Electronic fence early warning method based on DAS system

Abstract

The application discloses an electronic fence early warning method based on a DAS system and belongs to the technical field of electronic fences, and comprises the following steps: S1, multi-domain feature fusion extraction; S2, training sample set construction; S3, parameter optimization; S4, invasion target event determination; and S5, electronic fence early warning. The application adopts a recognition classification algorithm of a time domain, a frequency domain, a wavelet domain and a support vector machine (SVM) in the multi-domain feature fusion extraction, and combines a comprehensive attention convolutional neural network (CA-Net), so that the problems of high coherent fading probability, poor signal-to-noise ratio and weak environmental adaptability in the prior art are solved, and the invasion signal recognition accuracy and the early warning reliability under a complex environment are improved.

Description

Technical Field

[0001] This invention relates to the field of electronic fence technology, and more specifically to an electronic fence early warning method based on a DAS system. Background Technology

[0002] A DAS (Distributed Fiber Optic Acoustic Sensing) system is a technology that uses optical fiber as the sensing medium to continuously and distributedly monitor external disturbance signals such as vibrations and sound waves by detecting phase changes in the backscattered Rayleigh light in the fiber. This technology has advantages such as resistance to electromagnetic interference, corrosion resistance, good concealment, and long-distance monitoring capabilities, and is widely used in perimeter security, pipeline safety monitoring, and seismic wave detection. An electronic fence is a perimeter intrusion detection and early warning system, typically composed of a physical fence (such as barbed wire or a wall) and attached or built-in sensors to detect intrusion behaviors such as climbing, cutting, and touching. Traditional electronic fences often use electrical pulse, infrared beam, or tension sensing methods, while fiber optic electronic fences based on DAS systems utilize optical fiber as a passive sensing unit to achieve real-time spatial positioning and vibration pattern recognition of long-distance perimeter intrusion events, offering advantages such as strong environmental adaptability and convenient concealed construction.

[0003] Existing DAS systems used in electronic fence applications suffer from several drawbacks, including a high probability of coherent fading and poor signal-to-noise ratio, leading to low accuracy in identifying intrusion vibration signals. Furthermore, sampling rate loss limits the frequency response range, and traditional feature extraction methods are limited in scope and lack multi-classification capabilities, impacting early warning reliability. In addition, traditional systems are poorly adaptable to complex environments, prone to false alarms in multi-interference scenarios, and lack dynamic parameter adaptive adjustment mechanisms, making it difficult to cope with changes in signal characteristics under different geographical and climatic conditions, resulting in insufficient long-term operational stability. Summary of the Invention

[0004] The technical problem to be solved by this invention is: how to solve the problems of low accuracy of intrusion vibration signal identification caused by high coherent fading probability and poor signal-to-noise ratio in the prior art, as well as the single feature extraction method and insufficient multi-classification ability of traditional methods, which affect the reliability of early warning. This invention improves the accuracy of intrusion vibration signal identification and the reliability of early warning in complex environments. It provides an electronic fence early warning method based on DAS system, which extracts features from intrusion vibration signals from multiple perspectives, effectively improves the distinguishability of multiple types of intrusion vibration signals, thereby improving the accuracy of event classification and enhancing robustness in complex environments.

[0005] like Figure 4 As shown, the present invention solves the above-mentioned technical problems through the following technical solution, and the present invention includes the following steps:

[0006] S1: Multi-domain feature fusion extraction

[0007] In the fiber optic electronic fence, the intrusion vibration signal collected by the DAS system is subjected to multi-domain feature fusion extraction in the time domain, frequency domain, and wavelet domain to obtain a 9-dimensional feature vector group;

[0008] S2: Construction of Training Sample Set

[0009] The feature vector set is input into the support vector machine (SVM), and signal classification training is completed through hyperplane partitioning to construct a training sample set.

[0010] S3: Parameter Optimization

[0011] By employing grid search combined with K-CV cross-validation, the optimal penalty coefficient c and kernel function parameter g for the training sample set are determined in the LIBSVM software package, and an SVM classification training model is established.

