A method for identifying multiple dynamic impact signals by using a lightweight neural network

Abstract

This invention discloses a lightweight neural network identification method for high-speed dynamic impact signals. In the process of penetrating multi-layered hard targets, the presence of signal adhesion in the overload signals from multiple dynamic impacts makes identification exceptionally difficult. To address this problem, a lightweight network identification method based on an attention mechanism is proposed. First, time-frequency feature analysis is performed on the overload signal equivalent to that from multiple impact test benches, and continuous wavelet transform is used to extract the time-frequency features as input to the neural network. A lightweight network architecture based on an attention mechanism is designed, eliminating redundant layers and adding residual connection structures and a lightweight attention mechanism, thus ensuring recognition accuracy while significantly reducing parameters.

Description

Technical Field

[0001] This invention relates to the fields of dynamics and machine learning, and specifically to a method for identifying multiple dynamic impact signals using a lightweight neural network. Background Technology

[0002] Multiple impact loads are widely present in various industrial scenarios, such as aircraft wings, high-speed train bogies, and large mechanical transmission systems. The signal features extracted from multiple impact loads are helpful for achieving important functions such as load mode identification, fault diagnosis, and system prediction. However, the signals from multiple impact loads themselves have strong noise and random interference, making feature extraction of related load signals difficult.

[0003] In recent years, deep learning has shown great potential in the identification of multiple impact signals, opening a new frontier for signal-based methods for systems with multiple dynamic effects. These deep learning-based methods do not rely on particularly sophisticated filter design, making them more effective in prediction and identification in many practical engineering applications. Deep learning methods have achieved great success in a wide range of applications involving multiple dynamic effects.

[0004] Load identification of penetrating munitions is one of the most challenging aspects of military applications, across various impact load scenarios. Penetrating munitions need to penetrate multiple solid protective structures and detonate precisely at a designated layer to effectively damage a specific target. This designated detonation layer relies on layer identification of multiple impact load signals from a projectile acceleration sensor. However, in addition to strong noise and random interference similar to many other applications, penetrating munitions exhibit a unique "signal adhesion" problem in the time domain, further complicating load identification. Due to the extremely high impact velocity and complex internal structure of the munition, "signal adhesion" is essentially a mixture of the true penetration overload signal and the coupled oscillating signal. Besides "signal adhesion," the identification of multiple impact signals from penetrating munitions is also limited by hardware and operational conditions. Therefore, the application of deep learning methods in this field requires considerable additional effort. First, deep learning methods typically require extensive data recording, which is extremely difficult given the long and costly experimental cycles of penetrating munitions. Second, traditional large-scale neural networks are difficult to deploy directly due to hardware resource constraints. Therefore, lightweight pruning of neural network models and acceleration of operation under limited resources are crucial issues that must be addressed. Furthermore, since high-speed dynamic multiple impact signals are non-stationary random signals and are accompanied by severe signal adhesion, it is difficult to directly extract signal features. Summary of the Invention

[0005] Based on the above analysis and discussion, this invention discloses a method for identifying multiple dynamic impact signals using a lightweight neural network. First, a database of multiple dynamic impact signals is established and classified according to data characteristics. Second, features are extracted from different dynamic impact signals using continuous wavelet transform, and then a lightweight neural network based on an attention mechanism is designed based on the feature information. Finally, the neural network is trained and validated using the data features to ensure the effectiveness of the designed method.

[0006] A method for obtaining multi-missile cooperative countermeasure strategies based on a combination of virtual force method and war-inducing cooperation includes the following steps:

[0007] Step 1: Conduct dynamic impact tests on the target under different working conditions and collect dynamic impact signals under different working conditions. Establish a database of multiple dynamic impact signals of the target. Classify the dynamic impact signals in the database according to the impact characteristics to obtain the classified dynamic impact signals, and then proceed to Step 2.

[0008] Step 2: Use the continuous wavelet transform method to extract features from the classified dynamic impact signal, obtain the corresponding feature data, and use the feature data to form a training set and a test set, then proceed to step 3.

