A signal recognition system based on a biological shallow brain architecture neural network

Abstract

The application discloses a signal recognition system based on a neural network of a biological shallow brain architecture, and belongs to the field of signal recognition. The system adopts a light-weight shallow network architecture. The architecture can greatly reduce the resource requirements of a model for calculation, storage and energy consumption by significantly simplifying the network depth and parameter scale, so that the model can efficiently operate on a resource-limited platform. The signal recognition speed of the model is significantly improved to meet the requirements of high real-time applications, thereby providing a basic solution for efficient and reliable deployment of AI technology in a wide range of scenarios. Simulation verification shows that the system provided by the application can achieve basic convergence after one round of training.

Description

Technical Field

[0001] This invention belongs to the field of signal recognition, and more specifically, relates to a signal recognition system based on a neural network with a biological shallow brain architecture. Background Technology

[0002] In the field of wireless communication, artificial intelligence (AI) technology is profoundly transforming network architecture and functionality. Among these technologies, deep learning, with its powerful end-to-end learning capabilities and superior performance, has become a key technology enabling communication systems. Deeply integrating AI functions into network infrastructure such as base stations is an important development trend for improving network intelligence. However, current mainstream deep learning network architectures generally have a large number of parameters, posing significant challenges to deployment on resource-constrained base station sides.

[0003] On the one hand, base stations have relatively limited hardware resources such as computing power and storage capacity, making it difficult to support complex models with large-scale parameters used for signal recognition. This leads to resource bottlenecks when base stations run advanced AI algorithms for signal recognition, affecting their overall performance and stability. On the other hand, the large number of model parameters and calculations significantly increases the computational burden in the signal recognition process, inevitably leading to high latency and reducing the system's real-time response capability. These challenges severely restrict the efficient application of signal recognition models in many demanding scenarios in the communication field, such as sensor-integrated communication (ISAC), environmental perception, real-time control, and intelligent interaction. Summary of the Invention

[0004] In view of the above-mentioned defects or improvement needs of the existing technology, the present invention provides a signal recognition system based on a neural network with a biological shallow brain architecture, thereby solving the problems of large number of parameters and high computational complexity of existing signal recognition models.

[0005] To achieve the above objectives, according to a first aspect of the present invention, a signal recognition system based on a neural network with a biological superficial brain architecture is provided, comprising:

[0006] M cascaded neural networks based on the biological superficial brain architecture, f(θ1), f(θ2), ..., f(θ) M Each of the neural networks includes a feature extraction module and a squeezing and excitation block connected in sequence; the feature extraction module is used to extract the signal features of the input signal, and the squeezing and excitation block is used to enhance the signal features to obtain enhanced signal features; wherein, the input signal of f(θ1) is the signal to be identified, and the input signals of other neural networks besides f(θ1) all include the signal to be identified and the prediction vector output by the previous neural network cascaded with it; the signal to be identified is a communication signal, radar signal, or integrated radar communication signal; M>1;

[0007] The feature fusion layer is used to calculate the mean vector of the prediction vectors output by M neural networks, and the category corresponding to the largest element in the mean vector is taken as the category of the signal to be identified.

[0008] According to a second aspect of the present invention, a training method for a signal recognition system based on a biological shallow brain architecture neural network as described in the first aspect is provided, comprising:

[0009] The signal recognition system is trained by using each signal in the training set as a sample signal and the true category of the sample signal as a label.

[0010] According to a third aspect of the present invention, a signal recognition method based on a neural network with a biological superficial brain architecture is provided, comprising:

[0011] The signal to be identified is input into a signal recognition system based on a neural network with a biological shallow brain architecture as described in the first aspect, and the category of the signal to be identified is obtained.

[0012] According to a fourth aspect of the present invention, an electronic device is provided, comprising: a computer-readable storage medium and a processor;

[0013] The computer-readable storage medium is used to store executable instructions;

[0014] The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the method described in the second or third aspect.

[0015] According to a fifth aspect of the invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to perform the method as described in the second or third aspect.

[0016] According to a sixth aspect of the invention, a computer program product is provided, comprising a computer program or instructions that, when executed by a processor, implement the method as described in the second or third aspect.