[0012] S4: Intrusion Target Event Determination

[0013] Input the format-matched test samples into the SVM classification training model to predict the behavior category and determine the intrusion target event;

[0014] S5: Electronic Fence Warning

[0015] The intrusion vibration signal corresponding to the intrusion target event is encoded into two-dimensional information through time-frequency conversion, and then input into the comprehensive attention convolutional neural network CA-Net for recognition to complete the electronic fence early warning.

[0016] Furthermore, in step S1, the zero-crossing rate (ZCR) is used as the time-domain feature value for temporal feature extraction, and the ZCR threshold V is determined through environmental background noise analysis. th Within a set time window, if and only if the amplitude of the vibration signal decreases from less than -V th Travel to a place greater than +V th , or from greater than +V th Travel to less than -V th A valid zero-crossing is recorded as a single occurrence, and the zero-crossing rate ZCR is the number of valid zero-crossings within the time window. Frequency domain feature extraction involves calculating the power spectral density (PSD) of the four sub-bands of the vibration signal. The variance of the PSD value of each sub-band [Δp0, Δp1, Δp2, Δp3] is selected as the frequency domain feature value. The four sub-bands, from low to high, are low frequency band, mid-low frequency band, mid-high frequency band, and high frequency band, respectively. Wavelet domain feature extraction uses two-layer wavelet packet decomposition to decompose the vibration signal into four mutually orthogonal wavelet space nodes that cover the entire frequency band. Then, the wavelet energy values ​​[E0, E1, E2, E3] of the signals corresponding to the four wavelet space nodes are extracted as wavelet domain feature values.

[0017] Furthermore, in step S1, the final 9-dimensional original feature vector group obtained by multi-domain feature fusion extraction is x=[ZCR,Δp0,Δp1,Δp2,Δp3,E0,E1,E2,E3].

[0018] Furthermore, in step S2, the Support Vector Machine (SVM) is a binary classifier, and the training sample set of the intrusion vibration signal is represented as:

[0019] D={(x1,y1),(x2,y2),…,(x n ,y n )},y i ∈{-1,1};

[0020] Where i = 1, 2, 3, ..., n, x i The feature vector set representing the intrusion vibration signal, y i Labels indicating the type of intrusion vibration signal are represented by -1 and 1.

[0021] Furthermore, in step S3, within the LIBSVM software package, the penalty coefficient c and kernel function parameter g are divided into grid spaces within a preset range using a grid search method, and their values ​​are iterated and taken. For each set of penalty coefficient c and kernel function parameter g, the classification accuracy is calculated using the K-CV method, and the set with the highest accuracy is selected as the optimal parameter.

[0022] Furthermore, in step S4, the specific process of determining the intrusion target event based on the behavior category includes the following steps:

[0023] S41: The real-time acquired unknown vibration signal is subjected to multi-domain feature fusion extraction in accordance with the method of step S1 to generate a 9-dimensional feature vector group;

[0024] S42: Normalize the feature vector group using the scaling rules and factors used in the training sample set to make its feature scale consistent with the training sample set, thus completing the format matching.

[0025] S43: Input the normalized test samples into the SVM classification training model established in step S3 and optimized for parameters;

[0026] S44: The SVM classification training model calculates the mapping of test sample features in the feature space through a kernel function, makes a decision based on the optimal classification hyperplane, and outputs the predicted classification label of the test sample.

[0027] S45: Map the behavior category based on the output value of the predicted classification label; when the predicted classification label is -1, determine that the current signal is a non-intrusive background interference behavior, discard the signal and continue to maintain real-time monitoring; when the predicted classification label is 1, determine that the current signal is a suspected intrusion behavior, that is, capture an intrusion target event and intercept the intrusion vibration signal corresponding to the intrusion target event.

[0028] Furthermore, in step S5, the comprehensive attention convolutional neural network CA-Net is based on U-Net and introduces multiple attention mechanism modules to construct a multi-attention mechanism convolutional neural network.

[0029] Furthermore, the multiple attention mechanism modules include four spatial attention modules SA1, SA2, SA3, and SA4, four channel attention modules CA1, CA2, CA3, and CA4, and one scale attention module LA. The four spatial attention modules SA1, SA2, SA3, and SA4 are inserted into the skip connection path between the encoder and the decoder. The four channel attention modules CA1, CA2, CA3, and CA4 are inserted into the four layers of the decoder, respectively. The scale attention module LA is inserted before the output of the integrated attention convolutional neural network CA-Net.