[0009] Step 3: Construct a lightweight neural network architecture, then proceed to Step 4.

[0010] Step 4: Train the lightweight neural network architecture using the training set to obtain the lightweight neural network, then proceed to step 5.

[0011] Step 5: Input the test set into the lightweight neural network to verify the accuracy of the lightweight neural network.

[0012] Compared with the prior art, the advantages of this invention are:

[0013] (1) The signal features of the high-speed dynamic impact system were extracted using continuous wavelet transformation. This method can adaptively adjust according to the frequency characteristics of non-stationary signals and effectively extract features from the initial signal. By increasing the dimension of the features, the high-speed dynamic impact signal can be represented in both the time and frequency domains, providing good basic data for network training.

[0014] (2) An ultra-lightweight neural network was designed. To address the problem of the limited number of target features to be identified in the scenario of penetrating munitions, the network structure was further pruned. By using a jointly designed residual structure, the gradient vanishing problem of the pruned network was solved, thus accelerating the training speed of the network.

[0015] (3) A lightweight attention module was designed to address the characteristics of multiple impact processes. Based on this, a feature-based attention module was established by combining channel attention and spatial attention to further improve recognition accuracy. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of a multiple-impact signal training method.

[0017] Figure 2 These are schematic diagrams of measured overload impact signals and wavelet analysis under typical penetration conditions. Figure (a) shows the time domain diagram of a low-speed single-impact overload signal, Figure (b) shows the time domain diagram of a low-speed six-impact signal, Figure (c) shows the time domain diagram of a high-speed five-impact signal, Figure (d) shows the time domain diagram of a high-speed six-impact signal, Figure (e) shows the time-frequency domain diagram of a continuous wavelet transform of a low-speed single-impact overload signal, Figure (f) shows the time-frequency domain diagram of a low-speed six-impact overload signal, Figure (g) shows the time-frequency domain diagram of a high-speed five-impact overload signal, and Figure (h) shows the time-frequency domain diagram of a high-speed six-impact overload signal.

[0018] Figure 3 This is a diagram of a lightweight network structure based on the attention mechanism.

[0019] Figure 4 This is a diagram showing the layer identification results under low-speed conditions.

[0020] Figure 5 This is a diagram showing the layer identification results under medium-speed conditions.

[0021] Figure 6 This is a diagram showing the results of layer number recognition under high-speed conditions. Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0023] Combination Figures 1-4 The method for identifying multiple dynamic impact signals using a lightweight neural network as described in this invention includes the following steps:

[0024] Step 1: Establish a database of multiple dynamic impact signals and classify them according to impact characteristics, as follows:

[0025] Step 1.1: Classify the initial data according to the different impact velocities.

[0026] Due to the different types of penetrating munitions, their impact velocities vary significantly. Based on actual usage scenarios of penetrating munitions, laboratory equivalent simulation dynamic data were collected, yielding experimental data at different impact velocities. According to the velocity, the laboratory multiple impact equivalent data were divided into low-velocity, medium-velocity, and high-velocity datasets. The low-velocity dataset had a velocity of 200-350 m / s, with 178 data samples collected; the medium-velocity dataset had a velocity of 450-600 m / s, with 326 data samples collected; and the high-velocity dataset had a velocity of 800-1000 m / s, with 447 data samples collected.

[0027] Step 1.2: Classify the initial data according to the different impact velocities.

[0028] Due to the different types of attack targets, the number of target layers varies greatly. Considering the specific forms of most attack targets on the battlefield, and based on the different number of target layers, the samples in different speed datasets are further divided into layers 1-6. Specifically: Low-speed dataset: 30 samples in layer 1, 30 samples in layer 2, 30 samples in layer 3, 30 samples in layer 4, 29 samples in layer 5, and 29 samples in layer 6; Medium-speed dataset: 55 samples in layer 1, 55 samples in layer 2, 54 samples in layer 3, 54 samples in layer 4, 54 samples in layer 5, and 54 samples in layer 6; High-speed dataset: 74 samples in layer 1, 75 samples in layer 2, 74 samples in layer 3, 74 samples in layer 4, 75 samples in layer 5, and 74 samples in layer 6.