[0017] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0018] To address the common challenge of deploying deep learning models in resource-constrained environments, this invention, inspired by the efficient information processing structure of the shallow brain hypothesis and incorporating its flattened brain design principles, proposes a lightweight shallow network architecture with hierarchical perception and output. This architecture differs fundamentally from traditional deep networks, breaking away from strict hierarchical order and allowing deeper layers to directly integrate partially processed intermediate features as well as raw or low-order inputs. It can also synthesize feature information from different "depths" or "scales" of the network in parallel or adaptively. By significantly reducing network depth and parameter size, this architecture drastically reduces the model's resource requirements for computation, storage, and energy, enabling efficient operation on resource-constrained platforms. Furthermore, it significantly improves the model's signal recognition speed to meet the demands of high real-time applications, thus providing a fundamental solution for the efficient and reliable deployment of signal recognition technology in a wide range of scenarios (such as sensory integration (ISAC), environmental perception, real-time control, and intelligent interaction). Simulation results demonstrate that the system provided by this invention achieves basic convergence after one round of training. Attached Figure Description

[0019] Figure 1 A structural diagram of a signal recognition system based on a biological shallow brain architecture provided in an embodiment of the present invention;

[0020] Figure 2 A neural network structure diagram in a signal recognition system based on a biological shallow brain architecture, provided for an embodiment of the present invention;

[0021] Figure 3 This is a schematic diagram of data splicing provided in an embodiment of the present invention;

[0022] Figure 4 A graph showing the change in accuracy over training rounds provided in an embodiment of the present invention;

[0023] Figure 5 The confusion matrix diagram of a signal recognition system based on a biological shallow brain architecture neural network provided in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0025] The mainstream design paradigm of current deep learning model architectures is generally inspired by the early hierarchical processing mechanism of the visual cortex, typically employing a multi-layered, sequential feedforward structure. In this architecture, the raw input data is first processed by the lower-level network, and its output is passed layer by layer as features to higher levels for abstraction and integration, ultimately forming a decision or output. Classic and widely used models such as ResNet, CNN, DNN, LSTM, and Transformer all embody this hierarchical design concept.

[0026] However, current mainstream deep learning network architectures for signal recognition generally have a large number of parameters, which poses a significant challenge when deployed at resource-constrained base stations.

[0027] Based on this, embodiments of the present invention provide a signal recognition system based on a neural network with a biological superficial brain architecture, comprising:

[0028] M cascaded neural networks based on the biological superficial brain architecture, f(θ1), f(θ2), ..., f(θ) M Each of the neural networks includes a feature extraction module and a squeezing and excitation block connected in sequence; the feature extraction module is used to extract signal features of the input signal, and the squeezing and excitation block is used to enhance the signal features to obtain enhanced signal features; wherein, the input signal of f(θ1) is the signal to be identified, and the input signals of other neural networks besides f(θ1) all include the signal to be identified and the output of the previous neural network cascaded with it; the signal to be identified is a communication signal, radar signal, or integrated radar communication signal; M>1;

[0029] The feature fusion layer is used to calculate the mean vector of the prediction vectors output by M neural networks, and the category corresponding to the largest element in the mean vector is taken as the category of the signal to be identified.

[0030] Preferably, the system further includes a preprocessing module for preprocessing the signal to be identified;

[0031] The preprocessing module includes a frequency domain feature extraction module and a logarithmic normalization module; the frequency domain feature extraction module is used to perform a second FFT process on the signal to be identified to extract the frequency domain features of the signal to be identified, and the logarithmic normalization module is used to perform logarithmic normalization processing on the frequency domain features;

[0032] Preferably, the preprocessing module further includes a channel enhancement module connected to the logarithmic normalization module, used to expand the single-channel data output by the logarithmic normalization module into dual-channel data.

[0033] Preferably, the feature extraction module includes at least one basic residual block; the basic residual block includes a convolutional layer, a normalization layer and a feature splicing layer connected in sequence.

[0034] It should be noted that the training method for the signal recognition system based on the biological shallow brain architecture described above can be achieved using conventional training methods, such as end-to-end training. Based on this, this embodiment of the invention provides a training method for a signal recognition system based on the biological shallow brain architecture as described in any of the above embodiments, comprising:

[0035] The signal recognition system is trained by using each signal in the training set as a sample signal and the true category of the sample signal as a label.

[0036] That is, the signal recognition system is trained by minimizing the difference between the categories of sample signals predicted by M neural networks and their true categories.

[0037] Preferably, the loss function used for training includes the cross loss of each neural network and the supervised contrastive loss.

