Anti-interference open-set identification method, system and device for magnetoelectric composite detection signal

Abstract

The invention discloses an anti-interference open-set identification method, system and device for a magnetoelectric composite detection signal, and relates to the technical field of signal processing, and the method comprises the steps: collecting interference signal data containing known noise in a to-be-detected region; training a network model based on a hierarchical channel perception attention framework by using signals of the interference signal data after expansion in a time domain, a frequency domain, a wavelet domain and a decomposition domain; inputting signals of a to-be-detected unknown interference signal data set expanded in a time domain, a frequency domain, a wavelet domain and a decomposition domain into the trained network model based on the hierarchical channel perception attention framework, and respectively obtaining identification probabilities and confidence coefficients of the time domain, the frequency domain, the wavelet domain and the decomposition domain on each branch; outputting the current identification type according to the identification probability and the confidence coefficient of each branch; removing the identified noise and unknown noise to obtain a processed detection signal; according to the method, accurate identification and interference filtering of the target signal are realized, and the robustness and reliability of the system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of signal processing, and in particular to an anti-interference open set identification method, system and device for magneto-electric composite detection signals. Background Art

[0002] Non-visual passive target detection, this type of target detection method is mainly aimed at non-cooperative targets. Since it does not actively transmit signals to the outside, it has good concealment compared to traditional radar detection and infrared detection, and the detection distance is far, so it has a good development prospect. At present, electrostatic detection and magnetic detection are the main methods in the field. Electrostatic detection has the advantages of high sensitivity, moderate range and low false alarm rate, and shows its unique advantages in accurately identifying moving targets; while magnetic detection plays a key role in various environments and tactical requirements by virtue of its concealment, wide range of application, no influence of climate and close-range detection capabilities. Although electrostatic detection and magnetic detection have taken anti-interference measures in terms of system, working mode, signal processing, etc., a single detection mode has been difficult to obtain sufficient detection reliability in various complex application scenarios. Therefore, at this stage, it is explored to combine magnetic field detection with electrostatic field detection to form a composite detection.

[0003] In existing non-visual target detection technologies such as magnetic signal and electrostatic detection, the time-frequency domain characteristics of the target signal combined with the threshold are mostly used as recognition and judgment conditions, and almost no distinction is made between the identified targets. Under ideal conditions, the energy difference between the target to be detected and the background noise is large, so effective target recognition can be achieved.

[0004] However, in actual applications, due to the uncertainty of external factors such as the environment and similar targets, when identifying the magnetic field / electric field of a specific type of target, it is very easy to be disturbed by the external environment, resulting in false triggering and missed triggering. The particularity of its field application has led to relatively few related studies. Most of the existing views focus on two methods: increasing the number of sensors (multi-sensor composite detection) and real-time signal processing (such as filtering, noise reduction, etc.). Among them, the composite detection method has a better effect. Starting from multiple features of the target to be identified, in most cases, it can achieve the separation of conventional background noise and the target object, such as Figure 1 , demonstrating the influence of weak background interference signals and strong interference background signals on effective signal detection in the composite detection mode. However, the increase in the types of information captured also means the introduction of new dimensions of noise signals. For some complex working scenarios (such as near-ground detection, interference from similar targets, etc.), there are still many problems such as incomplete target recognition, unprocessed unknown signals, and single noise reduction effect. Summary of the invention

[0005] In view of the shortcomings of the prior art, such as weak anti-interference ability, low recognition accuracy and inability to identify unknown interference in complex environments, the present invention proposes an anti-interference open set recognition method, system and device for magneto-electric composite detection signals, which solves the problems existing in the prior art by combining electromagnetic composite detection with open set recognition technology, and introducing multi-domain feature extraction, hierarchical channel perception attention framework and relative entropy interference filtering methods.

[0006] An anti-interference open set identification method for a magneto-electric composite detection signal comprises the following steps:

[0007] Obtain historical interference signal data containing known noise;

[0008] The historical interference signal data containing known noise is expanded in the time domain, frequency domain, wavelet domain and decomposition domain, and the expanded signal is input into the network model based on the hierarchical channel-aware attention framework for training: after the information of the signal in different domains is converted into a feature representation tensor through the convolutional neural network layer, it is input into the convolutional attention mechanism layer to generate a weighted matrix; each feature is selected and biased according to the weighted matrix to generate a selectively enhanced feature tensor, which is input into the hierarchical channel-aware attention layer to obtain the weights of different channels, and the features are selected in the channel dimension by the weights; the generated features after selection are input into the fully connected layer, and the recognition probability is output to complete the training of the network model based on the hierarchical channel-aware attention framework; wherein, in the hierarchical channel-aware attention layer, the output of the convolution layer in ResNet18, the output of the Inception module and the original input of the module are spliced ​​in the channel dimension, and the spliced ​​features are processed by two convolution modules to obtain a set of weights, which are used to selectively enhance the features in the channel dimension, and the enhanced features are reduced in dimension and depth enhanced by the ResNet downsampling module;

[0009] The unknown interference signal data to be tested is expanded in the time domain, frequency domain, wavelet domain and decomposition domain, and the expanded signal is input into the trained network model based on the hierarchical channel-aware attention framework to obtain the recognition probability and confidence of each branch in the time domain, frequency domain, wavelet domain and decomposition domain respectively; according to the recognition probability and confidence, the recognition results of the current branches are judged, and the recognition results containing noise and unknown noise are removed to obtain the detection signal of the area to be tested.