[0029] Step 2: Use the continuous wavelet transform method to extract features from the classified dynamic impact signal, and obtain the corresponding feature data as follows:

[0030] The presence of different dynamic systems causes the projectile to generate strong oscillation signals during repeated target penetration. During this process, the strong nonlinearity and gap characteristics between the projectile and sensors cause these oscillation signals to constantly change. The main sources of these signals are the high-frequency oscillations transmitted within the projectile's casing, fuze chamber, and fuze as stress waves propagate through the penetration process, and the high-frequency output of the acceleration sensor during penetration. While the oscillation signals gradually decay over time, warheads with large aspect ratios cause the oscillation signals from adjacent layers to overlap with the actual impact penetration signals. This results in the rising and falling edges of the overload signals from multiple target layers becoming abnormally contiguous and difficult to directly identify, even submerging the effective signal envelope in the high-frequency oscillations. This leads to significant errors in the fuze's layer identification.

[0031] For the dataset in step 1, continuous wavelet transform can be used for time-frequency domain analysis. Wavelet transform can adaptively meet the requirements of time-frequency signals, focus on arbitrary details of the signal, subdivide time at high frequencies and subdivide frequency at low frequencies, which is beneficial for processing multiple impulse signals.

[0032] Measured overload impact signals under several typical penetration conditions, such as Figure 2 As shown, penetration velocity is a major factor contributing to the difficulty in identifying impact signal characteristics. Figure 2 As shown in (a) and (b), at lower velocities, layered structural features are easily identifiable from the signal in both single-layer and multi-layer impact scenarios. In contrast, at higher impact velocities, such as Figure 2 As shown in (c) and (d), the signal adhesion in the time domain is very severe, making it impossible to distinguish between operating conditions with different impact frequencies. Based on the results of continuous wavelet transform time-frequency domain analysis, the underlying causes of this signal adhesion problem and possible solutions are discussed. Figure 2 As shown in (e) and (f), the time-frequency domain results of low-velocity impact signals exhibit significant characteristics corresponding to the number of penetration layers. In contrast, under high-speed conditions, as... Figure 2 As shown in (g) and (h), the time-frequency domain results of the signal are only weakly correlated with the number of penetration layers in the mid-frequency range, and are completely indistinguishable in the high-frequency and low-frequency ranges. Under high-speed impact, the disturbances generated inside the shell by mechanical transmission and oscillation are high-frequency, while the noise generated by friction between the shell and the target is low-frequency. Therefore, mid-frequency information is the most valuable and deserves special attention from neural networks. Thus, continuous wavelet transform can increase the data dimension and further extract the time-frequency characteristics during the impact process. Then, 80% of the above feature data and 20% of the data are used to form the training set.

[0033] Step 3: Construct a lightweight neural network architecture, which includes a max-refire module and a lightweight attention mechanism module, as detailed below:

[0034] Step 3.1: Design the max-refire module in the lightweight network:

[0035] The max-refire module consists of a max pooling module, a fire module, and residual connections. Let the input feature be x, then the output y of the max pooling layer in max-refire is expressed as:

[0036] (1)

[0037] in This is a max pooling operation with a pooling kernel size of k and a pooling step size of s.

[0038] For the compression layer, the output y-feature map of its max-pooling layer is compressed in terms of channel count using 1x1 convolutions. This reduces the number of parameters and improves computational efficiency. The convolution kernel weights used in the compression layer are... The bias term is The output feature map of the compression stage Represented as:

[0039] (2)

[0040] in This represents a 1×1 convolution kernel. This indicates the sigmod activation function.