[0038] Preferably, the prediction vectors of the M neural networks are input into the Transformer attention module, which concatenates the prediction vectors output by each neural network and then sequentially performs position encoding, multi-head attention, and spatial attention calculations to obtain a feature representation that integrates information from different positions and spatial dimensions. The feature representation is then subjected to dimensional transformation and nonlinear activation processing to map it to a preset category dimension space and convert it into a probability distribution form. The category with the highest probability is taken as the signal category of the sample signal.

[0039] The loss function used for training also includes the cross-entropy loss of the Transformer attention module.

[0040] The signal recognition system based on a biological shallow brain architecture provided in this embodiment of the invention includes M sequentially cascaded neural networks f(θ1), f(θ2), ..., f(θ) based on a biological shallow brain architecture. M The specific value of M can be set according to the size of the training dataset, the network training speed, and the memory size of the deployed hardware devices.

[0041] The following example, with M=4, further illustrates the signal recognition system based on the biological shallow brain architecture neural network provided by this invention.

[0042] like Figure 1As shown, the input signal X is simultaneously input to four sequentially cascaded neural networks based on the biological shallow brain architecture: f(θ1), f(θ2), f(θ3), and f(θ4). The input of f(θ2) also includes the output of the previous neural network f(θ1) cascaded with f(θ2); the input of f(θ3) also includes the output of the previous neural network f(θ2) cascaded with f(θ3); and the input of f(θ4) also includes the output of the previous neural network f(θ3) cascaded with f(θ4). The outputs of the four neural networks are integrated by a Transformer attention module. Loss calculation is performed based on the outputs of the four neural networks and the Transformer attention module to train the signal recognition system.

[0043] The neural networks have the same structure, all including a feature extraction module and a squeeze-and-excitation block (SEB) connected in sequence.

[0044] The feature extraction module can use any existing feature extraction network, such as ResNet18, CNN, RNN, LSTM, etc.

[0045] As an example, the feature extraction module includes at least one basic residual block. When the number of basic residual blocks is 2, the structure of the neural network is as follows: Figure 2 As shown.

[0046] Each basic residual block has the same structure, including different types of layers, such as convolutional layers, normalization layers, and shortcat layers for feature concatenation. Each layer has a corresponding strides, kernel size, and activation function settings. In this example, the parameters of each layer of the basic residual block are shown in Table 1.

[0047] Table 1 Basic Residual Blocks

[0048]

[0049] To further enhance feature representation and prediction performance, a squeeze and excitation block (SEB) is connected after the feature extraction module.

[0050] The SEB calculation process includes the following stages:

[0051] First, recalibrate the channel features:

[0052]

[0053] Among them, U c ∈R H′×W′These are elements in the SEB output features, where H′ and W′ represent the height and width of the output features, respectively. This is achieved by combining the corresponding C convolutional kernels into V... c =[v c,1 ,…,v c,C The input features Q∈R of SEB C×H×W We obtain it by performing a convolution, where * represents the convolution operation, v c,t ∈R k×k Let Q represent the t-th convolutional kernel in the c-th combination. t ∈R H×W For the t-th input channel, H and W represent the height and width of the input feature.

[0054] Next, global average pooling is used to extract the spatial features U. c Compressed to channel descriptor Z c :

[0055]

[0056] Among them, U c (i,j) represents feature U c The value of the coordinate (i,j) in the middle, Z c For the convolutional features U c The scalar value obtained by global average pooling.

[0057] Finally, the channel incentive weights are generated:

[0058] Using channel descriptor Z = [Z1, Z2, ..., Z c The channel excitation weight S is calculated based on the formula S = SIGMOID(W2 * ReLU(W1Z)), where W1 ∈ R. (C / r)×C Channel dimension reduction is performed, accompanied by ReLU activation. The channel dimension is then restored, and channel incentive weights are finally generated through Sigomid activation. Then S is broadcast to a spatial dimension that matches Q (i.e., the weights of 1×1×C are expanded to C×H×W, with each channel sharing the same weights in its spatial position), and then multiplied element-wise with Q channel by channel as the output of the neural network.

[0059] When the feature extraction module includes at least one basic residual block, the system provided by this invention is also called a Shallow Brain ResNet Network (SBRN).

[0060] Preferably, the system provided in this embodiment of the invention further includes a preprocessing module, which is used to preprocess the input signal X before inputting it into each neural network, so as to improve the quality of the input data and the processing efficiency and feature extraction effect of the neural network.

