A signal recognition method based on gradient proof-of-stake mechanism convolution

Abstract

This invention discloses a signal recognition method based on gradient proof-of-stake convolution. The method is implemented through a signal recognition model, which is constructed from an input layer, at least one improved bottleneck module, an output layer, and a classification head. The improved bottleneck module includes extended convolution, gradient proof-of-stake convolution, an SE module, and a linear bottleneck. The gradient proof-of-stake convolution performs a standard convolution on the extended convolutional feature map of the signal to be recognized, calculating the feature map, channel weights, and residual coefficients. The SE module weights the calculated feature map to obtain a weighted feature map. The linear bottleneck compresses the weighted feature map, and the output layer transforms the weighted and compressed feature map into a global vector. The classification head performs signal recognition based on the global vector. This invention constructs a signal recognition model based on gradient proof-of-stake convolution, improving signal recognition accuracy without significantly increasing network parameters and computational load.

Description

Technical Field

[0001] This invention relates to the field of signal recognition technology, and in particular to a signal recognition method based on gradient proof-of-stake convolution. Background Technology

[0002] Signal recognition has significant application value in fields such as underwater security, marine resource exploration, and military defense. However, signals are significantly affected by complex environments, exhibiting problems such as multipath effects, environmental noise interference, and non-steady-state propagation. Traditional methods primarily rely on manual feature extraction and classification head design, whose performance is highly dependent on a combination of expert experience and signal processing techniques. This results in high computational complexity and poor stability, leading to low signal recognition efficiency and an inability to achieve accurate signal identification.

[0003] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this invention is to provide a signal recognition method based on gradient proof-of-stake mechanism convolution, aiming to solve the technical problem of how to achieve accurate signal recognition while improving signal recognition efficiency.

[0005] To achieve the above objectives, this invention provides a signal recognition method based on gradient proof-of-stake convolution:

[0006] Input the binary file of the signal to be identified into the signal recognition model, and output the signal recognition result;

[0007] The signal recognition model is constructed based on an input layer, at least one improved bottleneck module, an output layer, and a classification head. The improved bottleneck module includes extended convolution, gradient proof-of-stake convolution, SE module, and linear bottleneck.

[0008] The input layer converts the binary file into a channel-size feature map;

[0009] The extended convolution expands the channel size feature map to obtain an extended convolution feature map;

[0010] The gradient proof-of-stake mechanism convolution performs standard convolution on the extended convolution feature map, and performs convolution calculation on the standard convolution feature map, the weights of each channel, and the residual coefficients to obtain the depth convolution feature map;

[0011] The SE module and the linear bottleneck perform weighted and compressed processing on the deep convolutional feature map to obtain a compressed weighted feature map.

[0012] The output layer converts the compressed weighted feature map into a global vector, so that the classification head can perform signal recognition based on the global vector.

[0013] Optionally, before inputting the binary file of the signal to be identified into the signal recognition model and outputting the signal recognition result, the following steps are included:

[0014] Obtain the binary data sample corresponding to the signal sample;

[0015] The binary data sample is input into the initial recognition model for training, and the signal recognition result sample corresponding to the signal sample is output.

[0016] Calculate the loss function value corresponding to the initial recognition model based on the signal recognition result sample and the real recognition result corresponding to the signal sample;

[0017] Based on the loss function value, the model parameters in the initial recognition model are updated in reverse until the loss function value is less than or equal to a preset threshold, and the trained initial recognition model is used as the signal recognition model.

[0018] Optionally, the step of inputting the binary data sample into the initial recognition model for training includes:

[0019] Obtain the stakes parameters of each channel within the convolution of the gradient proof-of-stake mechanism;

[0020] The stakes parameter of each channel is normalized by using the clamp function to obtain the normalized stakes parameter of each channel.

[0021] Calculate the mean and standard deviation of the normalized stakes parameter for each channel;

[0022] The normalized scores parameters of each channel are standardized based on the mean and the standard deviation to obtain the standardized scores parameters of each channel.