[0041] For the expansion layer, the expansion layer uses 1x1 convolutions and 3x3 convolutions to combine the output features of the compression layer. The expansion layer increases non-linear expressive power; it consists of a 1x1 convolutional kernel and a 3x3 convolutional kernel, where the kernel weights of the first 1x1 convolutional layer are... The bias term is The kernel weights of the 3x3 convolutional layer are The bias term is The output feature map of the expand layer. Represented as:

[0042] (3)

[0043] in This indicates that the two feature maps are connected along the channel dimension; in formula (3), they are connected along the channel dimension through a 1x1 convolutional layer and a 3x3 convolutional layer to increase the non-linear expressive power of the network while maintaining its lightweight nature; on this basis, a residual structure is added, the mathematical expression of which is as follows:

[0044] (4)

[0045] in For residual structure output, This is the residual learning function; by utilizing this residual structure feature and combining it with the characteristics of the SqueezeNet network structure, the input is superimposed after the initial processing by the Fire module. Compared with the original residual network, this greatly reduces the number of parameters, training time, and network size.

[0046] Step 3.2: Design a lightweight attention mechanism module

[0047] Due to the diversity of impact features, especially the critical importance of feature selection after passing through network layers, an attention mechanism can be introduced to focus on the overload feature information during the penetration phase. Combining channel and spatial attention mechanisms, a lightweight attention mechanism module CBAM using convolutional modules is derived. The channel attention mechanism module CAM extracts spatial information of features through max pooling and average pooling layers; its channel attention features... The calculation method is as follows:

[0048] (5)

[0049] Where F represents the input of CAM, and MLP represents a fully connected layer. This represents an average pooling operation with a kernel size of k and a pooling step size of s. The weights of the average pooling layer in CAM. This is the bias of the average pooling layer in CAM. These are the weights of the max pooling layer in CAM. The bias of the max pooling layer in CAM is then used as a supplement, and the lightweight spatial attention module SAM is used to perform spatial attention features. The calculation method is as follows:

[0050] (6)

[0051] Where F ’ This represents the input to SAM. This represents the continuous operation of three 3×3 convolution kernels; This represents the weights of the 3×3 convolution kernel in SAM; This indicates the bias of the 3×3 convolution kernel in SAM;

[0052] Therefore, CAM can identify specific channels in the impact characteristics and provide enhanced layer number characteristics for SAM. This kind of serial expression can effectively increase the identification results of the number of impacts.

[0053] By combining the network structures designed in steps 3.1 and 3.2, and considering the characteristics of the residual structure and the network module structure based on the attention feature, a lightweight neural network model based on the attention mechanism can be obtained.

[0054] In step 4, the lightweight neural network architecture is trained using the training set to obtain the lightweight neural network, as detailed below:

[0055] Step 4.1: Update the parameter gradients of the compressed and expanded layers in max-refire:

[0056] Since the max pooling layer has no parameter changes, the gradient does not need to be updated. Using the output of the max pooling layer as the input to the compressed layer, the parameter update formula for the compressed layer is:

[0057] (7)

[0058] Wherein, the loss function is The learning rate is , This represents the weights of the updated 1×1 convolution kernel. This represents the bias of the updated 1×1 convolution kernel. This represents the weights of the current 1×1 convolution kernel. This indicates the bias of the current 1×1 convolution kernel. This indicates that the loss function is applied to the parameters. gradient, This indicates that the loss function is applied to the parameters. gradient, and All parameters are calculated using the backpropagation algorithm; then, the parameters of the compressed layer are used as input to the extended layer, and the parameter update formula for the extended layer is:

[0059] (8)

[0060] in, This represents the weights of the updated 1×1 convolution kernel. This represents the bias of the updated 1×1 convolution kernel. This represents the weights of the updated 3×3 convolution kernel. This indicates the bias of the updated 3×3 convolution kernel; This represents the weights of the current 1×1 convolution kernel. These represent the biases of the current 1×1 convolution kernel. This represents the weights of the current 3×3 convolution kernel. This indicates the bias of the current 3×3 convolution kernel; This indicates that the loss function is applied to the parameters. gradient, This indicates that the loss function is applied to the parameters. gradient, This indicates that the loss function is applied to the parameters. gradient, This indicates that the loss function is applied to the parameters. gradient, He Jun It was calculated using the backpropagation algorithm.

