Methods and Systems for Spectrum Data Compression and Backhaul and Accelerated Model Inference
By using a deep neural network for signal compression and recovery with joint training and structure optimization, the problems of high transmission overhead, low recognition accuracy, and slow inference speed in spectral data backhaul in wireless communication are solved, achieving efficient data transmission and recognition.
Patent Information
- Application Number
- CN202310059701.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-01-16
- Publication Date
- 2025-12-02
- Estimated Expiration
- 2043-01-16
AI Technical Summary
Existing technologies suffer from high transmission overhead, low recognition accuracy, and slow inference speed when transmitting spectrum data back in wireless communication. In particular, the training gradient of the upsampling module in the single-branch structure of deep neural networks is small and the convergence speed is slow.
A signal compression deep neural network and a signal recovery deep neural network are designed for joint training. An upsampling module with a multi-branch structure is adopted and fused into a single-branch structure during the inference stage. The signal compression and recovery deep neural networks are combined and deployed in sensing and recognition devices. The compressed data is transmitted through a wireless link for signal recovery and recognition.
It achieves low transmission overhead, high recognition accuracy, and fast inference speed, reducing the amount of data transmitted while improving recognition accuracy and training convergence speed.
Smart Images
Figure CN116070677B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of signal processing methods, and in particular to a method and system for spectral data compression and backhaul and model inference acceleration. Background Technology
[0002] Deep learning technology's powerful automatic feature extraction capabilities have led to its widespread application in image recognition, recommendation systems, object detection, and speech understanding. Driven by the vision of AI-enabled wireless communication, deep learning technology, particularly deep neural networks, has begun to be applied to typical scenarios such as modulation signal recognition, interference type recognition, and intrusion attack recognition, enhancing the intelligence level of wireless communication. For example, AMR-oriented Deep Neural Networks (ADNNs) use wireless communication signals as input data, employing techniques such as deep convolutional neural networks, recurrent neural networks, and attention mechanisms for data processing, and then outputting the modulation type corresponding to the signal. This achieves high recognition accuracy while significantly reducing human intervention and exhibiting strong generalization capabilities. However, transmitting sampled signal data from wireless sensing devices back to the ADNN-deployed recognition devices via wireless links incurs significant transmission overhead and latency.
[0003] Figure 1 A schematic diagram of a scenario including a radiation source, a sensing device, and a recognition device is provided. The sensing device receives the wireless communication signal from the radiation source and transmits the received data back to the recognition device via a wireless link. The recognition device then runs an ADNN model to determine the modulation type used by the radiation source. To reduce the amount of data transmitted via the wireless link, current methods often employ data compression coding. However, these methods cannot change the size of the original data and are often sensitive to packet loss along the path, making it difficult to cope with the challenges posed by small amounts of data loss. Furthermore, the single-branch upsampling modules commonly used in deep neural networks, such as nearest neighbor and transposed convolution, generally suffer from small training gradients, slow training convergence speed, and low accuracy. Summary of the Invention
[0004] The technical problem to be solved by the present invention is how to provide a spectrum data compression and backhaul and model inference acceleration processing method with low transmission overhead, high recognition accuracy and fast inference speed.
[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for spectrum data compression and backhaul and model inference acceleration, characterized by comprising the following steps:
[0006] S101: Joint Training of Deep Neural Network Models: Design deep neural network structures for signal compression and signal recovery, and perform joint training based on given signal data;
[0007] S102: Upsampling module structure inference optimization: The upsampling module with a multi-branch structure is merged into a single-branch structure to increase the inference speed of the signal recovery deep neural network;
[0008] S103: Deployment of deep neural network models: Deploy the trained signal compression deep neural network to the sensing device, and deploy the signal recovery deep neural network to the recognition device;
[0009] S104: Signal data acquisition and transmission: The sensing device inputs the acquired wireless communication signal into the signal compression deep neural network, and transmits the resulting compressed data to the identification device through a wireless link;
[0010] S105: Signal Data Recovery and Recognition: The recognition device inputs the shrinkage data received through the wireless link into the signal recovery deep neural network to obtain enhanced data, and then inputs it as signal data into the ADNN model to obtain the modulation type of the signal data.