[0023] The weights of each channel are calculated based on the standardized stakes parameters and adaptive temperature of each channel.

[0024] The binary data samples are input into the initial recognition model for training based on the channel weights and residual coefficients.

[0025] Optionally, before calculating the weights of each channel based on the standardized stakes parameters and adaptive temperature of each channel, the following steps are included:

[0026] Determine the current training round of the initial recognition model;

[0027] The temperature constraints of the gradient proof-of-stake mechanism convolution are constructed using a PyTorch tensor truncation function based on the maximum and minimum temperature values.

[0028] Based on the temperature constraints, an adaptive temperature is calculated according to the current training round and temperature parameters.

[0029] Optionally, the step of calculating the weight of each channel based on the standardized stakes parameter and adaptive temperature of each channel includes:

[0030] Determine the Gumbel noise in the convolution of the gradient proof-of-stake mechanism;

[0031] The weights of each channel are calculated using the softmax function based on the standardized stakes parameters, adaptive temperature, and Gumbel noise.

[0032] Optionally, before inputting the binary data sample into the initial recognition model for training based on the channel weights and residual coefficients, the process includes:

[0033] Determine the learnable residual parameters of the convolution in the gradient proof-of-stake mechanism;

[0034] The residual coefficients are calculated using the sigmoid function based on the learnable residual parameters.

[0035] Optionally, the input layer is constructed by combining a standard convolutional layer, a batch normalization (BN) layer, and an activation layer;

[0036] The standard convolutional layer performs standard convolution on the binary file to generate a standard convolutional feature map.

[0037] The BN layer normalizes the standard convolutional feature map to obtain a normalized feature map.

[0038] The activation layer obtains the channel size feature map based on the normalized feature map using the Hardswish activation function.

[0039] Optionally, the SE module performs weighted processing on the depthwise convolutional feature map to obtain a weighted feature map;

[0040] The linear bottleneck compresses the weighted feature map to obtain a compressed weighted feature map.

[0041] Optionally, the output layer is constructed by combining extended convolutional layers, batch normalization (BN) layers, activation layers, and global average pooling layers.

[0042] The extended convolutional layer performs extended convolution on the compressed weighted feature map to obtain an extended convolutional feature map.

[0043] The BN layer normalizes the expanded convolutional feature map to obtain a normalized feature map.

[0044] The activation layer obtains the activated normalized feature map based on the normalized feature map using the Hardswish activation function;

[0045] The global average pooling layer performs vector transformation on the activated normalized feature map to obtain a global vector.

[0046] Optionally, the classification head is constructed by combining a fully connected layer, an activation layer, a Dropout layer, and a Softmax layer;

[0047] The fully connected layer flattens the global vector to obtain a one-dimensional vector.

[0048] The activation layer obtains the activated one-dimensional vector based on the one-dimensional vector using the Hardswish activation function;

[0049] The Dropout layer randomly discards the activated one-dimensional vector;

[0050] The fully connected layer obtains the global feature vector based on the randomly lost one-dimensional vector;

[0051] The Softmax layer performs signal recognition based on the global feature vector.

[0052] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing a signal recognition program, which, when executed by a processor, implements the steps of the signal recognition method based on gradient proof-of-stake convolution as described above.

[0053] This invention discloses a signal recognition method based on gradient proof-of-stake convolution. The signal recognition method is implemented through a signal recognition model, which is constructed based on an input layer, at least one improved bottleneck module, an output layer, and a classification head. The improved bottleneck module includes extended convolution, gradient proof-of-stake convolution, an SE module, and a linear bottleneck. First, the extended convolution expands the channel size feature map corresponding to the signal to be recognized transmitted from the input layer to obtain an extended convolution feature map. Then, the gradient proof-of-stake convolution performs standard convolution on the extended convolution feature map and performs convolution calculation on the feature map of the standard convolution, the weights of each channel, and the residual coefficients to obtain a depth convolution feature map. After that, the SE module obtains a weighted feature map based on the depth convolution feature map using the SE weight formula. The linear bottleneck compresses the weighted feature map so that the output layer converts the compressed weighted feature map into a global vector. Finally, the classification head performs signal recognition on the signal to be recognized based on the global vector. The gradient proof-of-stake mechanism convolution proposed in this invention solves the problems of high computational complexity and low efficiency of traditional attention mechanisms. The signal recognition model constructed based on the combination of the input layer, at least one improved bottleneck module, output layer and classification head alleviates the numerical instability caused by mathematical transformation, thereby achieving accurate signal classification. Attached Figure Description