[0061] The preprocessing module performs the following preprocessing on the input signal X:

[0062] 1. Data partitioning

[0063] First, the data in the dataset are divided according to their respective categories, and then they are organized into samples of size M×N, with M=5 and N=256.

[0064] 2. Extract frequency domain features

[0065] Perform Fast Fourier Transform (FFT) calculations on data with input dimension M×N:

[0066]

[0067] in, Represents the original input signal. This represents the result obtained after the first FFT calculation. It is a two-dimensional spectrum obtained after a second FFT transformation.

[0068] X m,n ,Y m,n and Z m,n These represent the elements in the m-th row and n-th column of matrices X, Y, and Z, respectively.

[0069] 3. Logarithmic normalization

[0070] Apply a logarithmic transformation to each element of the complex matrix Z:

[0071]

[0072] in μ represents the real-valued matrix obtained after the transformation. Z It represents the average magnitude of all elements in the matrix.

[0073] 4. Data channel expansion

[0074] To accelerate model convergence and improve model robustness, a dual-channel data augmentation strategy is adopted, transforming the single-channel input obtained after the above transformation... Transformed into a two-channel representation using zero-fill enhancement. The second channel is initialized to null, which preserves the original feature distribution and provides additional capacity for gradient flow optimization.

[0075] The preprocessed data is fed into four neural networks. These four neural networks are organized according to a specific hierarchical structure: the output of the lower-level neural networks is further used as input to the higher-level neural networks. When two input sources are fed into a neural network, the data is concatenated, and this concatenation process is as follows: Figure 3 As shown, the data splicing process increases the number of data channels.

[0076] Furthermore, considering the potential dimensionality mismatch between input channels, the PyTorch function `Interpolate()` is used to appropriately resize the preprocessed data before data concatenation. This ensures that the dimensions of the input channels match, thereby guaranteeing the successful completion of the data concatenation operation.

[0077] A loss function is set to train the system provided in the embodiments of the present invention. The loss function includes:

[0078] 1. Cross-loss in neural networks

[0079] For each neural network f(θ) i ), calculate the corresponding cross-entropy loss. correspond Figure 1 h(φ) i Meanwhile, the loss output corresponds to Figure 1 q in i .set up Represents a neural network f(θ) i The predicted probability distribution across all categories, where Let y be the predicted probability of class c∈C; y is the true probability distribution. c Let C represent the true membership probability of class c, where C is the complete set of classes. Cross-entropy loss. The calculation formula is as follows:

[0080]

[0081] 2. Neural Network Supervised Contrast Loss

[0082] Supervisory comparison loss (correspond Figure 1 g(ψ) i The calculation is performed at two levels: instance level and batch level, and the output of the loss is... Figure 1 p shown i .

[0083] For the neural network f(θ) i The embedded feature vector q extracted from sample ji,j Its instance-level loss The definition is as follows:

[0084]

[0085] in, Indicates in batch B i The set of positive sample indices that have the same true class as sample j; The set of negative samples is shown. <·,·> are operators for calculating the cosine similarity between feature vectors, and τ>0 is the temperature hyperparameter.

[0086] Batch-level losses By analyzing batch B i The instance-level loss of all samples is aggregated, and the calculation formula is as follows:

[0087]

[0088] To improve the recognition accuracy of the system, preferably, the training process further includes: inputting the prediction vectors of the M neural networks into the Transformer attention module, which concatenates the prediction vectors output by each neural network, and then sequentially performs position encoding, multi-head attention and spatial attention calculations to obtain feature representations that fuse information from different positions and spatial dimensions. The feature representations are then subjected to dimensional transformation and nonlinear activation processing to map them to a preset category dimension space, and converted into a probability distribution form. The category with the highest probability is taken as the signal category of the sample signal.

[0089] Accordingly, the loss function used for training also includes the cross-entropy loss of the Transformer attention module.

[0090] Specifically, the features output by each neural network are fed into the Transformer attention module, which captures global dependencies between feature locations while emphasizing important local spatial information. Its computational process includes the following stages:

[0091] Location coding:

[0092] First, the input features (Where B represents the batch size, N represents the sequence length, and d represents the embedding dimension) Apply positional encoding, then inject sequence information:

[0093]

[0094] Here, G represents the input feature. It is a learnable positional encoding matrix. Indicates the outer product operation, 1 BIt is a B-dimensional all-1 vector, the enhanced position-aware feature tensor. The original feature dimensions were preserved.