[0011] A further technical solution is that the input of the signal compression deep neural network is signal data, the output is shrunken data, and it includes a data downsampling module; the input of the signal recovery deep neural network is shrunken data, the output is enhanced data, and it includes an upsampling module with a multi-branch structure.
[0012] A further technical solution is that step S101 specifically includes the following steps:
[0013] S1011: Design of a Deep Neural Network for Signal Compression: Design the model structure of a deep neural network for signal compression;
[0014] S1012: Design of a Deep Neural Network for Signal Recovery: Design the model structure of a deep neural network for signal recovery;
[0015] S1013: Joint training of neural network models: Jointly train signal compression deep neural networks and signal recovery deep neural networks.
[0016] A further technical solution is that, in step S1013:
[0017] Joint training refers to updating the overall model parameters of two or more network models using a related error loss function;
[0018] Training samples refer to a set of signal samples and their corresponding modulation type labels, which can be collected using public datasets or wireless communication devices.
[0019] A further technical solution is that the joint training includes the following steps:
[0020] The output layer of the signal compression deep neural network is connected to the input layer of the signal recovery deep neural network to form a whole network. The input is training samples and the output is augmented samples. Then, the error loss between the signal samples and the augmented data is calculated. The parameters of the signal compression deep neural network and the signal recovery deep neural network are updated simultaneously through backpropagation of the error loss, so as to achieve joint training.
[0021] A further technical solution is that, in step S102:
[0022] The upsampling module has a multi-branch structure during the training phase, simultaneously receiving data input and calculating the result data separately, and then concatenating the result data into the final output data;
[0023] Fusion into a single-branch structure refers to the lossless transformation of each branch in the upsampling module into a single-branch structure during the inference phase through structural reparameterization.
[0024] The structural reparameterization refers to first constructing a series of structures and then equivalently converting their parameters into another set of parameters, thereby converting this series of structures into another series of structures.
[0025] This invention also discloses a spectrum data compression and backhaul and model inference acceleration processing system, characterized by comprising:
[0026] The system includes a sensing device, an identification device, and a wireless communication network. The sensing device is used to receive signal data sent by a radiation source and to compress the received signal data and obtain compressed data through a signal compression deep neural network with a downsampling module.
[0027] Wireless communication network used to transmit shrinkage data to identification devices;
[0028] The identification device is used to receive contraction data, recover signal data through a signal recovery deep neural network with an upsampling module, and transmit the signal data to an ADNN model for processing. The ADNN model processes the signal data to obtain the modulation type of the signal data.
[0029] A further technical solution is that the signal compression deep neural network includes four compression modules: a first compression module of 128*512, a second compression module of 64*256, a third compression module of 32*128, and a fourth compression module of 16*64. The first to fourth compression modules are connected in sequence. The first to fourth compression modules include Conv convolutional layers, ReLU activation function layers, BN batch normalization processing layers, and MaxPool max pooling layers.
[0030] A further technical solution is that the signal recovery deep neural network includes four recovery modules: a first recovery module of 16*128, a second recovery module of 32*256, a third recovery module of 64*512, and a fourth recovery module of 2*1024. The first recovery module to the fourth recovery module are connected in sequence. The first recovery module to the fourth recovery module includes a Conv convolutional layer, a ReLU activation function layer, a BN batch normalization processing layer, and an Upsample upsampling layer.
[0031] A further technical solution is that the ADNN model includes: a first processing module of 64*512, a second processing module of 64*256, a third processing module of 64*128, a fourth processing module of 64*64, a fifth processing module of 64*32, a sixth processing module of 64*2, a seventh processing module of 1*64, and an eighth processing module of 1*24. The first to fifth processing modules include Conv convolutional layers, ReLU function layers, BN batch normalization layers, and MaxPool max pooling layers. The sixth processing module includes Conv convolutional layers, ReLU function layers, BN batch normalization layers, and AvgPool average pooling layers. The seventh to eighth processing modules are linear layers.
[0032] The beneficial effects of adopting the above technical solution are as follows: First, the transmission overhead is small. By introducing a signal compression deep neural network, the signal data is compressed into smaller dimensional shrinkage data for transmission, eliminating the need to send the collected signal data back to the recognition device, thus reducing the transmission overhead.