[0054] Figure 1 This is a block diagram of the signal recognition model structure of the first embodiment of the signal recognition method based on gradient proof-of-stake mechanism convolution of the present invention;

[0055] Figure 2 This is a structural block diagram of several improved bottleneck modules in the first embodiment of the signal recognition method based on gradient proof-of-stake mechanism convolution of the present invention.

[0056] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0057] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0058] This invention provides a signal recognition model based on gradient proof-of-stake convolution, referring to... Figure 1 , Figure 1 This is a block diagram of the signal recognition model structure of the first embodiment of the signal recognition method based on gradient proof-of-stake mechanism convolution of the present invention.

[0059] In this embodiment, an initial recognition model needs to be pre-built and trained. The trained model is then used as a signal recognition model. The signal recognition model is constructed based on an input layer 101, at least one improved bottleneck module 102, an output layer 103, and a classification head 104. The improved bottleneck module includes an extended convolution 1021, a gradient proof-of-stake convolution 1022, an SE module 1023, and a linear bottleneck 1024.

[0060] In the specific implementation, when training the initial recognition model, it is necessary to predetermine the signal type corresponding to the signal sample. The signal type can be 1D or 2D. The signal sample corresponding to 1D is time series data such as audio, and the signal sample corresponding to 2D is two-dimensional data such as images (the image is a grayscale image). Then, the signal sample is preprocessed (for example, if it is a 1D signal sample, the sampling rate is uniform; if it is a 2D signal sample, the resolution is uniform), and the preprocessed data is converted into binary file samples (i.e., binary data samples). The binary data samples are input into the initial recognition model for training, and the signal recognition result samples corresponding to the signal samples are output. Based on the signal recognition result samples and the corresponding real recognition results of the signal samples, the loss function value corresponding to the initial recognition model is calculated through the loss function. Based on the loss function value, the model parameters in the initial recognition model are updated in reverse until the gradient of the loss function value is less than or equal to the preset gradient threshold. Then, the trained initial recognition model is used as the signal recognition model.

[0061] It should be noted that the preset gradient threshold is a user-defined setting. Model parameters include the stakes parameter, channel weights, residual coefficients, and adaptive temperature, among others.

[0062] It should be understood that signals can be sound signals, light signals, electrical signals, etc.

[0063] The process of inputting binary data samples into the initial recognition model for training is as follows: The input layer converts the binary data samples into channel-size feature map samples; the extended convolution expands the channel-size feature map samples to obtain extended convolution feature map samples; the gradient proof-of-stake mechanism convolution performs standard convolution on the extended convolution feature map samples, and performs convolution calculation on the standard convolution feature map samples, the channel weights, and the residual coefficients to obtain depthwise convolution feature map samples; the SE module and the linear bottleneck perform weighted and compressed processing on the standard convolution feature map samples to obtain compressed weighted feature map samples; the output layer converts the compressed weighted feature map samples into global vector samples, so that the classification head can perform signal recognition based on the global vector samples to obtain the signal recognition result samples corresponding to the signal samples.

[0064] The gradient proof-of-stake mechanism convolution is a POSConv dynamic convolution (3x3 or 5x3 convolution), which is a dynamic convolution constructed based on the gradient proof-of-stake mechanism and standard convolution.

[0065] In this embodiment, the extended convolutional feature map sample X (the extended convolutional feature map sample X is the input of the standard convolutional layer, in the format (N,C,H,W) or (N,C,L), where N: batch size; C: number of input channels; H / W: feature map height / width; L: time series length).