[0010] Furthermore, the interference signal data is expanded to the frequency domain, wavelet domain and decomposition domain based on the time domain through Fourier transform, wavelet transform and VMD decomposition, which are specifically expressed as follows:

[0011] Frequency domain:

[0012] Wavelet domain:

[0013] VMD decomposition:

[0014] Among them, X(w) is the Fourier transform of the original signal x(t), which represents the representation of the signal in the frequency domain; j is the imaginary unit; w is the angular frequency; t is the time; X(a,τ) is the wavelet coefficient, which represents the wavelet transform value of the signal x(t) at scale a,τ, where a and τ are scale parameters and translation parameters respectively; is the wavelet coefficient shifted in time t; u k is the kth mode function, representing the kth component in the signal; ω k is the center frequency of the kth mode function; Indicates u k Hilbert transform on a time series t; represents the frequency modulation of the kth mode function; is the gradient operator; f(t) is the time series signal to be decomposed.

[0015] Furthermore, the convolutional neural network layer captures the local spatial correlation of the input signal through a convolution operation, generates a correlation matrix, establishes a nonlinear mapping relationship between features and targets through an activation function based on the correlation matrix, and generates a feature representation tensor with multi-dimensional features.

[0016] Furthermore, in the ResNet downsampling module, the input features are processed by the convolution layer, and the input features are added to the features processed by the convolution layer through jump connections to generate a residual map, thereby achieving dimensionality reduction and depth enhancement of the features.

[0017] Furthermore, judging the recognition results of the current branches according to the recognition probability and the confidence, removing the recognition results containing noise and unknown noise, and obtaining the detection signal of the area to be detected specifically includes the following steps:

[0018] According to the recognition probability of each branch in the time domain, frequency domain, wavelet domain and decomposition domain, the maximum index value of its recognition probability is obtained, and the recognition results of each current branch are judged; if the recognition results of each current branch are different, they are rejected as unknown type results; if they are the same, the confidence of each branch in the time domain, frequency domain, wavelet domain and decomposition domain is sorted to obtain the confidence thresholds of the time domain, frequency domain, wavelet domain and decomposition domain;

[0019] The confidence of each branch in the time domain, frequency domain, wavelet domain and decomposition domain is compared with the corresponding confidence threshold. If it is not greater than the corresponding confidence threshold, the current recognition type is output, otherwise the current sample is rejected as unknown noise;

[0020] The identified noise and unknown noise are removed to obtain the processed detection signal.

[0021] Furthermore, the confidence corresponding to each branch is determined by using the relative entropy between the recognition probability and the ideal probability in the time domain, frequency domain, composite domain and decomposition domain output by the network model based on the hierarchical channel-aware attention framework. The relative entropy is expressed as:

[0022]

[0023] Where p(x) is the probability distribution of the network model output, and q(x) represents the ideal probability distribution of the unknown interference signal. The probability distribution and the value of zero in the ideal probability distribution are updated to 1×10 -9 , determine whether the sum of the probability distribution is greater than 1. If it is greater than 1, subtract the part greater than 1 from the maximum probability value. x represents the subset signal in the set of signals X to be processed.

[0024] Furthermore, the interference signal data includes background steady-state interference and deceptive interference; the background steady-state interference includes self-friction and environmental thermal noise; the deceptive interference includes explosion, discharge, continuous pulse, intermittent pulse, and power frequency interference.

[0025] The present invention also includes an anti-interference open set identification system for magneto-electric composite detection signals, comprising:

[0026] An acquisition module, used for acquiring historical interference signal data containing known noise;

[0027] The model training module is used to expand the historical interference signal data containing known noise in the time domain, frequency domain, wavelet domain and decomposition domain, and input the expanded signal into the network model based on the hierarchical channel-aware attention framework for training: after the information of the signal in different domains is converted into a feature representation tensor through the convolutional neural network layer, it is input into the convolutional attention mechanism layer to generate a weighted matrix; each feature is selected and weighted according to the weighted matrix to generate a selectively enhanced feature tensor, which is input into the hierarchical channel-aware attention layer to obtain the weights of different channels, and the features are weighted in the channel dimension through the weights. Row selection; input the generated features after selection into the fully connected layer, output the recognition probability, and complete the training of the network model based on the hierarchical channel-aware attention framework; in the hierarchical channel-aware attention layer, the output of the convolution layer in ResNet18, the output of the Inception module, and the original input of the module are spliced ​​in the channel dimension, and the spliced ​​features are processed by two convolution modules to obtain a set of weights. The weights are used to selectively enhance the features in the channel dimension, and the enhanced features are reduced in dimension and depth enhanced by the ResNet downsampling module;

[0028] The generation module is used to expand the unknown interference signal data to be tested in the time domain, frequency domain, wavelet domain and decomposition domain, and input the expanded signal into the trained network model based on the hierarchical channel-aware attention framework to obtain the recognition probability and confidence of each branch in the time domain, frequency domain, wavelet domain and decomposition domain respectively; according to the recognition probability and confidence, the recognition results of each current branch are judged, and the recognition results containing noise and unknown noise are removed to obtain the detection signal of the area to be tested.