[0033] Second, it has high recognition accuracy. Compared with network models using a single-branch upsampling structure, the designed multi-branch upsampling structure network model can increase the gradient size during the training phase and improve the recognition accuracy of the ADNN model during the training phase.
[0034] Third, it has a fast inference speed. It integrates the upsampling module, which adopts a multi-branch structure, into a single-branch structure, which reduces the inference computation of the signal recovery deep neural network and improves the inference speed. Attached Figure Description
[0035] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.
[0036] Figure 1 This is a schematic diagram of the overall scene of the system described in the embodiment of the present invention;
[0037] Figure 2 This is a schematic diagram of the data processing flow of the system described in the embodiment of the present invention;
[0038] Figure 3 This is a flowchart of the method described in the embodiments of the present invention;
[0039] Figure 4 This is a schematic diagram of the I and Q channels of the wireless communication signal in the method described in the embodiment of the present invention;
[0040] Figure 5 This is a schematic diagram of the fusion of a multi-branch structure into a single-branch structure in the method described in the embodiments of the present invention;
[0041] Figure 6 This is a diagram of the ADNN model structure;
[0042] Figure 7 This is a schematic diagram of the design and training process of a deep neural network for signal compression and recovery;
[0043] Figure 8 This is a diagram of a deep neural network structure for signal compression.
[0044] Figure 9 This is a diagram of a deep neural network structure for signal recovery;
[0045] Figure 10 This is a flowchart of joint training of neural networks;
[0046] Figure 11 This is a visual representation of the feature map of compressed data;
[0047] Figure 12 This is a two-way diagram illustrating the IQ of augmented data;
[0048] Figure 13 This is a graph showing the error loss results during the training process of a deep neural network for signal compression and recovery.
[0049] Figure 14 This is a graph showing the accuracy results during the training process of the ADNN model;
[0050] Figure 15 This is a diagram of the structure of a single-branch transposed convolution model;
[0051] Figure 16 This is a diagram of the single-branch nearest neighbor model structure. Detailed Implementation
[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0053] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0054] like Figure 2 As shown, this embodiment of the invention discloses a spectrum data compression and backhaul and model inference acceleration processing system, comprising:
[0055] Sensing devices, identification devices, and wireless communication networks;
[0056] The sensing device is used to receive signal data sent by a radiation source, and to compress the received signal data and obtain compressed data through a signal compression deep neural network with a downsampling module.
[0057] Wireless communication network used to transmit shrinkage data to identification devices;
[0058] The identification device receives contraction data, recovers the signal data through a signal recovery deep neural network with an upsampling module, and transmits the signal data to an ADNN model for processing. The ADNN model then processes the signal data to obtain the modulation type of the signal data.
[0059] like Figure 3 As shown in the figure, this invention also discloses a method for spectrum data compression and backhaul and model inference acceleration, characterized by the following steps:
[0060] S101: Joint Training of Deep Neural Network Models: Design signal compression deep neural network and signal recovery deep neural network structures, and use the given signal data as samples for joint training. The input of the signal compression deep neural network is the signal data and the output is the shrunken data. The input of the signal recovery deep neural network is the shrunken data and the output is the enhanced data. It also includes an upsampling module with a multi-branch structure.
[0061] The signal data refers to the data representation of wireless communication signals, including IQ two-channel data, constellation diagrams, spectrum diagrams, etc.
[0062] For example, Figure 4 The signal data displayed uses a two-channel (IQ) representation method, which represents the amplitude and phase information of each wireless communication signal as two mutually orthogonal carrier components as IQ data. Each signal contains 1024 data points, and the IQ data together form a sample.
[0063] S102: Upsampling module structure inference optimization: The upsampling module with a multi-branch structure is merged into a single-branch structure to increase the inference speed of the signal recovery deep neural network;
[0064] The upsampling module has a multi-branch structure during the training phase. It simultaneously receives data input, calculates the result data separately, and concatenates the result data into the final output data. The branch structure can be a method such as linear interpolation, nearest neighbor interpolation, or transposed convolution.
[0065] The fusion into a single-branch structure refers to the lossless transformation of each branch in the upsampling module into a single-branch structure during the inference phase through structural reparameterization.