[0066] During training, the stakes parameter (range 0-1) also needs to be initialized for each output channel.

[0067] Furthermore, after performing standard convolution on the extended convolution feature map samples, the standard convolution outputs the standard convolution feature map samples conv_out = conv(x), obtaining the stakes parameters of each channel within the gradient proof-of-stake mechanism convolution; the stakes parameters of each channel are normalized using the clamp function to obtain the normalized stakes parameters of each channel; the mean and standard deviation of the normalized stakes parameters of each channel are calculated; the normalized stakes parameters of each channel are standardized based on the mean and standard deviation to obtain the standardized stakes parameters of each channel; and the weights of each channel are calculated based on the standardized stakes parameters of each channel and the adaptive temperature.

[0068] Normalized scores:

[0069]

[0070]

[0071] In the formula, η is the learning rate, and L is the loss function value. Let s_i be the gradient of the values ​​of the i-th channel, and let s_i be the values ​​parameter of the i-th channel (which can be regarded as a function based on the channel gradient, reflecting the contribution rate of each channel to the loss function, and thus reflecting the performance of each channel).

[0072] Standardized stakes:

[0073] norm_stakes_i = (stakes_i - mean) / std

[0074] In the formula, mean is the mean of scores_i, std is the standard deviation of scores_i, i = 1, 2, 3... number of channels, and norm_stakes_i is the standardized scores parameter of the i-th channel.

[0075] Furthermore, before calculating the weights of each channel based on the standardized stakes parameters and adaptive temperature of each channel, the current training round of the initial recognition model is determined; temperature constraints for gradient proof-of-stake convolution are constructed using the PyTorch tensor truncation function based on the maximum and minimum temperature values; and adaptive temperature is calculated based on the temperature constraints according to the current training round and temperature parameters.

[0076] Calculate the adaptive temperature:

[0077] T = clamp(t, min, max)

[0078]

[0079] In the formula, clamp is a PyTorch tensor truncation function that ensures the temperature is between the minimum and maximum temperatures (e.g., within the range of (0.15, 0.35)), and t = a + sigmoid(epoch / b). k+c (where a, b, k, and c are model parameters that can be set by the user, for example, t=0.2+sigmoid(epoch / 10)) (0.3+0.1), epoch is the current training round, and temp is the adaptive temperature.

[0080] Adaptive temperature function: controls the degree of randomness in Gumbel-Softmax.

[0081] The temperature is higher in the early stages of training (≈0.3), allowing for more random exploration. As the epoch increases, the temperature decreases (→0.2), resulting in more precise selection.

[0082] Furthermore, the processing method for calculating the weights of each channel based on the standardized stakes parameters and adaptive temperature is to determine the Gumbel noise of the gradient proof-of-stake mechanism convolution; and to calculate the weights of each channel using the softmax function based on the standardized stakes parameters, adaptive temperature, and Gumbel noise of each channel.

[0083] Mathematically, temperature τ affects the shape of the weight distribution; the larger τ is, the smoother the distribution.

[0084] Calculate channel weights:

[0085] Weights_i=gumbel_softmax(logits=norm_stakes_i / temp

[0086] τ = temp) = softmax( (logits + Gumbel noise) / τ )

[0087] Where, Weights_i is the feature weight of the i-th channel.

[0088] In specific implementation, it is also necessary to determine the learnable residual parameters of the proof-of-stake mechanism convolution; calculate the residual coefficient through the sigmoid function according to the learnable residual parameters.

[0089] Calculate the residual coefficient:

[0090] α = sigmoid(residual_coef)

[0091] Where, α is the residual coefficient, residual_coef is the learnable residual parameter, and sigmoid(x) = 1 / (1 + e^(-x)).

[0092] In this embodiment, the proof-of-stake mechanism convolution outputs: y = conv_out (weights + α), and the output is the output feature map (i.e., the depth convolution feature map) after the convolution operation, and the size remains unchanged.