[0029] The present invention also includes a computer device for anti-interference open-set identification of magneto-electric composite detection signals, comprising: a memory, a processor, and a computer program stored in the memory. When the processor executes the computer program, the steps of the anti-interference open-set identification method of the magneto-electric composite detection signal are implemented.

[0030] The present invention also includes a readable storage medium, which stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the anti-interference open set identification method of the magneto-electric composite detection signal.

[0031] The present invention provides an anti-interference open set identification method, system and device for magnetoelectric composite detection signals, which have the following beneficial effects:

[0032] Aiming at the problems of weak recognition and anti-interference capability and low interception efficiency of current near-ground magnetoelectric composite detection technology in complex environments due to background steady-state interference and deceptive interference, the present invention proposes an interference target open-set recognition and denoising method for passive electromagnetic composite detection. The method implements network pre-training by extending the detection signal to the frequency domain, wavelet domain and decomposition domain, and then combines the hierarchical channel-aware attention framework to effectively extract the features of signals in different domains. The method also utilizes the interference filtering method based on relative entropy and the open-set recognition module to achieve accurate recognition and interference filtering of target signals, thereby improving the robustness and reliability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 The noisy signal under weak background interference and the noisy signal under strong background interference in the embodiment of the present invention and their corresponding ideal effective target signals;

[0034] Figure 2 It is a framework diagram of the anti-interference open set identification method of the magneto-electric composite detection signal in an embodiment of the present invention;

[0035] Figure 3 It is a time domain schematic diagram of a noisy signal detected by magneto-electric composite detection in an embodiment of the present invention;

[0036] Figure 4It is a frequency domain schematic diagram of a noisy signal in magneto-electric composite detection according to an embodiment of the present invention;

[0037] Figure 5 Schematic diagram of the wavelet domain of the noisy signal of the magneto-electric composite detection in the embodiment of the present invention;

[0038] Figure 6 This is a VMD decomposition diagram of the explosion background noise signal detected by magneto-electric composite detection in an embodiment of the present invention;

[0039] Figure 7 This is a VMD decomposition diagram of the friction and thermal noise signal detected by magneto-electric composite detection in an embodiment of the present invention;

[0040] Figure 8 Schematic diagram of VMD decomposition of continuous pulse background noise in magneto-electric composite detection in an embodiment of the present invention;

[0041] Fig. 9 Schematic diagram of VMD decomposition of power frequency interference background noise in magneto-electric composite detection according to an embodiment of the present invention;

[0042] Fig.10 Schematic diagram of VMD decomposition of intermittent pulse background noise in magneto-electric composite detection in an embodiment of the present invention;

[0043] Fig.11 Schematic diagram of VMD decomposition of pulse discharge background noise in magneto-electric composite detection according to an embodiment of the present invention;

[0044] Fig.12 It is a network model architecture based on a hierarchical channel-aware attention framework in an embodiment of the present invention;

[0045] Fig.13 Schematic diagram of an open set identification module in an embodiment of the present invention;

[0046] Fig.14 It is a classification result diagram of each identification classification label in an embodiment of the present invention;

[0047] Fig.15 Schematic diagram of signal comparison before and after denoising in an embodiment of the present invention; (a) is a target detection signal under strong noise interference; (b) is a target detection signal after denoising;

[0048] Fig.16 Schematic diagram of the hardware system structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0050] The present invention proposes an anti-interference open set recognition method for magnetoelectric composite detection signals with high robustness and recognition accuracy. First, a part of a priori target signals and known noise interference signals are determined, and a network pre-training data set is established based on the components of the time-frequency domain, composite domain, and decomposition domain of the existing signal by using a signal decomposition and feature expansion module to complete the pre-training of a hierarchical channel perception attention network; secondly, the unknown interference signal actually collected is also input into the network after being decomposed and expanded, and based on the existing pre-training framework, the classification confidence of the input signal is calculated to identify the known interference component, the unknown interference component, and the effective component information of the target to be identified; then, the classification success probability, confidence and other indicators of each identified component information are voted and classified to realize the open set recognition of the unknown signal; finally, the effective component information of the target to be identified is restored separately in multiple domains, and all other noise classifications are discarded, so as to finally realize the output of the noise reduction signal.