[0066] The structure reparameterization refers to first constructing a series of structures (generally used for training) and then converting their parameters into another set of parameters (generally used for inference), thereby converting this series of structures into another series of structures.
[0067] For example, Figure 5 This shows a possible upsampling module design consisting of three branches. The first branch consists of a nearest neighbor interpolation layer, a convolutional layer, and a batch normalization layer. The second branch consists of a transposed convolutional layer and a batch normalization layer. The third branch consists of a transposed convolutional layer and a batch normalization layer. After structural reparameterization, the multi-branch structure is losslessly converted into a single-branch structure, which consists of a single transposed convolutional layer.
[0068] The transposed convolution refers to the inverse process of a convolution operation, which is usually used for sample generation and reconstruction to transform low-dimensional feature maps into high-dimensional feature maps.
[0069] S103: Deployment of deep neural network models: Deploy the trained signal compression deep neural network to the sensing device, and deploy the signal recovery deep neural network to the recognition device;
[0070] The sensing device refers to a device equipped with a signal receiving module that can collect wireless communication signal data;
[0071] The recognition device refers to a computing device configured with an ADNN model;
[0072] The ADNN model can be implemented using various deep neural networks, such as convolutional neural networks, recurrent neural networks, and a hybrid of the two.
[0073] For example, when a convolutional neural network is used as the ADNN model, its model structure is as follows: Figure 6 As shown, it includes 6 convolutional layers, activation function layers, batch normalization layers, max pooling layers, and 2 fully connected layers;
[0074] S104: Signal data acquisition and transmission: The sensing device inputs the acquired wireless communication signal into the signal compression deep neural network, and transmits the resulting compressed data to the identification device through a wireless link;
[0075] S105: Signal Data Recovery and Recognition: The recognition device inputs the shrinkage data received through the wireless link into the signal recovery deep neural network to obtain enhanced data, and then inputs it as signal data into the ADNN model to obtain the modulation type of the signal data.
[0076] Furthermore, such as Figure 7 As shown, the specific method for joint training of the deep neural network model includes the following steps:
[0077] S1011: Design of a Deep Neural Network for Signal Compression: Design the model structure of a deep neural network for signal compression;
[0078] The structure of the signal compression deep neural network can be selected from existing typical deep neural network structures, such as autoencoders, convolutional neural networks, recurrent neural networks, etc., or a new network structure can be designed.
[0079] For example, Figure 8 A schematic diagram of a signal compression deep neural network is given, which includes four compression modules. Each compression module consists of a convolutional layer, a batch normalization layer, an activation function layer, and a max pooling layer.
[0080] S1012: Design of a Deep Neural Network for Signal Recovery: Design the model structure of a deep neural network for signal recovery;
[0081] The signal recovery deep neural network can be structured from existing typical deep neural network structures, such as autoencoders, convolutional neural networks, recurrent neural networks, etc., or novel network structures can be designed.
[0082] For example, Figure 9 A schematic diagram of a deep neural network for signal recovery is given, which includes four recovery modules. Each recovery module consists of a convolutional layer, a batch normalization layer, an activation function layer, and an upsampling module.
[0083] The upsampling module refers to the upsampling module in step S102;
[0084] S1013: Joint training of neural network models: Joint training of signal compression deep neural network and signal recovery deep neural network;
[0085] Joint training refers to updating the overall model parameters of two or more network models using a related error loss function;
[0086] For example, Figure 10 A joint training implementation is presented, in which the output layers of a signal compression deep neural network and a signal recovery deep neural network are connected to form a unified network. Training samples are input, and augmented samples are output. Then, the error loss between the signal samples and the augmented data is calculated. The parameters of both the signal compression deep neural network and the signal recovery deep neural network are simultaneously updated through backpropagation of the error loss, thus achieving joint training.
[0087] The training samples refer to a set consisting of several signal samples and their corresponding modulation type labels, which can be collected using public datasets or through wireless communication devices;
[0088] For example, to verify the proposed method for spectrum data compression backhaul and model inference acceleration, this invention conducted experimental verification on the RADIOML2018.10A dataset, which contains 2,555,904 signal samples. The samples are divided into 24 modulation signals, each with a signal-to-noise ratio ranging from -20dB to 30dB, with a step size of 2dB. The number of samples for each modulation signal corresponding to each signal-to-noise ratio is 4,096. A single modulation signal sample is a 1024*2 vector, containing one in-phase component of length 1024 and one quadrature component of length 1024.
