CN-116361654-B - Method for identifying multiple dynamic impact signals by using lightweight neural network

CN116361654BCN 116361654 BCN116361654 BCN 116361654BCN-116361654-B

Abstract

The invention discloses a lightweight neural network identification method for high-speed dynamic impact signals. The situation that the overload signal with multiple impact dynamic impact is in signal adhesion during the process of penetrating the multilayer hard target makes identification extremely difficult. Based on the problems, a lightweight network identification method based on an attention mechanism is provided. Firstly, carrying out time-frequency characteristic analysis on overload signals which are simulated by the equivalent of a multi-impact experiment table, and extracting time-frequency characteristics by utilizing continuous wavelet transformation as the input of a neural network. A light-weight network framework based on an attention mechanism is designed, redundant layers are deleted, a residual error connection structure is increased, and the light-weight attention mechanism is designed, so that the framework ensures the identification precision under the condition of greatly reducing parameters.

Inventors

DAI KEREN
MA XIANG
SHI HUIFA
WEI DA
ZHANG HE

Assignees

南京理工大学

Dates

Publication Date: 20260505
Application Date: 20230403

Claims (3)

1. The method for identifying the multiple dynamic impact signals by using the lightweight neural network is characterized by comprising the following steps of: Step 1, performing dynamic impact tests on targets under different working conditions, collecting dynamic impact signals under different working conditions, establishing a database of multiple dynamic impact signals of the targets, and classifying the dynamic impact signals in the database according to impact characteristics to obtain classified dynamic impact signals; The use scene of the actual penetration ammunition is combined, the equivalent simulation dynamic data work of a laboratory is developed, and experimental data under different impact speeds are obtained; dividing experimental data under different impact speeds into a low-speed data set, a medium-speed data set and a high-speed data set according to different speeds, combining the specific forms of most attack targets of a battlefield, and dividing samples under different speed data sets into 1-6 layers according to different target layer numbers to obtain classified dynamic impact signals; Turning to step 2; Step 2, extracting the characteristics of the classified dynamic impact signals by using a continuous wavelet change method, correspondingly obtaining characteristic data, forming a training set and a testing set by the characteristic data, and turning to step 3; Step3, constructing a lightweight neural network framework, wherein the lightweight neural network framework comprises a max-refire module and a lightweight attention mechanism module, and the lightweight neural network framework comprises the following concrete steps: step 3.1, designing a max-refire module in a lightweight network: the max-refire module is composed of a max-pooling module, a Fire module and residual connection, and if the input characteristic is x, the max-refire max-pooling layer output y is expressed as: (1) Wherein the method comprises the steps of For maximum pooling operation, the pooling core size is k, and the pooling step length is s; for the compression layer, the output y characteristic diagram of the maximum pooling layer is subjected to channel number compression through 1x1 convolution, and the convolution kernel weight used by the compression layer is as follows The bias term is Output feature map of compression layer Expressed as: (2) Wherein the method comprises the steps of A convolution kernel of 1 x 1 is represented, Representation sigmod activates a function; For the extension layer, i.e. the extension layer, the extension layer compresses the output characteristics of the layer by 1x1 convolution and 3x3 convolution The expansion layer comprises a 1x1 convolution kernel and a 3x3 convolution kernel, wherein the convolution kernel weight of the first 1x1 convolution layer is The bias term is The convolution kernel weight of the 3x3 convolution layer is The bias term is Output feature map of expand layer Expressed as: (3) Wherein the method comprises the steps of The method is characterized in that two characteristic graphs are connected in a channel dimension, in the formula (3), a 1x1 convolution layer and a 3x3 convolution layer are connected in the channel dimension to increase the nonlinear expression capacity of a network, and meanwhile, the light weight of the network is also maintained, and a residual structure is added on the basis, wherein the mathematical expression is as follows: (4) Wherein the method comprises the steps of For the output of the residual structure, Is a residual learning function; Step 3.2, designing a light attention mechanism module The attention mechanism is introduced to focus on overload characteristic information in the penetration target penetrating stage, and a lightweight attention mechanism module CBAM using a convolution module is obtained by combining a channel and a spatial attention mechanism, wherein the channel attention mechanism module CAM extracts spatial information of characteristics through a maximum pooling layer and an average pooling layer, and the channel attention characteristics of the information The calculation method is as follows: (5) Where F represents the input of the CAM, MLP represents the fully connected layer, Represents average pooling operation, the pooling core size is k, the pooling step length is s, For the weights of the average pooling layer in the CAM, To average the bias of the pooling layer in the CAM, For the weights of the largest pooling layer in the CAM, Bias for maximum pooling layer in CAM, and then light-weight spatial attention module SAM is used as supplement, spatial attention characteristic The calculation mode of (2) is as follows: (6) Where F ' denotes the input of the SAM, Representing successive operations of 3 x 3 convolution kernels; Weights representing 3 x 3 convolution kernels in SAM; representing the offset of the 3 x 3 convolution kernel in the SAM; Therefore, the CAM can identify the specific channel in the impact characteristics and provide the SAM with enhanced layer number characteristics, and the serial expression can effectively increase the identification result of the impact times; Turning to step 4; and 4, training the lightweight neural network framework by using a training set to obtain the lightweight