[0051] like Figure 2 As shown, the method specifically comprises the following steps:

[0052] S1. Collect the initial noise data set, which consists of interference signals: background steady-state interference (self-friction, environmental thermal noise) and deceptive interference (explosion, discharge, continuous pulse, intermittent pulse, power frequency interference).

[0053] In order to more accurately capture the features of the signal, the signal is expanded in three dimensions: frequency domain, wavelet domain, and decomposition domain (with VMD decomposition as the main body) based on the time domain. The transformation formula is shown below to complete the construction of the pre-training data set. Figure 3 Schematic diagram of the time domain of magnetoelectric composite detection of noisy signals (a) is a schematic diagram of the time domain of explosion signals, (b) is a schematic diagram of the time domain of friction + thermal noise, (c) is a schematic diagram of the time domain of continuous pulses, (d) is a schematic diagram of the time domain of power frequency interference, (e) is a schematic diagram of the time domain of intermittent pulses, and (f) is a schematic diagram of the time domain of pulse discharge; Figure 4 Schematic diagram of the frequency domain of magnetoelectric composite detection of noisy signals, (a) is a schematic diagram of the frequency domain of explosion signals, (b) is a schematic diagram of the frequency domain of friction + thermal noise, (c) is a schematic diagram of the frequency domain of continuous pulses, (d) is a schematic diagram of the frequency domain of power frequency interference, (e) is a schematic diagram of the frequency domain of intermittent pulses, and (f) is a schematic diagram of the frequency domain of pulse discharges; Figure 5 Schematic diagram of the wavelet domain of the noisy signal of the magnetoelectric composite detection, (a) is the wavelet domain diagram of the explosion signal, (b) is the wavelet domain diagram of the friction + thermal noise, (c) is the wavelet domain diagram of the continuous pulse, (d) is the wavelet domain diagram of the power frequency interference, (e) is the wavelet domain diagram of the intermittent pulse, and (f) is the wavelet domain diagram of the pulse discharge; Figure 6 This is the VMD decomposition diagram of the explosion background noise signal detected by magneto-electric composite detection; Figure 7 This is the VMD decomposition diagram of the friction and thermal noise signal detected by magneto-electric composite detection; Figure 8Schematic diagram of VMD decomposition of continuous pulse background noise in magneto-electric composite detection; Fig. 9 This is a schematic diagram of VMD decomposition of power frequency interference background noise in magneto-electric composite detection; Fig.10 Schematic diagram of VMD decomposition of intermittent pulse background noise in magneto-electric composite detection; Fig.11 This is a schematic diagram of the VMD decomposition of the background noise of the pulse discharge in the magneto-electric composite detection.

[0054] Frequency domain:

[0055] Wavelet domain:

[0056] VMD decomposition:

[0057] Among them, in the frequency domain decomposition function, X(w) is the Fourier transform of the original signal x(t), which represents the representation of the signal in the frequency domain, e -jwt In the wavelet domain decomposition function, X(a,τ) is the wavelet coefficient, which represents the wavelet transform value of the signal x(t) at scale a,τ. a and τ are scale parameters and translation parameters respectively. is the wavelet coefficient shifted in time t. In the VMD decomposition domain, u k is the kth mode function, which is an AM-FM signal with a center frequency, representing the kth component in the signal; ω k is the center frequency of the kth mode function, usually obtained by Hilbert transform, indicating the frequency center of each mode function, with a total of K signals; As a whole, it represents u k The Hilbert transform of the time series t is represents the frequency modulation of the kth mode function, is the gradient operator, which means to obtain the gradient of the subsequent content, and f(t) is the time series signal to be decomposed.

[0058] S2. Build a basic network architecture based on the hierarchical channel-aware attention framework, randomly select 70% of the data set for network pre-training, and use the other 30% of the data set as a validation set to complete the pre-training of the classification recognition network.

[0059] Among them, the basic network architecture based on the hierarchical channel-aware attention framework is adopted, such as Fig.12As shown in the figure, the module adopts a cascade structure of a multi-layer channel attention mechanism module and a ResNet module. The purpose of setting up the multi-layer channel attention mechanism module is to enhance feature representation, emphasize important feature channels in the training set data, and suppress unimportant channels, so that the network pays more attention to information that is useful for achieving the final goal. By connecting multiple modules in series, richer feature representations can be learned at different levels of abstraction, thereby improving the ability to understand input data, enhancing feature expression capabilities, and performing hierarchical information fusion, but the cost is high computational complexity and an increase in the number of parameters. Therefore, in order to alleviate this cost, a ResNet downsampling module is inserted in the series connection of the multi-layer channel attention mechanism. Its function is to control the number of feature parameters. It can reduce the size of the feature map while maintaining useful information by increasing the depth of the feature map (i.e., the number of channels), which helps prevent the loss of important feature information during downsampling.

