Method for identifying abnormal events in perimeter security zones based on fiber optic signals

Abstract

The present application belongs to the field of distributed optical fiber signal vibration event classification, and discloses a perimeter security area abnormal event identification method based on optical fiber signals. The phase-sensitive optical time domain reflectometer is used to collect optical fiber vibration event data sets and perform preprocessing to form a training set. The 1DCNN model is improved. The training sample is input into the improved 1DCNN deep learning model for training to obtain the trained improved 1DCNN model. And the abnormal event identification is performed. The present application improves the identification accuracy of perimeter security area abnormal events, solves the problems of traditional feature extraction methods of optical fiber signals, such as the need for artificial feature generation, serious dependence on expert knowledge, weak generalization ability, and large calculation amount in the identification process of existing models using optical fiber signals.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of distributed optical fiber signal vibration event classification, and particularly relates to a perimeter security area abnormal event identification method based on an optical fiber signal. BACKGROUND

[0002] Nowadays, with the rapid development of the information age, cables, submarine cables, oil and gas pipelines, and train tracks are being laid more and more frequently, and the resulting safety monitoring problems have become a major issue that needs to be addressed. Vibration information is the carrier of various safety hazard events. If the vibration information of the outside world can be monitored in real time, and the characteristics of different events can be distinguished from the vibration signal and then identified, it will certainly play an important role in safety monitoring. For example, identifying whether there are large mechanical devices vibrating near the cable to prevent the cable from being damaged due to construction, identifying whether there are people climbing over the protective net around the railway to achieve the purpose of safety warning, etc.

[0003] Distributed optical fiber vibration sensing (DVS) technology can continuously measure the vibration signals around the sensing optical fiber, and is a kind of distributed optical fiber sensing technology. Compared with traditional sensors, this kind of technology has the characteristics of being distributed, anti-electromagnetic interference, corrosion-resistant, and not requiring power supply. At the same time, the distributed optical fiber vibration sensing system also has the characteristics of high sensitivity, simple structure, and long detection distance. It is very sensitive to external vibrations and has a large monitoring dynamic range, and is very suitable for detecting vibration events in a large area. Its unique application scenarios make it the first choice in the fields of border security, oil and gas pipeline leakage detection and positioning, building structure damage monitoring, and border monitoring in multiple environments. Phase-sensitive optical time domain reflectometry (Φ-OTDR) is a typical representative of distributed optical fiber sensing systems, and is widely used in security and monitoring fields such as national defense borders and military bases. However, high nuisance alarm rate (NAR) is a major drawback of Φ-OTDR systems. Due to the sensitivity of the sensing optical fiber to the environment and the similarity of some intrusion signals and non-intrusion signals, the system is prone to make poor judgments in the vibration signal recognition process, resulting in a high nuisance alarm rate and wasting a lot of manpower and resources.

[0004] Traditional shallow fiber event classification methods require manual extraction of features from the collected vibration signals during feature extraction. These extraction methods are tedious and complex, and determining whether the extracted features can help the classifier to identify requires certain experience and knowledge accumulation. In addition, some of the extracted features may be invalid in other scenarios, resulting in unstable generalization ability of the model. Once the environment changes, the false alarm rate of the system will significantly increase. SUMMARY

[0005] The perimeter security area abnormal event recognition method based on the optical fiber signal can directly extract features from the original optical fiber vibration event data set and identify the type of the vibration event without manual participation, improves the recognition accuracy of the perimeter security area abnormal event, and solves the problems of the traditional feature extraction method of the optical fiber signal, such as the need for artificial feature generation, serious dependence on expert knowledge, weak generalization ability, and large calculation amount in the process of recognition by the existing model using the optical fiber signal.

