected to the bus 1104.

[0235] Multiple components in the electronic device 1100 are connected to the I / O interface 1105, including an input unit 1106, such as a keyboard, a mouse, etc.; an output unit 1107, such as various types of displays, speakers, etc.; a storage unit 1108, such as a magnetic disk, an optical disk, etc.; and a communication unit 1109, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 1109 allows the electronic device 1100 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0236] The computing unit 1101 can be a variety of general-purpose and / or specialized processing components with processing and computing capabilities. Some examples of the computing unit 1101 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units that run machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 1101 performs the various methods and processes described above, such as the channel equalization model training method or the equalization processing method for fading signals. For example, in some embodiments, the channel equalization model training method or the equalization processing method for fading signals can be implemented as a computer software program that is tangibly contained in a machine-readable medium, such as the storage unit 1108. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 1100 via the ROM 1102 and / or the communication unit 1109. When the computer program is loaded into the RAM 1103 and executed by the computing unit 1101, one or more steps of the channel equalization model training method or the equalization processing method for fading signals described above can be performed. Alternatively, in other embodiments, the computing unit 1101 may be configured to execute the channel equalization model training method or the equalization processing method of the fading signal in any other appropriate manner (for example, by means of firmware).

[0237] Various embodiments of the systems and techniques described above can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), system-on-chip systems (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0238] The program code for implementing the methods of the present application can be written in any combination of one or more programming languages. Such program code can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when the program code is executed by the processor or controller, the functions / operations specified in the flow charts and / or block diagrams are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0239] In the context of the present application, a machine-readable medium can be a tangible medium that can contain or store a program for use by an instruction execution system, device or equipment or used in combination with an instruction execution system, device or equipment. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0240] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0241] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0242] A computer system may include a client and a server. The client and server are generally remote from each other and typically interact through a communication network. The client-server relationship arises through computer programs running on the respective computers and having a client-server relationship with each other. The server may be a cloud server, a server in a distributed system, or a server integrated with a blockchain.

[0243] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this application can be achieved. This is not a limitation herein.

[0244] The above specific embodiments do not constitute a limitation on the scope of protection of this application. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application shall be included within the scope of protection of this application.

Claims

1. A channel equalization model training method, characterized in that: include: Determine an original signal, and simulate the original signal to generate a sample data set; wherein the sample data in the sample data set includes: a noise-free signal used as label data and a noise signal used as a training sample, wherein the noise signal is a modulated signal carrying the influence of Rayleigh fading; An initial model based on a hybrid attention mechanism is selected; the initial model includes: a dual-branch encoder and a single-branch decoder, wherein the encoding block main trunk in the dual-branch encoder is used to extract detailed spatial features, and the time residual dilation convolution RTDC auxiliary branch is used to construct temporal feature dependencies; the initial model is an autoencoder; the autoencoder based on the hybrid attention mechanism includes: a dual-branch encoder, a multi-head self-attention mechanism MHSA fusion feature module, and a single-branch decoder; wherein the dual-branch encoder includes: an encoding block main trunk and an RTDC auxiliary branch, and the encoding block main trunk includes 4 encoding blocks; the single-branch decoder includes 3 stacked decoding blocks; the encoding block internally includes: a depthwise separable convolution DSC, and a 3-layer DSC cascade compression-excitation SE attention mechanism module; Inputting the training sample data in the sample data set into the initial model, extracting all-round feature information of the noise signal using a dual-branch encoder, and performing dimension-upgrading on the feature information using a decoder to output a reconstructed signal; Based on the SmoothL1Loss loss value of the reconstructed signal and the noise-free signal in the training sample data, the initial model is repeatedly iterated, the learning rate is adjusted to determine the model parameters, and a trained channel equalization model is obtained.

2. The method according to claim 1, wherein Determining an original signal and simulating the original signal to generate a sample data set specifically includes: Generate a random binary bit stream and simulate the original signal to be transmitted; Based on the original signal, simulate modulation according to a plurality of preset different modulation modes to generate corresponding multiple different noise-free signals, and perform preprocessing operations on the noise-free signals, wherein the preprocessing operations at least include: waveform adjustment and power normalization processing; Simulating a Rayleigh fading channel, simulating a noise-free signal after preprocessing, a fading signal transmitted through the Rayleigh fading channel, and simulating a Gaussian white noise channel to add noise power to the fading signal to obtain a noise signal; A variety of different noise-free signals and their corresponding noise signals are summarized as a sample data set.

3. The method according to claim 1, wherein The encoding block trunk routing of the dual-branch encoder is composed of the first convolutional layer and four encoding blocks connected in sequence, and a residual connection of the CBAM attention mechanism is set between the first convolutional layer and the third encoding block, and residual connections are set in the first and third layers within the encoding block and after the last scaling operation; the single-branch decoder is composed of three decoding blocks and a transposed convolutional layer in series, a residual connection is set between the first decoding block and the second decoding block, and residual connections are set in the first and third layers within the decoding block and after the last scaling operation.

4. The method according to claim 1, wherein Each DSC layer is followed by a batch normalization layer and an activation function layer. The compression-excitation SE attention mechanism module consists of a pooling layer, the first convolutional layer, the LeakyReLU activation function, the second convolutional layer, and the Sigmod activation function. On the trunk side of the coding block, the dual-branch encoder is used to extract comprehensive feature information from the noise signal, specifically including: The first layer of DSC is used to expand the channel dimension of the feature information of the noise signal input to the coding block to obtain the corresponding feature map; Use the second layer of DSC to downsample the feature map by controlling the step size and padding; The third layer DSC is used to further adjust the number of channels of the input feature map; The input feature map is compressed using the pooling layer and the first convolutional layer to compress the feature dimension; The activation function and the second convolutional layer are used to stimulate the compressed features, transform the feature dimensions, and assign weights to obtain the feature weight size.