[0060] This paper proposes a basic network architecture based on a hierarchical channel-aware attention framework, such as Fig.12 As shown in FIG. 1 , the network architecture includes a convolutional neural network layer, a convolutional attention mechanism layer, a hierarchical channel-aware attention layer and a fully connected layer. The network model based on the hierarchical channel-aware attention framework is trained by inputting the signal after the interference signal data containing known noise is expanded in the time domain, frequency domain, wavelet domain and decomposition domain. The training process includes the following steps:

[0061] The signal data containing known noise is expanded in the time domain, frequency domain, wavelet domain and decomposition domain and then input into the network model based on the hierarchical channel-aware attention framework:

[0062] The convolutional neural network layer extracts features from the input signal and converts the information of the signal in different domains (time domain, frequency domain, wavelet domain, decomposition domain) into feature representation tensors. The convolution operation can capture the local spatial correlation of the signal, form a correlation matrix, and preliminarily establish a nonlinear mapping relationship between features and targets through activation functions, and generate a tensor with multi-dimensional features as input for subsequent processing.

[0063] The output of the convolutional neural network layer is input into the convolutional attention mechanism layer. The attention mechanism generates a weighted matrix to select and emphasize different features and enhance the focus on key features. The attention mechanism generates a weight distribution based on the feature importance of different channels. These weights are used to enhance the features with high correlation in the channel dimension while suppressing redundant features. The output result is a selectively enhanced feature tensor.

[0064] The output of the convolutional attention mechanism layer is input into the hierarchical channel-aware attention layer to obtain the weights of different channels, and the features are selected in the channel dimension through the weights; the generated features of the hierarchical channel-aware attention layer are input into the fully connected layer to output the recognition probability;

[0065] Among them, the hierarchical channel-aware attention layer includes multiple multi-layer channel attention mechanism modules and Res Net downsampling modules. In each multi-layer channel attention mechanism module, the output result of the convolutional attention mechanism layer is input into the hierarchical channel-aware attention layer. The output of the convolutional layer in ResNet18, the output of the Inception module, and the original input x of the module are spliced ​​in the channel dimension, and the spliced ​​features are processed through two convolution modules to output features.

[0066] The output of the convolutional attention mechanism layer is input into the hierarchical channel-aware attention layer. The multi-layer channel attention mechanism module is connected in parallel with an Inception module based on ResNet18. At the same time, the output of the convolution in ResNet18, the output of the Inception structure, and the input x of the module are spliced ​​in the channel dimension. The output is passed through two convolution modules, and the weights of the output of the convolution in ResNet18, the output of the Inception structure, and the input data of the module are output. The features are selected in the channel dimension through the weights. The characteristics of the residual block are retained, and the diversity of feature extraction and active selection of features are realized.

[0067] The ResNet downsampling module achieves dimensionality reduction and depth enhancement of feature tensors through residual connections and convolution operations. First, the input feature tensor is processed by the convolution layer, the size is reduced, and the number of channels is increased to retain useful feature information. Then, the original input features are added to the convolutional features through skip connections to form a residual map, which helps alleviate the gradient vanishing problem in deep networks and enables the network to better learn important features.

[0068] The present invention improves the design of the multi-layer channel attention mechanism module structure in the structure. The module is connected in parallel with an Inception module on the basis of ResNet, and its purpose is to enhance the feature extraction capability of the network.

[0069] The feature extraction and processing process is as follows:

[0070] First, the convolutional layer output in ResNet18, the output of the Inception module, and the original input x of the module are concatenated in the channel dimension. The purpose of this step is to fuse the feature information from different network layers together to form a richer feature representation; then, the concatenated features are processed by two convolution modules, which are used to learn a set of weights that can reflect the importance of different feature channels; finally, these learned weights are used to selectively enhance the features in the channel dimension, with the goal of allowing the network to selectively emphasize certain features based on the importance of each channel's features to the task, while ignoring or reducing the weights of other features.

[0071] By improving the parallel structure in the multi-layer channel attention mechanism module structure, not only the advantage of the residual block in ResNet in effectively training deep networks is retained, but also the diversity and flexibility of feature extraction are increased. This ability to actively select features enables the network to more accurately identify and extract information that is critical to the task, thereby improving the overall recognition efficiency and accuracy, and better serving the recognition of interference features in modal component signals under different levels of decomposition.

[0072] In order to further improve the recognition accuracy of the model, the present invention adopts a noise open set recognition method based on relative entropy. The core idea of ​​this method is to use the recognition probability output by the model to evaluate whether the sample belongs to a known category or is an unknown interference.

[0073] To explain further, the recognition probability calculated by the model, especially the output of the softmax layer, can be interpreted as the probability distribution of the model for the input sample belonging to each category, where the sum of the probabilities of all categories is 1. Ideally, for a clear category, the output of the model should be that the probability of judging the category as the correct category is close to 1, and the probability of judging all other categories as wrong categories is close to 0.