[0095] Multi-head attention calculation:

[0096] Using a multi-head attention mechanism, the position-encoded input G′ is transformed by H parallel attention heads. The formula for calculating a single attention head h is:

[0097]

[0098] By outputting the projection matrix The outputs of multiple attention heads are aggregated to form an integrated representation.

[0099] O = [CONCAT(O1,…,O] H )]W O

[0100] in, and These are projection matrices, used to convert the input X′ into query, key, and value representations, respectively. k and d v Each represents its own projection dimension. CONCAT(·) means concatenating vectors along the last dimension, preserving the independent contribution of each attention head.

[0101] Spatial attention calculation:

[0102] Spatial attention is computed based on the multi-head attention output to capture location-sensitive features:

[0103]

[0104] in, It can be used as the weight matrix of a 3-size 1D convolutional filter for dimension-adjusted features. Parameterize, W c,k O′ represents the scalar element in the c-th row and k-th column of the weight matrix W. b,c,n+k Features and bias terms at index batch b, channel c, and time position (n+k) (with zero-padding for boundary conditions). Spatial attention weights are generated by sigmoid activation and modulated through the Hadamard product (⊙) to obtain the output.

[0105] Will The input is fed into the fully connected layer of the Transformer attention module for dimensional transformation and nonlinear activation processing to map it to a preset category dimensional space, and then converted into a probability distribution form. The category with the highest probability is taken as the signal category of the sample.

[0106] The cross-entropy loss of the Transformer attention layer output is calculated using the formula described above.

[0107] Final total loss function The loss consists of the cross-entropy loss of each neural network, the supervised contrastive loss, and the cross-entropy loss of the Transformer attention layer. The calculation formula is as follows:

[0108]

[0109] The following simulation experiment demonstrates the effectiveness of radar signal recognition for UAVs in Sensor-Integrated Communication (ISAC) by using radar signals as an example.

[0110] The training dataset used in the simulation is derived from a publicly available dataset acquired by a SAAB SIRS1600 frequency-modulated continuous wave (FMCW) radar. This radar operates in the 77 GHz band and possesses high-precision range and velocity measurement capabilities, effectively capturing the micro-Doppler features of small targets. The targets in the dataset range from 5 to 200 meters. This wide range of distance variation causes the signal to experience varying degrees of attenuation and interference during propagation, resulting in diverse signal-to-noise ratio (SNR) conditions. This characteristic significantly enhances the complexity and realism of the dataset, enabling the system to learn a broader range of target feature representations.

[0111] The training dataset contains data from ten different categories. The specific classifications and sample size distribution are shown in Table 2. The drones include six models, five of which are quadcopter drones and one is a hexacopter drone.

[0112] Table 2 Training Dataset

[0113]

[0114] The system training platform and training-related parameter settings are shown in Table 3.

[0115] Table 3 System Training Platform and Training-Related Parameters

[0116] Training hardware platform NVIDIA GeForce RTX 3090 Dataset split ratio Training set 80%, test set 20% Network weight initialization Kaiming initialization Gradient backpropagation Adam optimizer Learning rate <![CDATA[1×10 -3 ]]> Weight decay <![CDATA[1×10 -4 ]]> SEB hyperparameters 4 Transformer module hyperparameters 8

[0117] Simulation experiments were conducted to record and compare the accuracy changes of different models during the training process. Specific details are as follows: Figure 4 And as shown in Table 4.

[0118] Figure 4 This demonstrates the accuracy performance of each model at each training epoch throughout the entire training process until convergence, clearly showing the evolution of the performance of different models as training progresses. From Figure 4 It can be seen that the SBRN network provided by this invention has basically converged after the first round of training; and compared with the accuracy of each network after convergence, the SBRN network provided by this invention has the highest accuracy.

[0119] Table 4 compares the accuracy of each model at specific sampling training intervals (epochs 1, 5, 10, and 20). The rightmost column shows the peak accuracy of each model after convergence. The following table compares common models such as ResNet18, ResNet34, ViT, and ResNet-MOSE with the model proposed in this invention:

[0120] Table 4 Simulation Data

[0121]

[0122] To quantitatively evaluate the performance of the SBRN network model provided in this invention Figure 5 The confusion matrix of all model outputs is shown, where the vertical axis represents the true class label and the horizontal axis shows the distribution of the model's predicted classes. Figure 5 As can be seen, within the system framework provided by this invention, the diagonal elements of these matrices consistently exceed 0.94 for each neural network, indicating their inherent classification ability. More importantly, when the mean-aggregated feature vectors from all four neural networks are used as the classification criterion, the confusion matrix exhibits even higher diagonal element values, exceeding 0.96. This reflects that by fusing the outputs of each submodule, the accuracy of each category is enhanced.