[0093] In the POS convolution, the input of each forward propagation is the feature map x (conv_out) and the current stakes parameter. Therefore, the stakes also need to be continuously updated during training (i.e., updated through standard backpropagation).

[0094] It should also be noted that the stakes determine the weight, α is a fixed value, the channels with high stakes (>0.7) are enhanced when y > conv_out in the output calculation, the channels with medium stakes are normally operated when y is close to conv_out in the output calculation, and the channels with low stakes are suppressed when y << conv_out in the output calculation.

[0095] Furthermore, after the model training is completed, it is necessary to preprocess the signal to be recognized to obtain a binary file; input the binary file into the signal recognition model to output the signal recognition result.

[0096] The signal type corresponding to the signal to be recognized is determined in advance. The signal type can be 1D or 2D. The signal to be recognized corresponding to 1D is time series data such as audio, and the signal to be recognized corresponding to 2D is two-dimensional data such as images (the image is a grayscale image). Then, preprocess the signal to be recognized (for example, if it is a signal sample corresponding to 1D, unify the sampling rate; if it is a signal sample corresponding to 2D, unify the resolution), and convert the preprocessed data into a binary file.

[0097] Since the signal recognition model is constructed based on a combination of an input layer, at least one improved bottleneck module, an output layer, and a classification head, the improved bottleneck module includes extended convolution, gradient proof-of-stake convolution, SE module, and linear bottleneck.

[0098] It should also be noted that the input layer is constructed by combining standard convolutional layers, batch normalization (BN) layers, and activation layers.

[0099] Furthermore, the standard convolutional layer performs standard convolution on the binary file to generate a standard convolutional feature map; the BN layer normalizes the standard convolutional feature map to obtain a normalized feature map; and the activation layer obtains the channel size feature map based on the normalized feature map using the Hardswish activation function.

[0100] It should be understood that the standard convolutional feature map corresponding to the binary file of 2D data is in (N,C,H,W) format, while the standard convolutional feature map corresponding to the binary file of 1D data is in (N,C,L) format.

[0101] The channel size feature map can be understood as the feature map corresponding to the channel size. For example, if the channel size is 16, then the corresponding feature map size is 16. 16.

[0102] Extended convolution expands the channel-size feature map corresponding to the signal to be identified transmitted from the input layer to obtain an extended convolution feature map.

[0103] It should also be noted that the channels need to be expanded based on the channel size feature map. The channel size feature map is split and input into multiple channels for feature processing. The output is an expanded convolutional feature map after merging the features of each channel (i.e., the feature map after expanding the channels).

[0104] In the specific implementation, expanded convolution (pw) is the number of channels expanded in a 1x1 convolution.

[0105] The gradient proof-of-stake mechanism convolution performs standard convolution on the expanded convolution feature map, and then performs convolution calculation on the standard convolution feature map, the weights of each channel, and the residual coefficients to obtain the depth convolution feature map.

[0106] The SE module obtains a weighted feature map from the depthwise convolutional feature map using the SE weight formula.

[0107] In the specific implementation, if multiple bottleneck modules are used in series, some of them can be selected to use SE, while the rest are not used. If layers 1, 4, 5, 7, and 8 are enabled, ReLU and Hardsigmoi activation functions are used alternately.

[0108] SE weight calculation formula:

[0109]

[0110] In the formula, W1, W2, b1, and b2 are the parameters of the first and second fully connected layers, respectively, and w all For the weighted feature map, x c is the feature map of channel c, and zall is the global average pooling result for all channels.

[0111] The linear bottleneck 1024 compresses the weighted feature map so that the output layer 103 can convert the compressed weighted feature map into a global vector.

[0112] It should also be noted that the linear bottleneck is a 1x1 convolutional compression channel. The weighted feature map can be compressed to a specified size (which can be larger than, less than or equal to, the size of the base feature map) through the linear bottleneck to obtain the compressed weighted feature map.