[0074] In the solution of the present invention, the difference between the actual output probability and this ideal probability distribution is quantified by calculating the relative entropy between them. Relative entropy, also known as KL divergence, is an indicator that measures the difference between two probability distributions. In this solution, the probability distribution of the model output is compared with the ideal probability distribution, and the calculated relative entropy value can be regarded as a confidence indicator for judging whether the input sample belongs to an unknown interference type. The higher the relative entropy value, the greater the difference between the output probability distribution of the model and the ideal distribution, suggesting that the sample may be a new interference that the model has never seen. In this way, the network can more accurately identify and filter out unknown interference, thereby improving the recognition accuracy of the model.

[0075] The expression of relative entropy used is as follows:

[0076]

[0077] Relative entropy measures the difference between a sample p(x) and its ideal distribution q(x), where p(x) is the probability distribution of the output of a deep neural network, and q(x) is the ideal probability distribution of a certain type of interference signal.

[0078] The steps are:

[0079] 1) Obtain the judgment probability output by the deep neural network.

[0080] 2) Obtain the type determined by probability.

[0081] 3) The relative entropy of the calculated output probability and the ideal probability of the current category.

[0082] Since some data are zero in the ideal probability distribution, and the calculation of relative entropy requires the ideal probability as a divisor, the data needs to be preprocessed:

[0083] 1) Update the probability distribution and the value of zero in the ideal probability distribution to 1×10 -9 .

[0084] 2) Determine whether the sum of the probability distribution is greater than 1. If it is greater than 1, subtract the part greater than 1 from the maximum probability value. This operation ensures that the sum of the probability distribution is 1, and also avoids the interference of the 0 value on the calculation.

[0085] S3, input the time domain, frequency domain, wavelet domain and decomposed signal of the unknown interference signal to be tested into the trained deep neural network to obtain the recognition probability P of the time domain branch TF , confidence D TF ; Identification probability P of frequency domain branch RD , confidence D RD ; Identification probability of the branch in wavelet domain P WR , confidence D WR ; The recognition probability P of the decomposed signal VMD , confidence D VMD . TF , D RD , D WR and D VMD Perform sorting statistics to obtain the confidence threshold th TF、 th RD、 th WR and VMD , where 90% of the data in the training set is less than this threshold, and the data greater than this threshold can be considered rejected. The process is as follows Figure 7 shown.

[0086] S4. Obtain the maximum index value of the recognition probability output by each branch, and judge it as the recognition result of each branch at present, which are D TF,I , D RD,I , D WR,I and D VMD,I If D TF,I =D RD,I =D WR,I =D VMD,I , then go to S5, otherwise directly reject the sample as an unknown type sample.

[0087] S5, the recognition probability confidence D on each branch in S3 TFF,I , D RD,I , D WR,I and D VMD,I Respectively with D TF,I =D RD,I =D WR,I =D VMD,I The threshold of the determined type is compared. TF ≤th TF &&D RD ≤th RD &&D WR,I ≤th WR &&D VMD,I ≤th VMD , the input sample is considered to be the current recognition category, otherwise the current sample is rejected as unknown noise; the process of its open set recognition module is as follows Fig.13 As shown, Fig.14 Schematic diagram of classification results for each identification classification label.

[0088] S6. Remove the identified noise and unknown noise to obtain a processed detection signal. Fig.15 The figure shows the restoration of the effective signal according to the effective sub-target signal component in the classification result, and the schematic diagram of the signal comparison before and after removing the background noise. (a) is the schematic diagram of the intersection signal under noise, and (b) is the intersection signal after processing.

[0089] The method mentioned in the design of the present invention can rely on the following systems, including sensor drive interface, signal conditioning, embedded data acquisition and processing, power supply module, etc. Each system adopts an embedded and scalable design, and the algorithm can be directly integrated into the data analysis module of the main control unit during the development process. The specific structure of the system is as follows Fig.16 shown.

[0090] When the hardware system of this structure is working, firstly, the magnetic detection module uses the tunneling magnetoresistive sensor (TMR magnetic sensor) as the core device to obtain magnetic field information in real time. After the sensor output signal is processed by the hardware, it is input into the embedded processing unit through the A / D conversion module to further process the signal and analyze the magnetic field data obtained by the sensor. Secondly, the electrostatic field detection module works based on the electrostatic field theory, and the induction electrode charge is monitored in real time. According to the change of QV conversion voltage, the position of the bomb relative to the target is judged and the output of the detonation signal is controlled. The substrate voltage power supply adjustment module outputs a small current to detect the substrate signal, and continuously adjusts the substrate signal current according to the state of the induction electrode to make its voltage value output stable. It is mainly used to achieve stable operation under different environmental conditions and improve the system's adaptability. Repeat this process until the acquired magnetic signal meets the target magnetic characteristics and composite magnetic target recognition criteria.

[0091] The present invention proposes an open-set recognition and denoising method for interference targets for passive electromagnetic composite detection. The detection signal is expanded to the frequency domain, wavelet domain and decomposition domain to realize network pre-training. Then, the hierarchical channel-aware attention framework is combined to effectively extract the features of signals in different domains. The interference filtering method based on relative entropy and the open-set recognition module are used to realize accurate recognition and interference filtering of target signals, thereby improving the robustness and reliability of the system.