[0123] In summary, the system provided by this invention proposes a novel network architecture for signal recognition. This network architecture has significant advantages, with a small number of network parameters, and can be applied to various general scenarios that require signal recognition tasks.

[0124] This invention provides a signal recognition method based on a neural network with a biological superficial brain architecture, comprising:

[0125] The signal to be identified is input into a signal recognition system based on a neural network with a biological shallow brain architecture as described in any of the above embodiments to obtain the category of the signal to be identified.

[0126] This invention provides an electronic device, including: a computer-readable storage medium and a processor;

[0127] The computer-readable storage medium is used to store executable instructions;

[0128] The processor is used to read executable instructions stored in the computer-readable storage medium and execute the training method or signal recognition method as described in any of the above embodiments.

[0129] This invention provides a computer-readable storage medium storing computer instructions for causing a processor to execute a training method or signal recognition method as described in any of the above embodiments.

[0130] This invention provides a computer program product, including a computer program or instructions, which, when executed by a processor, implement the training method or signal recognition method as described in any of the above embodiments.

[0131] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A signal recognition system based on a neural network with a biological superficial brain architecture, characterized in that, include: M cascaded neural networks based on the biological shallow brain architecture , ,…, Each of the aforementioned neural networks includes a feature extraction module and a squeezing and excitation block connected in sequence; the feature extraction module is used to extract signal features from the input signal, and the squeezing and excitation block is used to enhance the signal features to obtain enhanced signal features; wherein... The input signal is the signal to be identified, except The input signals of all other neural networks include the signal to be identified and the prediction vector output by the previous neural network cascaded with it; the signal to be identified is a communication signal, radar signal, or integrated radar-communication signal; M>1; The feature fusion layer is used to calculate the mean vector of the prediction vectors output by M neural networks, and the category corresponding to the largest element in the mean vector is taken as the category of the signal to be identified. The loss function used in training includes the cross loss of each neural network, the supervised contrast loss, and the cross-entropy loss of the Transformer attention module. During training, the prediction vectors of the M neural networks in the signal recognition system are input into the Transformer attention module. The module concatenates the prediction vectors output by each neural network and then performs position encoding, multi-head attention, and spatial attention calculations in sequence to obtain a feature representation that integrates information from different positions and spatial dimensions. The feature representation is then subjected to dimensional transformation and nonlinear activation to map it to a preset category dimension space and convert it into a probability distribution form. The category with the highest probability is taken as the signal category of the sample signal.

2. The system as described in claim 1, characterized in that, It also includes a preprocessing module for preprocessing the signal to be identified; The preprocessing module includes a frequency domain feature extraction module and a logarithmic normalization module; the frequency domain feature extraction module is used to perform a second FFT process on the signal to be identified to extract the frequency domain features of the signal to be identified, and the logarithmic normalization module is used to perform logarithmic normalization processing on the frequency domain features.

3. The system as described in claim 2, characterized in that, The preprocessing module also includes a channel enhancement module connected to the logarithmic normalization module, used to expand the single-channel data output by the logarithmic normalization module into dual-channel data.

4. The system as described in claim 1, characterized in that, The feature extraction module includes at least one basic residual block; the basic residual block includes a convolutional layer, a normalization layer and a feature splicing layer connected in sequence.

5. A training method for a signal recognition system, characterized in that, The signal recognition system is a signal recognition system based on a neural network with a biological superficial brain architecture as described in any one of claims 1-4, and the training method includes: The signal recognition system is trained by using each signal in the training set as a sample signal and the true category of the sample signal as a label.

6. A signal recognition method based on a neural network with a biological superficial brain architecture, characterized in that, include: The signal to be identified is input into the signal recognition system based on the neural network architecture of the biological shallow brain as described in any one of claims 1-4, and the category of the signal to be identified is obtained.

7. An electronic device, characterized in that, include: Computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is used to read executable instructions stored in the computer-readable storage medium and execute the training method as described in claim 5 or the recognition method as described in claim 6.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a processor to execute the training method as described in claim 5 or the recognition method as described in claim 6.

Images