5. The method according to claim 3, wherein The RTDC auxiliary branch consists of 3 layers of DSC cascaded with 4 layers of time dilation convolution and CBAM; On the auxiliary branch side of the RTDC, the dual-branch encoder is used to extract comprehensive feature information from the noise signal, specifically including: The feature information of the input noise signal is sequentially expanded, dimensionally reorganized, and the feature map channels and spatial dimensions are compressed using DSC with kernel sizes of 3×3, 1×2, and 3×3. The feature dimensions are then matched once using DSC with a kernel size of 2×4, and the residual operation is performed on the compressed output. The time series features obtained by residual operation are convolved N times in sequence: the kernel size is 1×3 and the expansion rate is 2 N-1 The time-expanded convolution TCN performs padding convolution and downsampling, where N is a positive integer; The CBAM mechanism is used to assign feature weights to the extracted time series feature information.

6. The method according to claim 5, wherein The formula of the CBAM is: Among them, the calculation formula of the channel attention module is: The calculation formula of spatial attention is: Where W0 and W1 are the weight values ​​calculated in the multi-layer perception network MLP, σ is the Sigmoid function, and the output is multiplied by the input feature map F to obtain the channel-optimized feature F′, which is used as the input feature map of the spatial attention module. The output F is the feature map F″ with spatial channel attention after CBAM processing; Represents the feature map obtained after the input feature map F performs the maximum pooling operation on each channel C; Represents the feature map obtained after performing the average pooling operation on each channel C; Represents the feature map obtained after the maximum pooling operation of feature F′ in the channel S dimension, represents the feature map obtained after the average pooling operation of feature F′ in the channel S dimension; f 7×7 Represents a convolution operation with a convolution kernel size of 7×7.

7. The method according to claim 4 or 5, characterized in that The multi-head self-attention mechanism MHSA formula is as follows: Q=FW q K=FW k V=FW v Among them, F represents the feature map input to MHSA, Q, K, and V represent the query matrix, key matrix, and value matrix obtained according to the weight matrix respectively; W q 、W k 、W v is the weight matrix of the corresponding matrix; Attn is the multi-head attention mechanism operation process, D K is the length of F; The multi-head self-attention mechanism includes two parallel attention heads; each attention head has an independent projection weight matrix; After extracting features based on the coding block main road side and the RTDC auxiliary branch side, the following feature fusion operations are also included: For each attention head, multiply the query, key, and numerical vector with the corresponding weight matrix to obtain the query, key, and numerical vector corresponding to each attention head; Perform attention calculation on each attention head to obtain the attention weights and weighted numerical vectors of multiple attention heads; The weighted numerical vectors of multiple attention heads are connected or spliced ​​to obtain the global feature information output by the multi-head attention.

8. A method for equalizing a fading signal, characterized in that: include: Acquiring test data to be processed, wherein the test data is a modulated signal carrying Rayleigh fading; Inputting the test data into a channel equalization processing model for processing to obtain an equalized prediction signal; The channel equalization processing model is obtained by training based on the channel equalization model training method according to any one of claims 1 to 7.

9. A channel equalization model training device, characterized in that: include: A simulation module, configured to determine an original signal and simulate the original signal to generate a sample data set; wherein the sample data in the sample data set includes: a noise-free signal used as label data and a noise signal used as a training sample, wherein the noise signal is a modulated signal carrying the influence of Rayleigh fading; A selection module is used to select an initial model based on a hybrid attention mechanism; the initial model includes: a dual-branch encoder and a single-branch decoder, wherein the encoding block main trunk in the dual-branch encoder is used to extract detailed spatial features, and the time residual dilation convolution RTDC auxiliary branch is used to construct temporal feature dependencies; the initial model is an autoencoder; the autoencoder based on the hybrid attention mechanism includes: a dual-branch encoder, a multi-head self-attention mechanism MHSA fusion feature module, and a single-branch decoder; wherein the dual-branch encoder includes: an encoding block main trunk and an RTDC auxiliary branch, and the encoding block main trunk includes 4 encoding blocks; the single-branch decoder includes 3 stacked decoding blocks; the encoding block internally includes: a depthwise separable convolution DSC, and a 3-layer DSC cascade compression-excitation SE attention mechanism module; a processing module, configured to input the training sample data from the sample data set into the initial model, extract all-round feature information of the noise signal using a dual-branch encoder, and perform dimension upgrading on the feature information using a decoder, and output a reconstructed signal; The training module is used to repeatedly iterate the initial model and adjust the learning rate to determine the model parameters based on the SmoothL1Loss loss value of the reconstructed signal and the noise-free signal in the training sample data, so as to obtain a trained channel equalization model.

10. An equalization processing device for fading signals, characterized in that: include: An acquisition module, configured to acquire test data to be processed, wherein the test data is a modulated signal carrying Rayleigh fading; A prediction module, configured to input the test data into a channel equalization processing model for processing to obtain an equalized prediction signal; The channel equalization processing model is obtained by training based on the channel equalization model training method according to any one of claims 1 to 7.

11. A computer-readable storage medium, characterized in that The computer-readable storage medium stores one or more programs, which, when executed by an electronic device including multiple applications, enable the electronic device to perform the steps of the channel equalization model training method described in any one of claims 1 to 7, and / or, perform the steps of the equalization processing method for fading signals described in claim 8.

12. A computer program product, characterized in that The computer-readable storage medium stores one or more programs, which, when executed by an electronic device including multiple applications, enable the electronic device to perform the steps of the channel equalization model training method described in any one of claims 1 to 7, and / or, perform the steps of the equalization processing method for fading signals described in claim 8.

Images