[0092] Based on the same inventive concept, the present invention also proposes an anti-interference open set identification system for magneto-electric composite detection signals, comprising:

[0093] The acquisition module is used to acquire historical interference signal data containing known noise.

[0094] The model training module is used to train the network model based on the hierarchical channel-aware attention framework by inputting the signal after expanding the historical interference signal data containing known noise in the time domain, frequency domain, wavelet domain and decomposition domain: after the information of the signal in different domains is converted into a feature representation tensor through the convolutional neural network layer, it is input into the convolutional attention mechanism layer to generate a weighted matrix; each feature is selected and weighted according to the weighted matrix to generate a selectively enhanced feature tensor, which is input into the hierarchical channel-aware attention layer to obtain the weights of different channels, and the features are selected in the channel dimension through the weights. Selection; the generated features after selection are input into the fully connected layer, and the recognition probability is output to complete the training of the network model based on the hierarchical channel-aware attention framework; wherein, in the hierarchical channel-aware attention layer, the output of the convolutional layer in ResNet18, the output of the Inception module, and the original input of the module are spliced ​​in the channel dimension, and the spliced ​​features are processed by two convolution modules to obtain a set of weights, and the weights are used to selectively enhance the features in the channel dimension, and the enhanced features are reduced in dimension and enhanced in depth through the ResNet downsampling module.

[0095] A generation module is used to input the signals of the unknown interference signal data to be tested after expansion in the time domain, frequency domain, wavelet domain and decomposition domain into the trained network model based on the hierarchical channel perception attention framework, and obtain the recognition probability and confidence of each branch in the time domain, frequency domain, wavelet domain and decomposition domain respectively; according to the recognition probability and confidence, the recognition results of the current branches are judged, and the recognition results containing noise and unknown noise are removed to obtain the detection signal of the area to be tested.

[0096] The present invention also proposes a computer device for anti-interference open-set identification of magnetoelectric composite detection signals, comprising: a memory, a processor, and a computer program stored in the memory. When the processor executes the computer program, the steps of the anti-interference open-set identification method of the magnetoelectric composite detection signal are implemented.

[0097] The present invention also proposes a readable storage medium, which stores a computer program. The computer program includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of an anti-interference open set identification method for magneto-electric composite detection signals.

[0098] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for anti-interference open set identification of magneto-electric composite detection signals, characterized in that: The following steps are involved: Obtain historical interference signal data containing known noise; The historical interference signal data containing known noise is expanded in the time domain, frequency domain, wavelet domain and decomposition domain, and the expanded signal is input into the network model based on the hierarchical channel-aware attention framework for training: after the information of the signal in different domains is converted into a feature representation tensor through the convolutional neural network layer, it is input into the convolutional attention mechanism layer to generate a weighted matrix; each feature is selected and biased according to the weighted matrix to generate a selectively enhanced feature tensor, which is input into the hierarchical channel-aware attention layer to obtain the weights of different channels, and the features are selected in the channel dimension by the weights; the generated features after selection are input into the fully connected layer, and the recognition probability is output to complete the training of the network model based on the hierarchical channel-aware attention framework; wherein, in the hierarchical channel-aware attention layer, the output of the convolution layer in ResNet18, the output of the Inception module and the original input of the module are spliced ​​in the channel dimension, and the spliced ​​features are processed by two convolution modules to obtain a set of weights, which are used to selectively enhance the features in the channel dimension, and the enhanced features are reduced in dimension and depth enhanced by the ResNet downsampling module; The unknown interference signal data to be tested is expanded in the time domain, frequency domain, wavelet domain and decomposition domain, and the expanded signal is input into the trained network model based on the hierarchical channel-aware attention framework to obtain the recognition probability and confidence of each branch in the time domain, frequency domain, wavelet domain and decomposition domain respectively; according to the recognition probability and confidence, the recognition results of the current branches are judged, and the recognition results containing noise and unknown noise are removed to obtain the detection signal of the area to be tested.

2. The anti-interference open set identification method of a magneto-electric composite detection signal according to claim 1 is characterized in that: Through Fourier transform, wavelet transform and VMD decomposition, the interference signal data is expanded to frequency domain, wavelet domain and decomposition domain based on the time domain, which are specifically expressed as follows: Frequency domain: Small wave range: VMD decomposition: Among them, X(w) is the Fourier transform of the original signal x(t), which represents the representation of the signal in the frequency domain; j is the imaginary unit; w is the angular frequency; t is the time; X(a, τ) is the wavelet coefficient, which represents the wavelet transform value of the signal x(t) at scale a, τ, where a and τ are scale parameters and translation parameters respectively; is the wavelet coefficient shifted in time t; u k is the kth mode function, representing the kth component in the signal; ω k is the center frequency of the kth mode function; Indicates u k Hilbert transform on a time series t; represents the frequency modulation of the kth mode function; is the gradient operator; f(t) is the time series signal to be decomposed.