[0089] The experimental system environment was Windows 10, the deep learning framework was PyTorch 1.6.0, and the software packages used included NumPy, Matplotlib, and Pandas.
[0090] use Figure 8 The signal compression deep neural network model shown Figure 9 The signal recovery deep neural network model shown is as follows. Figure 6 The ADNN model shown;
[0091] The signal data sample has a dimension of 1024*2, while the shrunken data sample has a dimension of 64*16. Therefore, the amount of data transmitted is reduced by 50%.
[0092] Figure 11 It gives the use of Figure 2 When signal data is used as input data for a signal compression deep neural network, the resulting image is a visualization of the flattened shrunken data.
[0093] Without using the S102 upsampling module structure inference optimization step, the computation time required for the signal recovery deep neural network model to infer 10 times for 500 signal samples is 4.7 seconds. After using the S102 upsampling module structure inference optimization step, the computation time required for the signal recovery deep neural network model to infer 10 times for 500 signal samples is 2.84 seconds, and the inference computation time is reduced by 39.57%.
[0094] When 29,496 samples were randomly selected from the dataset, the average MSE of the enhanced data obtained by the signal recovery deep neural network and the signal data input to the signal compression deep neural network was 0.0206, which achieved a high data recovery accuracy.
[0095] Figure 12 It gives the use of Figure 4 The enhanced data corresponding to the signal data shows that the IQ signals are highly similar, which meets the requirements for enhanced data.
[0096] Figure 13 and Figure 14 Other network structures remain unchanged, only the following are given: Figure 5 The multi-branch structure shown is changed to Figure 15 or Figure 16 The diagram illustrates the error loss of the signal recovery deep neural network and the recognition accuracy of the ADNN model during training when using a single-branch structure based on transposed convolution or nearest neighbor. Here, "multi-branch structure" indicates the use of the multi-branch upsampling scheme designed in this application; "single-branch-transposed convolution" indicates the use of... Figure 15 The upsampling scheme shown; "single branch - nearest neighbor" indicates that the upsampling method is adopted. Figure 16 The upsampling scheme shown is... Figure 13 and Figure 14 It can be seen that the training error loss of the network model using the single-branch upsampling scheme is higher than that of the multi-branch upsampling network model designed in the embodiment of the present invention. The signal recovered by the multi-branch network model has higher recognition accuracy in the ADNN model than that of the single-branch network model, and the training convergence speed is faster.
[0097] The nearest neighbor refers to a data interpolation algorithm that directly inserts data points from low-dimensional data into the nearest neighbor positions in high-dimensional data. It calculates and replaces the data points in high-dimensional data by using the coordinates of the data points in low-dimensional data and the scaling factor. For example, to scale 2×2 low-dimensional data to 4×4 high-dimensional data, the scaling factor is 2. For the data point at position (2,2) in the high-dimensional data, its position in the original data is mapped to (1,1) based on the scaling factor. Therefore, the data at position (1,1) in the original data is assigned to position (2,2) in the high-dimensional data. The same applies to other positions until all data points in the high-dimensional data have been traversed.
Claims
1. A method for compressed transmission and accelerated model inference processing of spectrum data, characterized in that... Includes the following steps: S101: Joint Training of Deep Neural Network Models: Design deep neural network structures for signal compression and signal recovery, and perform joint training based on given signal data; S102: Upsampling module structure inference optimization: The upsampling module with a multi-branch structure is merged into a single-branch structure to increase the inference speed of the signal recovery deep neural network; S103: Deployment of deep neural network models: Deploy the trained signal compression deep neural network to the sensing device, and deploy the signal recovery deep neural network to the recognition device; S104: Signal data acquisition and transmission: The sensing device inputs the acquired wireless communication signal into the signal compression deep neural network, and transmits the resulting compressed data to the identification device through a wireless link; S105: Signal Data Recovery and Recognition: The recognition device inputs the shrinkage data received through the wireless link into the signal recovery deep neural network to obtain enhanced data, and then inputs it as signal data into the ADNN model to obtain the modulation type of the signal data.