neural network, wherein the lightweight neural network is specifically as follows: And 4.1, carrying out parameter gradient update on a compression layer and an expansion layer in max-refire: The gradient does not need to be updated because the maximum pooling layer has no parameter change, and the output obtained by the maximum pooling layer is used as the input of the compression layer, so that the parameter updating formula of the compression layer is as follows: (7) Wherein the loss function is The learning rate is , Representing the weights of the updated 1 x 1 convolution kernel, Representing the offset of the updated 1 x 1 convolution kernel, Representing the weight of the current 1 x 1 convolution kernel, Representing the offset of the current 1 x 1 convolution kernel, Representing the parameters of the loss function pair Is used for the gradient of (a), Representing the parameters of the loss function pair Is used for the gradient of (a), And And then taking the parameters of the compression layer as the input of the expansion layer, and then updating the parameters of the expansion layer by the following formula: (8) Wherein, the Representing the weights of the updated 1 x 1 convolution kernel, Representing the offset of the updated 1 x 1 convolution kernel, Representing the weights of the updated 3 x 3 convolution kernel, Representing the offset of the updated 3 x 3 convolution kernel; representing the weights of the current 1x 1 convolution kernel, Respectively represent the offsets of the current 1 x 1 convolution kernels, Representing the weight of the current 3x3 convolution kernel, Representing the offset of the current 3x3 convolution kernel; Representing the parameters of the loss function pair Is used for the gradient of (a), Representing the parameters of the loss function pair Is used for the gradient of (a), Representing the parameters of the loss function pair Is used for the gradient of (a), Representing the parameters of the loss function pair Is used for the gradient of (a), And all are equal to Calculated by using a back propagation algorithm; And 4.2, updating the parameter gradient of the lightweight attention mechanism module: since the update of the parameter gradient is not involved in the maximum pooling and the average pooling, and since the channel attention is connected to the fully connected layer after the average pooling and the maximum pooling layer are spliced, the weights and the bias of the two parts are updated according to the following formula respectively: (9) Wherein, the Representing updated weights of the full connection layer after pooling by averaging, Representing the updated bias of the fully connected layer after pooling by averaging, Representing updated weights of the full connection layer after passing the max pooling layer, Representing updated offsets of the full connection layer after passing through the max pooling layer; representing the current weight of the fully connected layer after pooling by average, Representing the current bias of the fully connected layer after pooling by averaging, Representing the current weight of the fully connected layer after passing through the max-pooling layer, Representing the current bias of the fully connected layer after passing through the maximum pooling layer; Representing the parameters of the loss function pair Is used for the gradient of (a), Representing the parameters of the loss function pair Is used for the gradient of (a), Representing the parameters of the loss function pair Is used for the gradient of (a), Representing the parameters of the loss function pair Is used for the gradient of (a), And All are calculated by a back propagation algorithm; For spatial attention, it is spliced by the average pooling and maximum pooling layers in series and then enters into three 3×3 convolution layers, so the weights and offsets after the average pooling and maximum pooling are updated as follows: (10) Wherein, the Representing the weights of the updated 3 x 3 convolution kernel, Representing the offset of the updated 3 x 3 convolution kernel, Representing the weight of the current 3x3 convolution kernel, Representing the offset of the current 3x3 convolution kernel, Representing the parameters of the loss function pair Is used for the gradient of (a), Representing the parameters of the loss function pair Is used for the gradient of (a), And The method comprises the steps of calculating through a back propagation algorithm; Turning to step 5; and 5, inputting the test set into a lightweight neural network, and verifying the accuracy of the lightweight neural network.
2. The method for identifying multiple dynamic impact signals by using a lightweight neural network according to claim 1, wherein in step 2, the feature extraction is performed on the classified dynamic impact signals by using a continuous wavelet change method, so as to obtain feature data correspondingly, which is specifically as follows: Performing time-frequency domain analysis on the classified dynamic impact signals in the data set by using a continuous wavelet change method to obtain a wavelet transformation time-frequency domain analysis result, namely characteristic data, wherein 80% of data in the characteristic data are used as a training set, and the rest 20% of data are used as the training set; The continuous wavelet transformation can meet the requirement of time-frequency signals in a self-adaptive mode, any details of the time-frequency signals are focused, the time is subdivided at a high frequency, the frequency is subdivided at a low frequency, the processing of multiple impact signals is facilitated, according to the result of wavelet transformation time-frequency domain analysis, the time-frequency domain result of the low-speed impact signals has obvious characteristics corresponding to the penetration layer number, under high-speed impact, disturbance generated by mechanical transmission and oscillation in a shell is high-frequency, noise generated by friction between the shell and a target is low-frequency, and in addition, medium-frequency information is the most valuable, and the neural network is particularly interesting.
3. The method for identifying multiple dynamic impact signals by using a lightweight neural network according to claim 1, wherein the inputting of the test set into the lightweight neural network in step 5 verifies the accuracy of the network, specifically as follows: and obtaining an accuracy evaluation chart according to a calculation method taking the ratio of the number of identification layers to the actual number of layers through the network as accuracy.