[0030] Furthermore, the spatial attention module receives the feature map upsampled by 2D deconvolution from the previous level of the decoder as the query vector, and receives the feature map output from the corresponding level of the encoder as the key vector. The output of the spatial attention module is concatenated with the upsampled feature map of the decoder in the channel dimension. Within each level of the decoder, the feature map is concatenated and subjected to the first convolution, and then input into the corresponding channel attention module for recalibration of the channel weights. The recalibrated feature map is then input into the second convolution of the current level. The feature maps output from the four levels of the decoder are adjusted to a uniform spatial resolution through bilinear interpolation to obtain four multi-scale feature maps. The four multi-scale feature maps are concatenated and input into the scale attention module to complete the fusion of multi-scale features. The fused features are then judged by the output layer to obtain the final classification result.

[0031] The present invention has the following advantages over the prior art:

[0032] 1. This invention integrates multi-domain features, resulting in more comprehensive feature representation and stronger anti-interference capabilities;

[0033] 2. This invention integrates multiple attention mechanisms, resulting in higher accuracy in fine-grained event classification;

[0034] 3. This invention cascades SVM and CA-Net, balancing real-time performance with low computational cost, and can significantly reduce the false alarm rate in complex environments. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the binary classification principle of SVM in an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram illustrating the implementation principle of LIBSVM in an embodiment of the present invention;

[0037] Figure 3 This is a schematic diagram of the integrated attention convolutional neural network in an embodiment of the present invention;

[0038] Figure 4 This is a flowchart illustrating the electronic fence early warning method based on the DAS system of the present invention. Detailed Implementation

[0039] The embodiments of the present invention are described in detail below. These embodiments are implemented based on the technical solution of the present invention, and provide detailed implementation methods and specific operation processes. However, the scope of protection of the present invention is not limited to the following embodiments.

[0040] Example 1

[0041] By studying the detection and early warning technology of fence intrusion signals using DAS system, we establish a recognition and classification algorithm based on multi-domain feature fusion extraction and support vector machine (SVM), as well as a neural network feature recognition and prediction algorithm based on channel attention mechanism.

[0042] To improve the accuracy of target event classification and enhance robustness in complex environments, this invention uses multi-domain features as feature vectors for classification, including feature extraction in the time domain, frequency domain, and wavelet domain.

[0043] Furthermore, by comprehensively comparing the advantages and disadvantages of various pattern classifiers, Support Vector Machine (SVM) was selected as the pattern classifier used in this invention. Multi-domain feature fusion extraction, which extracts features from intrusion vibration signals from multiple perspectives, effectively improves the distinguishability of various types of intrusion vibration signals, thereby improving event classification accuracy and enhancing robustness in complex environments.

[0044] This invention extracts representative feature values ​​in the time domain, frequency domain, and wavelet domain:

[0045] In the time domain, the zero-crossing rate (ZCR) was selected as the feature value, and a suitable ZCR threshold was determined through environmental noise analysis. Specifically, to eliminate invalid zero-crossing triggering caused by environmental background noise, this invention sets an amplitude threshold V. th (For example, take 1.5 to 2 times the average amplitude of the ambient background noise). Within the set time window, if and only if the amplitude of the vibration signal decreases from less than -V th Travel to a place greater than +V th, or from greater than +V th Travel to less than -V th A zero-crossing is recorded as a valid zero-crossing within a given time window. The zero-crossing rate (ZCR) is the number of valid zero-crossings within that time window.

[0046] In the frequency domain, four sub-bands of the original signal were calculated, specifically divided into: low-frequency band (e.g., 0~50Hz), mid-low-frequency band (e.g., 50~200Hz), mid-high-frequency band (e.g., 200~500Hz), and high-frequency band (e.g., above 500Hz). These four sub-bands correspond to the main frequency band distribution of different intrusion behaviors (e.g., slow climbing corresponds to low frequency, metal shearing corresponds to high frequency). The power spectral density (PSD) of the signal was calculated, and the variance of the PSD value of each sub-band, i.e., [Δp0, Δp1, Δp2, Δp3], was selected as the eigenvalue.