[0061] Step 4.2: Update the parameter gradients of the lightweight attention mechanism module:

[0062] Since max pooling and average pooling do not involve updating parameter gradients, and since channel attention is formed by concatenating average pooling and max pooling layers and then connecting them to the fully connected layer, the weights and biases of the two parts are updated according to the following formula:

[0063] (9)

[0064] in, This represents the updated weights of the fully connected layer after the average pooling layer. This represents the updated bias of the fully connected layer after the average pooling layer. This represents the updated weights of the fully connected layer after the max pooling layer. This represents the updated bias of the fully connected layer after the max pooling layer; This represents the current weights of the fully connected layer after passing through the average pooling layer. This indicates the current bias of the fully connected layer after passing through the average pooling layer. This represents the current weights of the fully connected layer after the max pooling layer. This indicates the current bias of the fully connected layer after passing through the max pooling layer; This indicates that the loss function is applied to the parameters. gradient, This indicates that the loss function is applied to the parameters. gradient, This indicates that the loss function is applied to the parameters. gradient, This indicates that the loss function is applied to the parameters. gradient, and All results were calculated using the backpropagation algorithm.

[0065] For spatial attention, it consists of concatenated average pooling and max pooling layers, followed by three 3×3 convolutional layers. Therefore, the weights and biases are updated after average pooling and max pooling as follows:

[0066] (10)

[0067] in, This represents the weights of the updated 3×3 convolution kernel. This indicates the bias of the updated 3×3 convolution kernel. This represents the weights of the current 3×3 convolution kernel. This indicates the bias of the current 3×3 convolution kernel. This indicates that the loss function is applied to the parameters. gradient, This indicates that the loss function is applied to the parameters. gradient, and It was calculated using the backpropagation algorithm.

[0068] Step 5: Input the test set into the lightweight neural network obtained in Step 4 to verify the network's accuracy, as follows:

[0069] The test set data obtained in step 2 is input into the lightweight neural network obtained in step 4 for testing. To highlight the advantages of the network proposed in this invention, comparison networks are used, including SqueezeNet, GoogleNet, ResNet, and the lightweight neural network proposed in this paper.

[0070] From both a theoretical perspective of signal processing and a practical perspective of penetrating munition functionality, the accuracy of impact type identification is the most important performance indicator. The layer identification results of the proposed method compared with three other methods at different velocities are as follows: Figure 4-6 As shown. At lower speeds, signal adhesion caused by nonlinear interference is minimized, and all four methods achieve 100% layer recognition accuracy, such as... Figure 4 As shown. However, the layer recognition performance of the comparative method deteriorates significantly with increasing penetration rate and intensified nonlinear interference. In stark contrast, the proposed method maintains ideal layer recognition performance under these harsh conditions. Figure 5 The results show the layer number identification results of different methods at a medium speed. The overall recognition rate of the ResNet architecture is 88.3%, while that of the traditional SqueezeNet architecture is 99.2%. It is worth noting that the layer identification accuracy is lower for tests affecting specific multi-layer structures. The recognition accuracy of the ResNet and traditional SqueezeNet architectures drops to 78.7% and 75%, respectively, under their worst-case conditions. Under these medium-speed conditions, both the proposed method and GoogleNet maintain 100% recognition accuracy, but the GoogleNet architecture is clearly much larger. Figure 6 The results of different methods for layer number identification under high-speed conditions are shown. The overall identification rate of the ResNet architecture is 80%, and that of the traditional SqueezeNet architecture is 82.2%. These two architectures can no longer meet the functional requirements of practical penetrating munitions. Under these extreme conditions, the proposed lightweight network architecture can maintain the same overall identification accuracy of 95.5% as GoogleNet, which has a very large network scale.