[0113] The output layer is constructed by combining extended convolutional layers, batch normalization (BN) layers, activation layers, and global average pooling layers: 1x1 extended convolution + BN + Hardswish + global average pooling.

[0114] The extended convolutional layer extends the compressed weighted feature map through convolution to obtain an extended convolutional feature map; the BN layer normalizes the extended convolutional feature map to obtain a normalized feature map; the activation layer uses the Hardswish activation function to obtain an activated normalized feature map based on the normalized feature map; the global average pooling layer transforms the activated normalized feature map into a vector to obtain a global vector.

[0115] The classification head performs signal identification on the signal to be identified based on the global vector.

[0116] The classification head is constructed from a combination of a fully connected layer, an activation layer, a Dropout layer, and a Softmax layer: fully connected layer + Hardswish + Dropout + fully connected layer + Softmax layer.

[0117] In this embodiment, the fully connected layer flattens the global vector to obtain a one-dimensional vector; the activation layer obtains an activated one-dimensional vector by applying the Hardswish activation function to the one-dimensional vector; the Dropout layer randomly drops one-dimensional vectors after activation; the fully connected layer obtains a global feature vector based on the randomly dropped one-dimensional vector; and the Softmax layer performs signal recognition based on the global feature vector.

[0118] The output signal recognition result is a certain object category with the highest probability value of the signal to be recognized.

[0119] In this embodiment, reference Figure 2 , Figure 2This is a structural block diagram of multiple improved bottleneck modules in the first embodiment of the signal recognition method based on gradient proof-of-stake convolution of the present invention. If multiple improved bottleneck modules exist, and if the multiple improved bottleneck modules are respectively improved bottleneck module A, improved bottleneck module B, ... improved bottleneck module N, the input layer inputs the channel size feature map into improved bottleneck module A for processing to obtain a compressed weighted feature map. Then, the compressed weighted feature map is input into improved bottleneck module B for processing again until improved bottleneck module N outputs a compressed weighted feature map. The compressed weighted feature map output by improved bottleneck module N is then input into the output layer for subsequent processing.

[0120] It should also be noted that the compressed weighted feature map output from the previous iteration will be used as the input to the bottleneck module for the next iteration in the loop.

[0121] In this embodiment, firstly, the extended convolution expands the channel-size feature map corresponding to the signal to be identified transmitted from the input layer, resulting in an extended convolutional feature map. Then, the gradient proof-of-stake mechanism convolution performs a standard convolution on the extended convolutional feature map, and calculates the convolutional features map, channel weights, and residual coefficients to obtain a depth convolutional feature map. Next, the SE module obtains a weighted feature map based on the depth convolutional feature map using the SE weight formula. A linear bottleneck compresses the weighted feature map so that the output layer transforms the compressed weighted feature map into a global vector. Finally, the classification head performs signal recognition based on the global vector. The gradient proof-of-stake mechanism convolution proposed in this embodiment solves the problems of high computational complexity and low efficiency of traditional attention mechanisms. The signal recognition model constructed based on the input layer, at least one improved bottleneck module, output layer, and classification head alleviates the numerical instability caused by mathematical transformations, thereby achieving accurate signal recognition.