[0047] In the wavelet domain, the wavelet packet decomposition algorithm decomposes not only the low-frequency part of the signal but also the high-frequency part, achieving higher time-frequency resolution compared to traditional wavelet transform. This invention employs two-layer wavelet packet decomposition to decompose the original signal into four mutually orthogonal wavelet spatial nodes covering the entire frequency band (i.e., four wavelet domain sub-bands: [2,0], [2,1], [2,2], [2,3]). Then, the wavelet energy values ​​(i.e., [E0, E1, E2, E3]) corresponding to these four nodes are extracted as feature values; this quantifies the energy distribution characteristics of the intrusion vibration signal in different frequency bands.

[0048] Finally, a 9-dimensional feature vector set x=[ZCR,Δp0,Δp1,Δp2,Δp3,E0,E1,E2,E3] is obtained.

[0049] Support Vector Machines (SVMs) have shown many unique advantages in dealing with small-sample nonlinear classification problems. They can map the feature vectors after feature fusion to a high-dimensional space and achieve accurate classification by dividing a hyperplane with a "maximum margin".

[0050] Support Vector Machine (SVM) is essentially a binary classifier. By fitting multiple binary classifiers together, a multi-classification problem can be solved.

[0051] Suppose the training sample set of the given intrusion vibration signal is D={(x1,y1),(x2,y2),…,(x n ,y n )},y i ∈{-1,1}, i=1,2,3,……,n, where x i This represents the set of all eigenvectors of the intrusion vibration signal, y iThe labels representing the types of intrusive vibration signals are represented by -1 and 1. -1 represents negative samples (i.e., non-intrusive natural or environmental background noise interference such as wind and rain, vehicle passage, etc.), and 1 represents positive samples (i.e., suspected intrusive vibration signals such as people climbing, damaging, or knocking).

[0052] like Figure 1 As shown, the red and blue circles represent two different classes of samples. The goal of the Support Vector Machine (SVM) supervised learning method is to find a hyperplane that can distinguish between these two different classes of samples. (See figure) Let represent the optimal hyperplane that completely separates the two classes of samples. Here, ω is the normal vector of the hyperplane, determining its direction; b is the displacement term, determining the distance between the hyperplane and the origin. (Dashed line) and These represent the boundary planes containing the closest positive and negative class samples to the optimal classification hyperplane, respectively. The sample points located on these two dashed boundaries (i.e., the red and blue circles with crosses) are "support vectors," which play a decisive role in determining the optimal classification hyperplane.

[0053] Once the training sample set for the DAS system is ready, parameter optimization is required:

[0054] This invention employs the LIBSVM software package, which primarily seeks the optimal penalty coefficient *c* and kernel function parameter *g* to ensure the best performance of the trained model. The penalty coefficient *c* (Cost) and kernel function parameter *g* (Gamma) are two key hyperparameters of the Radial Basis Function (RBF) kernel function in SVM. The penalty coefficient *c* represents the tolerance for error; a larger *c* indicates a greater penalty for training error, making overfitting more likely, while a smaller *c* indicates a greater likelihood of underfitting. The kernel function parameter *g* determines the distribution of data after mapping to the new feature space; a larger *g* results in fewer support vectors, and a smaller *g* results in more support vectors.

[0055] A common method for finding the optimal penalty coefficient c and kernel function parameter g is the grid search method. The principle is to divide the penalty coefficient c and kernel function parameter g into a grid space within a certain range. For the penalty coefficient c and kernel function parameter g, all possible values ​​are traversed within the grid. For each set of penalty coefficient c and kernel function parameter g, the K-CV method is used to calculate the validation classification accuracy of the training sample set under that set of penalty coefficient c and kernel function parameter g. Finally, the set of penalty coefficient c and kernel function parameter g with the highest classification accuracy is taken as the final optimal parameters for the training sample set.