Claims

1. A method for identifying multiple dynamic impact signals using a lightweight neural network, characterized in that, Includes the following steps: Step 1: Conduct dynamic impact tests on the target under different working conditions, collect dynamic impact signals under different working conditions, establish a database of multiple dynamic impact signals of the target, classify the dynamic impact signals in the database according to the impact characteristics, and obtain the classified dynamic impact signals. Based on the actual use scenarios of penetrating munitions, laboratory equivalent simulation dynamic data work was carried out to obtain experimental data under different impact velocities. According to the different velocities, the experimental data under different impact velocities were divided into low-velocity datasets, medium-velocity datasets and high-velocity datasets. Based on the specific forms of most attack targets on the battlefield, according to the different number of target layers, the samples under different velocity datasets were further divided into 1-6 layers, thus obtaining the classified dynamic impact signals. Proceed to step 2; Step 2: Use the continuous wavelet transform method to extract features from the classified dynamic impact signal, obtain the corresponding feature data, and use the feature data to form a training set and a test set, then proceed to step 3. Step 3: Construct a lightweight neural network architecture, which includes a max-refire module and a lightweight attention mechanism module, as detailed below: Step 3.1: Design the max-refire module in the lightweight network: The max-refire module consists of a max pooling module, a fire module, and residual connections. Let the input feature be x, then the output y of the max pooling layer in max-refire is expressed as: (1) in This is a max pooling operation with a pooling kernel size of k and a pooling step size of s. For the compression layer, the output y-feature map of its max-pooling layer is compressed in terms of channel number using a 1x1 convolution. The convolution kernel weights used in the compression layer are... The bias term is The output feature map of the compressed layer Represented as: (2) in This represents a 1×1 convolution kernel. This represents the sigmoid activation function; For the expansion layer, the expansion layer uses 1x1 convolutions and 3x3 convolutions to combine the output features of the compression layer. The expansion layer increases non-linear expressive power; it consists of a 1x1 convolutional kernel and a 3x3 convolutional kernel, where the kernel weights of the first 1x1 convolutional layer are... The bias term is The kernel weights of the 3x3 convolutional layer are The bias term is The output feature map of the expand layer. Represented as: (3) in This indicates that the two feature maps are connected along the channel dimension; in formula (3), they are connected along the channel dimension through a 1x1 convolutional layer and a 3x3 convolutional layer to increase the non-linear expressive power of the network while maintaining its lightweight nature; on this basis, a residual structure is added, the mathematical expression of which is as follows: (4) in For residual structure output, For residual learning function; Step 3.2: Design a lightweight attention mechanism module An attention mechanism is introduced to focus on the overload feature information during the penetration phase. By combining channel and spatial attention mechanisms, a lightweight attention mechanism module CBAM using convolutional modules is derived. The Channel Attention Mechanism (CAM) module extracts spatial information of features through max pooling and average pooling layers. Its channel attention features The calculation method is as follows: (5) Where F represents the input of CAM, and MLP represents a fully connected layer. This represents an average pooling operation with a kernel size of k and a pooling step size of s. The weights of the average pooling layer in CAM. This is the bias of the average pooling layer in CAM. These are the weights of the max pooling layer in CAM. The bias of the max pooling layer in CAM is then used as a supplement, and the lightweight spatial attention module SAM is used to perform spatial attention features. The calculation method is as follows: (6) Where F ’ This represents the input to SAM. This represents the continuous operation of three 3×3 convolution kernels; This represents the weights of the 3×3 convolution kernel in SAM; This indicates the bias of the 3×3 convolution kernel in SAM; Therefore, CAM can identify specific channels in the impact characteristics and provide enhanced layer number features for SAM. This kind of serial expression can effectively increase the identification results of the number of impacts. Proceed to step 4; Step 4: Train the lightweight neural network architecture using the training set to obtain the lightweight neural network, as detailed below: Step 4.1: Update the parameter gradients of the compressed and expanded layers in max-refire: Since the max pooling layer has no parameter changes, the gradient does not need to be updated. Using the output of the max pooling layer as the input to the compressed layer, the parameter update formula for the compressed layer is: (7) Wherein, the loss function is The learning rate is , This represents the weights of the updated 1×1 convolution kernel. This represents the bias of the updated 1×1 convolution kernel. This represents the weights of the current 1×1 convolution kernel. This indicates the bias of the current 1×1 convolution kernel. This indicates that the loss function is applied to the parameters. gradient, This indicates that the loss function is applied to the parameters. gradient, and All parameters are calculated using the backpropagation algorithm; then, the parameters of the compressed layer are used as input to the extended layer, and the parameter update formula for the extended layer is: (8) in, This represents the weights of the updated 1×1 convolution kernel. This represents the bias of the updated 1×1 convolution kernel. This represents the weights of the updated 3×3 convolution kernel. This indicates the bias of the updated 3×3 convolution kernel; This represents the weights of the current 1×1 convolution kernel. These represent the biases of the current 1×1 convolution kernel. This represents the weights of the current 3×3 convolution kernel. This indicates the bias of the current 3×3 convolution kernel; This indicates that the loss function is applied to the parameters. gradient, This indicates that the loss function is applied to the parameters. gradient, This indicates that the loss function is applied to the parameters. gradient, This indicates that the loss function is applied to the parameters. gradient, He Jun It was calculated using the backpropagation algorithm; Step 4.2: Update the parameter gradients of the lightweight attention mechanism module: Since max pooling and average pooling do not involve updating parameter gradients, and since channel attention is formed by concatenating average pooling and max pooling layers and then connecting them to the fully connected layer, the weights and biases of the two parts are updated according to the following formula: (9) in, This represents the updated weights of the fully connected layer after the average pooling layer. This represents the updated bias of the fully connected layer after the average pooling layer. This represents the updated weights of the fully connected layer after the max pooling layer. This represents the updated bias of the fully connected layer after the max pooling layer; This represents the current weights of the fully connected layer after passing through the average pooling layer. This indicates the current bias of the fully connected layer after passing through the average pooling layer. This represents the current weights of the fully connected layer after the max pooling layer. This indicates the current bias of the fully connected layer after passing through the max pooling layer; This indicates that the loss function is applied to the parameters. gradient, This indicates that the loss function is applied to the parameters. gradient, This indicates that the loss function is applied to the parameters. gradient, This indicates that the loss function is applied to the parameters. gradient, and All were calculated using the backpropagation algorithm; For spatial attention, it consists of concatenated average pooling and max pooling layers, followed by three 3×3 convolutional layers. Therefore, the weights and biases are updated after average pooling and max pooling as follows: (10) in, This represents the weights of the updated 3×3 convolution kernel. This indicates the bias of the updated 3×3 convolution kernel. This represents the weights of the current 3×3 convolution kernel. This indicates the bias of the current 3×3 convolution kernel. This indicates that the loss function is applied to the parameters. gradient, This indicates that the loss function is applied to the parameters. gradient, and Calculated using the backpropagation algorithm; Proceed to step 5; Step 5: Input the test set into the lightweight neural network to verify the accuracy of the lightweight neural network.