Description

Method for identifying multiple dynamic impact signals by using lightweight neural network Technical Field The invention relates to the field of dynamics and machine learning, in particular to a method for identifying multiple dynamic impact signals by using a lightweight neural network. Background The multi-impact load is widely applied to various industrial scenes such as aerowings, high-speed train bogies, large mechanical transmission systems and the like, and the signal characteristics extracted from the multi-impact load are beneficial to realizing important functions such as load mode identification, fault diagnosis, system prediction and the like. However, the multiple impact load signal itself has strong noise and random interference, making feature extraction of the relevant load signal difficult. In recent years, deep learning has shown great potential in the aspect of multi-impact signal identification, and opens up a new front for a signal-based multi-dynamic influence system method. These deep learning based approaches do not rely on particularly elaborate filter designs, which makes them more efficient prediction and identification in many practical engineering applications, and deep learning approaches have met with great success in a wide range of applications involving a variety of dynamic effects. In a variety of different scenarios of impact loads, load identification of the penetrating ammunition is most difficult in military use. Penetration ammunition is required to penetrate multiple solid protective structures and accurately detonate on a given layer to achieve effective destruction of a particular target. The designated detonation layer is dependent on layer identification of the multiple impact load signals of the projectile acceleration sensor. However, the aggressive ammunition multi-impact load signal presents unique "signal attachment" problems in the time domain in addition to strong noise and random interference similar to many other applications, making its load identification more difficult. Due to the ultra high impact velocity and complex internal structure of the ammunition, the "signal blocking" is essentially a real penetration overload signal mixed with the coupled oscillating signal. In addition to "signal blocking," multiple impact signal identification of the penetrating ammunition is also limited by hardware and some factors of operating conditions. Therefore, the application of the deep learning method in this field requires much additional effort. First, deep learning methods typically require a large number of data records, and the long and costly experimental period of the penetration ammunition makes data recording very difficult. Second, conventional large-scale neural networks are difficult to deploy directly due to hardware resource limitations. Therefore, lightweight clipping of neural network models and acceleration of operation under limited resources are important issues that must be studied. Again, since the high-speed dynamic multiple impact signal is a non-stationary random signal and is accompanied by severe signal blocking, it is difficult to directly extract the signal characteristics. Disclosure of Invention Based on the above analysis discussion, the invention discloses a method for identifying multiple dynamic impact signals by using a lightweight neural network, which comprises the steps of firstly establishing a multiple dynamic impact signal database and classifying according to data characteristics. And secondly, carrying out feature extraction on different dynamic impact signals by using a continuous wavelet transformation method, and designing a light neural network based on an attention mechanism according to feature information. And finally, training and verifying the neural network by combining the data characteristics, so as to ensure the validity of the designed method. A multi-bullet collaborative countermeasure strategy acquisition method based on the combination of a virtual force method and a primer cooperation comprises the following steps: Step 1, performing dynamic impact tests on targets under different working conditions, collecting dynamic impact signals under different working conditions, establishing a database of multiple dynamic impact signals of the targets, classifying the dynamic impact signals in the database according to impact characteristics to obtain classified dynamic impact signals, and turning to step 2. And 2, carrying out feature extraction on the classified dynamic impact signals by using a continuous wavelet change method, correspondingly obtaining feature data, forming a training set and a testing set by the feature data, and turning to the step 3. And 3, constructing a lightweight neural network framework, and turning to step 4. And 4, training the lightweight neural network framework by using the training set to obtain the lightweight neural network, and turning to step 5. And 5, inputting the test s