[0056] Before inputting the training sample set into the Support Vector Machine (SVM) model, the 9-dimensional feature vector group in the training sample set is first normalized. Specifically, for each feature in the feature vector group x=[ZCR,Δp0,Δp1,Δp2,Δp3,E0,E1,E2,E3], the maximum and minimum values ​​of that feature in the training sample set are calculated, and the min-max normalization method is used to scale each feature to a uniform numerical range. The normalization formula is: x'=(xx min ) / (x max -x min ); where x represents the eigenvalue before normalization, x' represents the eigenvalue after normalization, and x' represents the eigenvalue after normalization. min x represents the minimum value of the corresponding feature in the training sample set. max This represents the maximum value of the corresponding feature in the training sample set. After normalization, the x-values ​​corresponding to each feature in the training sample set are saved. min and x max As a scaling factor, it is used for format matching processing of subsequent test samples. After the DAS system acquires unknown vibration signals in real time, it extracts a 9-dimensional feature vector group in the same way as the training samples, and calls the feature scaling factors of each dimension saved during the training phase to normalize the test samples, so that the test samples and training samples maintain the same feature scale and data format. Subsequently, the normalized test samples are input into the SVM classification training model after parameter optimization for classification prediction. The entire LIBSVM training and prediction process is as follows: Figure 2 As shown.

[0057] First, feature values ​​are extracted from the training samples in the time domain, frequency domain, and wavelet domain to form a feature vector group. Then, the optimal penalty coefficients c and g of the training sample set are obtained by grid search in the LIBSVM software package, thereby establishing a good training model.

[0058] The test samples were processed using the same feature extraction method as the training sample set to obtain a 9-dimensional feature vector set. The normalization scaling rules and scaling factors saved during the training phase were then used for format matching before being fed into the SVM classification training model. Finally, the behavior category of the test samples was determined. The output behavior categories consist of only two classes: background interference behavior (label -1) and suspected intrusion behavior (label 1).

[0059] Finally, the DAS system data is fed into a pre-trained Support Vector Machine (SVM) model for classification according to a predetermined data format. If the SVM prediction result is -1, the system determines it to be environmental interference and discards the signal without further processing. If the prediction result is 1, the system determines that an "intrusion target event" has been captured. Subsequently, the system encodes and maps the original one-dimensional time-series signal corresponding to the "intrusion target event" into a two-dimensional feature map (i.e., two-dimensional information) using time-frequency transformation methods such as Short Time Fourier Transform (STFT) and Gram Angular Field (GAF). This feature map is then used as the input to a Comprehensive Attention Convolutional Neural Network (CA-Net) for the next level of multi-class fine-grained recognition (such as specifically identifying intrusion behaviors like climbing, cutting, and knocking).

[0060] On the other hand, the U-Net network, based on convolutional neural networks (CNNs), has achieved outstanding performance in image detection and segmentation. Compared with traditional methods, the U-Net network exhibits good robustness and performance in the field of two-dimensional signal processing and can automatically learn the most salient features from large datasets. However, when processing with the U-Net network, due to the complexity and diversity of signals, it can still misclassify under certain fault conditions, affecting detection.

[0061] This invention starts with improving the U-Net network to achieve accurate intelligent event recognition. By introducing multiple attention mechanism modules and constructing a dataset, it realizes a multi-attention mechanism convolutional neural network for intelligent event recognition. Furthermore, it utilizes neural network interpretability techniques, represented by Grad-Channel Attention (Grad-CAM), to characterize the network's inference basis and improve the network's interpretability.

[0062] Attention mechanisms are expected to improve the performance of U-Net networks and enhance intelligent event recognition because they focus on the most salient information in the feature map while suppressing irrelevant parts.

[0063] In this invention, a Comprehensive Attention Convolutional Neural Network (CA-Net) is employed. This network widely uses multiple attention mechanisms in CNN architectures, utilizing spatial, channel, and scale attention modules to explain and understand how the network obtains pixel-level prediction results.

[0064] The spatial attention mechanism module in the Comprehensive Attention Convolutional Neural Network (CA-Net) can increase the receptive field of the network and find more potential correlations between pixels; the channel attention mechanism module can weight multiple channels of features, highlighting the information of significant channels and suppressing insignificant channels; the scale attention mechanism module can effectively fuse features of different scales, making it easier to output all the potential effective information in the features of each decoder.

[0065] The U-Net network, improved based on the integrated attention mechanism, can achieve stronger expressive power, ultimately improving the accuracy of intelligent event recognition.

[0066] The integrated attention convolutional neural network (CA-Net) structure used in this invention is as follows: Figure 3 As shown:

[0067] This network uses the robust architecture of U-Net as its backbone, adding specialized convolutional blocks to achieve comprehensive attention guidance on the spatial, channel, and scale of the feature maps.