2. The method for identifying multiple dynamic impact signals using a lightweight neural network according to claim 1, characterized in that, In step 2, the continuous wavelet transform method is used to extract features from the classified dynamic impact signal, and the corresponding feature data is obtained as follows: The continuous wavelet transform method is used to perform time-frequency domain analysis on the classified dynamic impact signals in the dataset, and the results of the wavelet transform time-frequency domain analysis, i.e., feature data, are obtained. Use 80% of the feature data as the training set and the remaining 20% ​​of the data as the training set. Continuous wavelet transform can adaptively meet the requirements of time-frequency signals, focusing on arbitrary details of the time-frequency signals, subdividing time at high frequencies and frequency at low frequencies, which is beneficial for processing multiple impact signals. According to the results of wavelet transform time-frequency domain analysis, the time-frequency domain results of low-speed impact signals have significant characteristics corresponding to the number of penetration layers. Under high-speed impact, the disturbances generated by mechanical transmission and oscillation inside the shell are high-frequency, while the noise generated by friction between the shell and the target is low-frequency. In addition, the mid-frequency information is the most valuable and deserves special attention from neural networks.

3. The method for identifying multiple dynamic impact signals using a lightweight neural network according to claim 1, characterized in that, Step 5 involves inputting the test set into the lightweight neural network to verify the network's accuracy, as detailed below: The accuracy evaluation chart is obtained by using the ratio of the number of recognized layers through the network to the actual number of layers as the accuracy calculation method.

Images