[0122] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0123] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0124] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0125] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A signal recognition method based on gradient proof-of-stake convolution, characterized in that, The method includes: Obtain the stakes parameters of each channel within the gradient proof-of-stake mechanism convolution, where the stakes parameters of each channel represent the contribution rate of each channel to the loss function. The stakes parameter of each channel is normalized by using the clamp function to obtain the normalized stakes parameter of each channel. Calculate the mean and standard deviation of the normalized stakes parameter for each channel; The normalized scores parameters of each channel are standardized based on the mean and the standard deviation to obtain the standardized scores parameters of each channel. The weights of each channel are calculated based on the standardized stakes parameters and adaptive temperature of each channel. Determine the learnable residual parameters of the convolution in the gradient proof-of-stake mechanism; The residual coefficients are calculated using the sigmoid function based on the learnable residual parameters. Based on the channel weights and residual coefficients, the binary data samples corresponding to the signal samples are input into the initial recognition model for training; Based on the loss function value of the initial recognition model, the model parameters within the initial recognition model are updated in reverse until the loss function value is less than or equal to a preset threshold, and the trained initial recognition model is used as the signal recognition model. Input the binary file of the signal to be identified into the signal recognition model, and output the signal recognition result; The signal recognition model is constructed based on an input layer, at least one improved bottleneck module, an output layer, and a classification head. The improved bottleneck module includes extended convolution, gradient proof-of-stake convolution, SE module, and linear bottleneck. The input layer converts the binary file into a channel-size feature map; The extended convolution expands the channel size feature map to obtain an extended convolution feature map; The gradient proof-of-stake mechanism convolution performs standard convolution on the extended convolution feature map, and performs convolution calculation on the standard convolution feature map, the weights of each channel, and the residual coefficients to obtain the depth convolution feature map; The SE module and the linear bottleneck perform weighted and compressed processing on the deep convolutional feature map to obtain a compressed weighted feature map. The output layer converts the compressed weighted feature map into a global vector, so that the classification head can perform signal recognition based on the global vector.

2. The method as described in claim 1, characterized in that, After inputting the binary data samples corresponding to the signal samples into the initial recognition model for training based on the channel weights and residual coefficients, the process includes: Obtain the signal recognition result sample corresponding to the signal sample; The loss function value corresponding to the initial recognition model is calculated based on the signal recognition result sample and the actual recognition result corresponding to the signal sample.

3. The method as described in claim 1, characterized in that, Before calculating the weights of each channel based on the standardized stakes parameters and adaptive temperature, the following steps are included: Determine the current training round of the initial recognition model; The temperature constraints of the gradient proof-of-stake mechanism convolution are constructed using a PyTorch tensor truncation function based on the maximum and minimum temperature values. Based on the temperature constraints, an adaptive temperature is calculated according to the current training round and temperature parameters.

4. The method as described in claim 1, characterized in that, The calculation of channel weights based on the standardized stakes parameters and adaptive temperature of each channel includes: Determine the Gumbel noise in the convolution of the gradient proof-of-stake mechanism; The weights of each channel are calculated using the softmax function based on the standardized stakes parameters, adaptive temperature, and Gumbel noise.

5. The method as described in claim 1, characterized in that, The input layer is constructed by a combination of standard convolutional layers, batch normalization (BN) layers, and activation layers. The standard convolutional layer performs standard convolution on the binary file to generate a standard convolutional feature map. The BN layer normalizes the standard convolutional feature map to obtain a normalized feature map. The activation layer obtains the channel size feature map based on the normalized feature map using the Hardswish activation function.

6. The method as described in claim 1, characterized in that, The SE module performs weighted processing on the depthwise convolutional feature map to obtain a weighted feature map; The linear bottleneck compresses the weighted feature map to obtain a compressed weighted feature map.

7. The method as described in claim 1, characterized in that, The output layer is constructed by combining extended convolutional layers, batch normalization layers, activation layers, and global average pooling layers. The extended convolutional layer performs extended convolution on the compressed weighted feature map to obtain an extended convolutional feature map. The BN layer normalizes the expanded convolutional feature map to obtain a normalized feature map. The activation layer obtains the activated normalized feature map based on the normalized feature map using the Hardswish activation function; The global average pooling layer performs vector transformation on the activated normalized feature map to obtain a global vector.

8. The method as described in claim 1, characterized in that, The classification head is constructed from a combination of a fully connected layer, an activation layer, a Dropout layer, and a Softmax layer. The fully connected layer flattens the global vector to obtain a one-dimensional vector. The activation layer obtains the activated one-dimensional vector based on the one-dimensional vector using the Hardswish activation function; The Dropout layer randomly discards the activated one-dimensional vector; The fully connected layer obtains the global feature vector based on the randomly lost one-dimensional vector; The Softmax layer performs signal recognition based on the global feature vector.

Images