[0068] The blue rectangles labeled 3×3 or 1×1 and numbers (16, 32, 64, 128, and 256 or categories) correspond to the kernel size and output channels.

[0069] The network contains four spatial attention modules (SA1, SA2, SA3, SA4), four channel attention modules (CA1, CA2, CA3, CA4), and one scale attention module (LA).

[0070] F1, F2, F3, and F4 correspond to resampled versions of feature maps at different scales in the decoder. These feature maps are concatenated and used as input to the scale attention module. By encoding multi-channel data information from the sensing fiber into two-dimensional information and inputting it into the network, intelligent identification and early warning of external intrusion targets can be achieved after training.

[0071] In this embodiment, four spatial attention modules (SA1, SA2, SA3, SA4) are inserted into the skip connection path between the encoder and decoder. Each spatial attention module receives a feature map upsampled by a 2D deconvolution from the layer above the decoder as a query vector and receives the feature map output from the corresponding layer of the encoder as a key vector. The output of the spatial attention module is concatenated with the upsampled feature map from the decoder along the channel dimension (concatenation layer). Four channel attention modules (CA1, CA2, CA3, CA4) are inserted into the four layers of the decoder, respectively. Within each decoder layer, after concatenation and the first convolution (Convolution Layer + BN + ReLU), the feature map is input to the corresponding channel attention module for channel weight recalibration. The recalibrated feature map is then input into the second convolution of that layer. A scale attention module (LA) is inserted before the network output. The feature maps output from the four layers of the decoder are adjusted to a uniform spatial resolution using bilinear interpolation, resulting in four multi-scale feature maps F1, F2, F3, and F4. These four feature maps are concatenated and then input into the scale attention module (LA) to complete the adaptive fusion of multi-scale features. The fused features are then processed by the output layer to obtain the final classification result.

[0072] In this embodiment, the final warning results include: the specific physical location that triggered the alarm (corresponding to the specific fiber optic distance and zone number of the DAS system), the specific type of intrusion behavior (e.g., personnel climbing, mechanical cutting, hard object striking, etc.), and the precise timestamp of the intrusion behavior. Based on the warning results, a pop-up alarm is triggered in the host computer monitoring software, and the corresponding on-site audible and visual alarm devices in the zone can also be activated.

[0073] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. An electronic fence early warning method based on a DAS system, characterized in that, Includes the following steps: S1: Multi-domain feature fusion extraction In the fiber optic electronic fence, the intrusion vibration signal collected by the DAS system is subjected to multi-domain feature fusion extraction in the time domain, frequency domain, and wavelet domain to obtain a 9-dimensional feature vector group; S2: Construction of Training Sample Set The feature vector set is input into the support vector machine (SVM), and signal classification training is completed through hyperplane partitioning to construct a training sample set. S3: Parameter Optimization By employing grid search combined with K-CV cross-validation, the optimal penalty coefficient c and kernel function parameter g for the training sample set are determined in the LIBSVM software package, and an SVM classification training model is established. S4: Intrusion Target Event Determination Input the format-matched test samples into the SVM classification training model to predict the behavior category and determine the intrusion target event; S5: Electronic Fence Warning The intrusion vibration signal corresponding to the intrusion target event is encoded into two-dimensional information through time-frequency conversion, and then input into the comprehensive attention convolutional neural network CA-Net for recognition to complete the electronic fence early warning.

2. The electronic fence early warning method based on a DAS system according to claim 1, characterized in that, In step S1, the zero-crossing rate (ZCR) is used as the time-domain feature value for temporal feature extraction, and the ZCR threshold V is determined through environmental background noise analysis. th Within a set time window, if and only if the amplitude of the vibration signal decreases from less than -V th Travel to a place greater than +V th , or from greater than +V th Travel to less than -V th A valid zero-crossing is recorded as a single occurrence, and the zero-crossing rate ZCR is the number of valid zero-crossings within the time window. Frequency domain feature extraction involves calculating the power spectral density (PSD) of the four sub-bands of the vibration signal. The variance of the PSD value of each sub-band [Δp0, Δp1, Δp2, Δp3] is selected as the frequency domain feature value. The four sub-bands, from low to high, are low frequency band, mid-low frequency band, mid-high frequency band, and high frequency band, respectively. Wavelet domain feature extraction uses two-layer wavelet packet decomposition to decompose the vibration signal into four mutually orthogonal wavelet space nodes that cover the entire frequency band. Then, the wavelet energy values ​​[E0, E1, E2, E3] of the signals corresponding to the four wavelet space nodes are extracted as wavelet domain feature values.