2. The method for spectrum data compression and backhaul and model inference acceleration as described in claim 1, characterized in that: The input of a signal compression deep neural network is signal data, and the output is shrunken data. It also includes a data downsampling module. The input of a signal recovery deep neural network is shrunken data, and the output is enhanced data. It also includes an upsampling module with a multi-branch structure.
3. The method for spectrum data compression and backhaul and model inference acceleration as described in claim 1, characterized in that, Step S101 specifically includes the following steps: S: Signal compression deep neural network design: Design the model structure of a signal compression deep neural network; S: Design of a deep neural network for signal recovery: Design the model structure of a deep neural network for signal recovery; S: Joint training of neural network models: Joint training of signal compression deep neural network and signal recovery deep neural network.
4. The method for spectrum data compression and backhaul and model inference acceleration as described in claim 3, characterized in that, In step S1013: Joint training refers to updating the overall model parameters of two or more network models using a related error loss function; Training samples refer to a set of signal samples and their corresponding modulation type labels, which can be collected using public datasets or wireless communication devices.
5. The method for spectrum data compression and backhaul and model inference acceleration as described in claim 4, characterized in that, The joint training includes the following steps: The output layer of the signal compression deep neural network is connected to the input layer of the signal recovery deep neural network to form a whole network, which takes training samples as input and outputs augmented samples. Then, the error loss between the signal sample and the augmented data is calculated, and the parameters of the signal compression deep neural network and the signal recovery deep neural network are updated simultaneously through backpropagation of the error loss, so as to achieve joint training.
6. The method for spectrum data compression and backhaul and model inference acceleration as described in claim 1, characterized in that, In step S102: The upsampling module has a multi-branch structure during the training phase, simultaneously receiving data input and calculating the result data separately, and then concatenating the result data into the final output data; Fusion into a single-branch structure refers to the lossless transformation of each branch in the upsampling module into a single-branch structure during the inference phase through structural reparameterization. The structural reparameterization refers to first constructing a series of structures and then equivalently converting their parameters into another set of parameters, thereby converting this series of structures into another series of structures.
7. A spectrum data compression and backhaul and model inference acceleration processing system, characterized in that... include: The system includes a sensing device, an identification device, and a wireless communication network. The sensing device is used to receive signal data sent by a radiation source and to compress the received signal data and obtain compressed data through a signal compression deep neural network with a downsampling module. Wireless communication network used to transmit shrinkage data to identification devices; The identification device is used to receive contraction data, recover signal data through a signal recovery deep neural network with an upsampling module, and transmit the signal data to an ADNN model for processing. The ADNN model processes the signal data to obtain the modulation type of the signal data.
8. The spectrum data compression and backhaul and model inference acceleration processing system as described in claim 7, characterized in that... The signal compression deep neural network includes four compression modules: a first compression module of 128*512, a second compression module of 64*256, a third compression module of 32*128, and a fourth compression module of 16*64. The first to fourth compression modules are connected in sequence. The first to fourth compression modules include Conv convolutional layers, ReLU activation function layers, BN batch normalization processing layers, and MaxPool max pooling layers.
9. The spectrum data compression and backhaul and model inference acceleration processing system as described in claim 7, characterized in that... The signal recovery deep neural network includes four recovery modules: a first recovery module of 16*128, a second recovery module of 32*256, a third recovery module of 64*512, and a fourth recovery module of 2*1024. The first to fourth recovery modules are connected in sequence. The first to fourth recovery modules include Conv convolutional layers, ReLU activation function layers, BN batch normalization processing layers, and Upsample upsampling layers.
10. The spectrum data compression and backhaul and model inference acceleration processing system as described in claim 7, characterized in that... The ADNN model includes: The system comprises a first processing module of 64*512, a second processing module of 64*256, a third processing module of 64*128, a fourth processing module of 64*64, a fifth processing module of 64*32, a sixth processing module of 64*2, a seventh processing module of 1*64, and an eighth processing module of 1*24. The first to fifth processing modules include Conv convolutional layers, ReLU function layers, BN batch normalization layers, and MaxPool max pooling layers. The sixth processing module includes Conv convolutional layers, ReLU function layers, BN batch normalization layers, and AvgPool average pooling layers. The seventh and eighth processing modules are linear layers.