3. The anti-interference open set identification method of a magneto-electric composite detection signal according to claim 1 is characterized in that: The convolutional neural network layer captures the local spatial correlation of the input signal through the convolution operation, generates a correlation matrix, establishes a nonlinear mapping relationship between the feature and the target through the activation function according to the correlation matrix, and generates a feature representation tensor with multi-dimensional features.

4. The anti-interference open set identification method of a magneto-electric composite detection signal according to claim 1, characterized in that: In the ResNet downsampling module, the input features are processed by the convolution layer, and the input features are added to the features processed by the convolution layer through jump connections to generate residual mapping, thereby achieving dimensionality reduction and depth enhancement of the features.

5. The anti-interference open set identification method of a magneto-electric composite detection signal according to claim 1, characterized in that: The method of judging the recognition results of each current branch according to the recognition probability and the confidence, removing the recognition results containing noise and unknown noise, and obtaining the detection signal of the area to be detected specifically includes the following steps: According to the recognition probability of each branch in the time domain, frequency domain, wavelet domain and decomposition domain, the maximum index value of its recognition probability is obtained, and the recognition results of each current branch are judged; if the recognition results of each current branch are different, they are rejected as unknown type results; if they are the same, the confidence of each branch in the time domain, frequency domain, wavelet domain and decomposition domain is sorted to obtain the confidence thresholds of the time domain, frequency domain, wavelet domain and decomposition domain; The confidence of each branch in the time domain, frequency domain, wavelet domain and decomposition domain is compared with the corresponding confidence threshold. If it is not greater than the corresponding confidence threshold, the current recognition type is output, otherwise the current sample is rejected as unknown noise; The identified noise and unknown noise are removed to obtain the processed detection signal.

6. The anti-interference open set identification method of a magneto-electric composite detection signal according to claim 1, characterized in that: The recognition results of each branch are judged by using the relative entropy between the recognition probability and the ideal probability in the time domain, frequency domain, composite domain and decomposition domain output by the network model based on the hierarchical channel-aware attention framework. The relative entropy is expressed as: Where p(x) is the probability distribution of the network model output, and q(x) represents the ideal probability distribution of the unknown interference signal. The probability distribution and the value of zero in the ideal probability distribution are updated to 1×10 -9 , determine whether the sum of the probability distribution is greater than 1. If it is greater than 1, subtract the part greater than 1 from the maximum probability value. x represents the subset signal in the set of signals X to be processed.

7. The anti-interference open set identification method of a magneto-electric composite detection signal according to claim 1, characterized in that: The interference signal data includes background steady-state interference and deceptive interference; the background steady-state interference includes self-friction and environmental thermal noise; the deceptive interference includes explosion, discharge, continuous pulse, intermittent pulse, and power frequency interference.

8. An anti-interference open set identification system for magneto-electric composite detection signals, characterized in that: include: An acquisition module, used for acquiring historical interference signal data containing known noise; The model training module is used to expand the historical interference signal data containing known noise in the time domain, frequency domain, wavelet domain and decomposition domain, and input the expanded signal into the network model based on the hierarchical channel-aware attention framework for training: after the information of the signal in different domains is converted into a feature representation tensor through the convolutional neural network layer, it is input into the convolutional attention mechanism layer to generate a weighted matrix; each feature is selected and weighted according to the weighted matrix to generate a selectively enhanced feature tensor, which is input into the hierarchical channel-aware attention layer to obtain the weights of different channels, and the features are weighted in the channel dimension through the weights. Row selection; input the generated features after selection into the fully connected layer, output the recognition probability, and complete the training of the network model based on the hierarchical channel-aware attention framework; in the hierarchical channel-aware attention layer, the output of the convolution layer in ResNet18, the output of the Inception module, and the original input of the module are spliced ​​in the channel dimension, and the spliced ​​features are processed by two convolution modules to obtain a set of weights. The weights are used to selectively enhance the features in the channel dimension, and the enhanced features are reduced in dimension and depth enhanced by the ResNet downsampling module; The generation module is used to expand the unknown interference signal data to be tested in the time domain, frequency domain, wavelet domain and decomposition domain, and input the expanded signal into the trained network model based on the hierarchical channel-aware attention framework to obtain the recognition probability and confidence of each branch in the time domain, frequency domain, wavelet domain and decomposition domain respectively; according to the recognition probability and confidence, the recognition results of each current branch are judged, and the recognition results containing noise and unknown noise are removed to obtain the detection signal of the area to be tested.

9. A computer device for anti-interference open set recognition of magneto-electric composite detection signals, characterized in that: include: A memory, a processor and a computer program stored in the memory, wherein when the processor executes the computer program, the steps of the anti-interference open set identification method for magneto-electric composite detection signals according to any one of claims 1 to 7 are implemented.

10. A readable storage medium, characterized in that: The readable storage medium stores a computer program, which includes program instructions. When the program instructions are executed by a processor, they are used to execute the steps of the anti-interference open set identification method for magneto-electric composite detection signals as described in any one of claims 1 to 7.

Images