3. The electronic fence early warning method based on a DAS system according to claim 2, characterized in that, In step S1, the final 9-dimensional original feature vector group obtained by multi-domain feature fusion extraction is x=[ZCR,Δp0,Δp1,Δp2,Δp3,E0,E1,E2,E3].

4. The electronic fence early warning method based on a DAS system according to claim 3, characterized in that, In step S2, the Support Vector Machine (SVM) is a binary classifier, and the training sample set of the intrusion vibration signal is represented as follows: D={(x1,y1),(x2,y2),…,(x n ,y n )},y i ∈{-1,1}; Where i = 1, 2, 3, ..., n, x i The feature vector set representing the intrusion vibration signal, y i Labels indicating the type of intrusion vibration signal are represented by -1 and 1.

5. The electronic fence early warning method based on a DAS system according to claim 4, characterized in that, In step S3, within the LIBSVM software package, the penalty coefficient c and kernel function parameter g are divided into grid spaces within a preset range using a grid search method, and their values ​​are iterated and taken. For each set of penalty coefficient c and kernel function parameter g, the classification accuracy is calculated using the K-CV method, and the set with the highest accuracy is selected as the optimal parameter.

6. The electronic fence early warning method based on a DAS system according to claim 1, characterized in that, In step S4, the specific process of determining the intrusion target event based on behavior category includes the following steps: S41: The real-time acquired unknown vibration signal is subjected to multi-domain feature fusion extraction in accordance with the method of step S1 to generate a 9-dimensional feature vector group; S42: Normalize the feature vector group using the scaling rules and factors used in the training sample set to make its feature scale consistent with the training sample set, thus completing the format matching. S43: Input the normalized test samples into the SVM classification training model established in step S3 and optimized for parameters; S44: The SVM classification training model calculates the mapping of test sample features in the feature space through a kernel function, makes a decision based on the optimal classification hyperplane, and outputs the predicted classification label of the test sample. S45: Map the behavior category based on the output value of the predicted classification label; when the predicted classification label is -1, determine that the current signal is a non-intrusive background interference behavior, discard the signal and continue to maintain real-time monitoring; when the predicted classification label is 1, determine that the current signal is a suspected intrusion behavior, that is, capture an intrusion target event and intercept the intrusion vibration signal corresponding to the intrusion target event.

7. The electronic fence early warning method based on a DAS system according to claim 1, characterized in that, In step S5, the comprehensive attention convolutional neural network CA-Net is based on U-Net and introduces multiple attention mechanism modules to construct a multi-attention mechanism convolutional neural network.

8. The electronic fence early warning method based on a DAS system according to claim 7, characterized in that, The multiple attention mechanism modules include four spatial attention modules SA1, SA2, SA3, and SA4, four channel attention modules CA1, CA2, CA3, and CA4, and one scale attention module LA. The four spatial attention modules SA1, SA2, SA3, and SA4 are inserted into the skip connection path between the encoder and the decoder. The four channel attention modules CA1, CA2, CA3, and CA4 are inserted into the four layers of the decoder, respectively. The scale attention module LA is inserted before the output of the integrated attention convolutional neural network CA-Net.

9. The electronic fence early warning method based on a DAS system according to claim 8, characterized in that, The spatial attention module receives the feature map upsampled by 2D deconvolution from the previous level of the decoder as the query vector, and receives the feature map output from the corresponding level of the encoder as the key vector. The output of the spatial attention module is concatenated with the upsampled feature map of the decoder along the channel dimension. Within each level of the decoder, the feature map is concatenated and subjected to the first convolution, and then input into the corresponding channel attention module for recalibration of channel weights. The recalibrated feature map is then input into the second convolution of the current level. The feature maps output from the four levels of the decoder are adjusted to a uniform spatial resolution through bilinear interpolation to obtain four multi-scale feature maps. The four multi-scale feature maps are concatenated and input into the scale attention module to complete the fusion of multi-scale features. The fused features are then judged by the output layer to obtain the final classification result.

Images