[0006] To achieve the above object, the technical scheme adopted by the present application is:

[0007] A perimeter security area abnormal event recognition method based on an optical fiber signal, the perimeter security area abnormal event recognition method based on the optical fiber signal comprising:

[0008] Collecting optical fiber vibration event data sets by a phase-sensitive optical time domain reflectometer and preprocessing the optical fiber vibration event data sets to form a training set, wherein the training samples in the training set are the preprocessed optical fiber vibration event data;

[0009] Inputting the training samples into an improved 1DCNN deep learning model, wherein the improved 1DCNN deep learning model comprises four convolution modules, a long short-term memory network, a global mean pooling layer, a full connection layer, and a SVM classifier optimized by a grey wolf optimization algorithm, each convolution module comprises a convolution layer, a batch normalization layer, a Relu activation function layer, and a maximum pooling layer, and an attention mechanism module is connected between the first convolution module and the second convolution module;

[0010] Calculating a first loss function based on the output of the full connection layer in the improved 1DCNN deep learning model and the label of the training sample, and updating the weight parameters of the improved 1DCNN deep learning model before the SVM classifier optimized by the grey wolf optimization algorithm until the training is completed according to the first loss function;

[0011] Fixing the weight parameters of the improved 1DCNN deep learning model before the SVM classifier optimized by the grey wolf optimization algorithm, calculating a second loss function based on the output of the SVM classifier optimized by the grey wolf optimization algorithm and the label of the training sample, and adjusting the parameters of the SVM classifier optimized by the grey wolf optimization algorithm according to the second loss function until the training is completed;

[0012] Taking the improved 1DCNN deep learning model trained, outputting an abnormal event recognition result for the optical fiber vibration event data to be identified.

[0013] The following also provides several optional modes, but not as an additional limitation of the above general scheme, just a further supplement or preferred, without technical or logical contradiction, each optional mode can be combined alone for the above general scheme, but also can be combined between multiple optional modes.

[0014] As preferred, the optical fiber vibration event data contains a plurality of abnormal event categories, each training sample under each abnormal event category is composed of m time domain points and n adjacent space points, that is, each training sample is an m row n column matrix.

[0015] As preferred, the input of the training sample into the improved 1DCNN deep learning model comprises:

[0016] The data points in the m row n column matrix are spliced in order of rows to obtain a spliced signal;

[0017] The spliced signal is input into the improved 1DCNN deep learning model.

[0018] As preferred, the process of optimizing the SVM classifier by using the grey wolf optimization algorithm is as follows:

[0019] The size of the wolf pack and the maximum number of iterations are set, the penalty parameter and the kernel function parameter in the SVM classifier are set as the two-dimensional coordinates of the wolf individual position, and the penalty parameter and the kernel function parameter are randomly initialized.

[0020] The training samples of the training set are input into the SVM classifier, the fitness value is calculated according to the output of the SVM classifier, and the rank of the wolf is classified according to the fitness value, the rank of the wolf includes wolf, wolf, wolf and wolf, and the formula for calculating the fitness value is as follows:

[0021]

[0022] In the formula, is the fitness value, is the number of correct classifications, is the number of incorrect classifications;

[0023] The mathematical model of grey wolf hunting is as follows:

[0024]

[0025]

[0026] In the formula, is the distance between the gray wolf and the prey, is the iteration number, is the current position of the prey, represents the current search position of the wolf, is the updated position of the search wolf, is the first coefficient, is the second coefficient; calculated by the following formula:

[0027]

[0028]

[0029] wherein, is the convergence factor, which decreases linearly from 2 to 0 with the iteration number is updated; and is a random variable, ranging from [0, 1];

[0030] the distance between the wolf and the wolf, the wolf, the wolf is shown as follows:

[0031]

[0032] The step size and direction of the wolf's advance are mathematically modeled as follows:

[0033]

[0034] wherein, , , are the distances between the wolf, wolf, wolf and the wolf, , , are the first coefficients corresponding to the wolf, wolf, wolf, , , are the current positions of the wolf, wolf, wolf, , , are the updated positions of the wolf, wolf, The second coefficient corresponding to the wolf, 、 、 Respectively The wolf is updated Wolf, Wolf, The updated position of the wolf;

[0035] The mathematical formula of the final updated position of the wolf is:

[0036]

[0037] The fitness value of the updated individual is calculated, and the optimal fitness value of the current iteration number is recorded as If , The fitness value of the wolf is , then is updated to , and the optimal position is recorded as the position of the wolf If , The fitness value of the wolf is , then is assigned to Wolf, and the optimal position is recorded as the position of the wolf If , The fitness value of the wolf is , then is assigned to Wolf, and the optimal position is recorded as the position of the wolf ;

[0038] Determine whether the iteration number reaches the maximum iteration number, if it reaches, the loop is terminated, and the best penalty parameter And the best kernel function parameter , and the SVM classifier optimized by the grey wolf optimization algorithm is obtained using the best penalty parameter And the best kernel function parameter If not, return to continue iteration.

[0039] The perimeter security area abnormal event identification method based on optical fiber signal provided by the application has the following beneficial effects compared with the prior art:

[0040] (1) The application inserts BN (batch normalization) between the convolution layer and the activation function layer of the convolution module of the 1DCNN deep learning model. Adding a batch normalization layer can speed up the training of the network and enhance the generalization ability.

[0041] (2) The present invention inserts a CBAM attention mechanism module between the first convolution module and the second convolution module of the 1DCNN deep learning model, assigning different weights to each channel data, solving the problem that the 1DCNN deep learning model regards each channel data as equally important, and enhancing the local feature extraction capability of the 1DCNN deep learning model.

[0042] (3) Since the optical fiber vibration event data is time series data, and the 1DCNN deep learning model cannot capture the time series characteristics of the data, the present invention uses the long short-term memory network LSTM that can capture the time correlation between the data after the 1DCNN deep learning model extracts the spatial characteristics between the data, thereby fully extracting the spatiotemporal characteristics of the optical fiber vibration data, solving the problem that the 1DCNN deep learning model cannot capture the time correlation between the optical fiber data, thereby greatly improving the accuracy of vibration event recognition and solving the problem of high false alarm rate in the current optical fiber vibration measurement perimeter security.

[0043] (4) After the 1DCNN deep learning model extracts features, the traditional method is to directly input them into the fully connected layer, flatten these feature maps, and then use 2 to 3 layers of fully connected layers for dimensionality reduction. However, the amount of fiber vibration data is too large, and the amount of data after the features extracted by the model are flattened is too large, leading to problems such as excessive model parameters, large computational complexity, and overfitting. The present invention uses global mean pooling to replace the first few layers of the fully connected layer, retaining only the last layer, greatly reducing the number of parameters in the network and avoiding the problem of excessive number of parameters caused by the fully connected layer.

[0044] (5) Due to the slow calculation speed of the Softmax classifier, the large influence of noise interference, and the simple normalization operation of the fully connected output data in accordance with the probability distribution, the classification performance of the convolutional neural network cannot be further improved. These problems ultimately lead to less than ideal classification results and are not suitable for optical fiber vibration data sets. Therefore, the present invention uses the SVM classifier optimized by the Gray Wolf Optimization Algorithm to replace the conventional Softmax classifier in 1DCNN to identify vibration events based on the extracted features, further improving the recognition accuracy of vibration events. The Gray Wolf Optimization Algorithm is used to improve the SVM classifier and find the optimal parameter combination of the SVM classifier, which improves the classification performance compared with the conventional SVM classifier.

[0045] (6) The method of the present invention is an end-to-end recognition method. The entire process does not require human intervention. It can directly extract features from the original optical fiber signal and identify the type of vibration event. It solves the problems of traditional feature extraction methods that require artificial feature creation, are heavily dependent on expert knowledge, and have weak generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1A flow chart of a perimeter security area abnormal event identification method based on an optical fiber signal according to the present application;

[0047] Figure 2 A training flow chart of an improved 1DCNN deep learning model according to the present application;

[0048] Figure 3 A feature visualization effect diagram of a first experimental object extracted in an experiment according to the present application;

[0049] Figure 4 A feature visualization effect diagram of a second experimental object extracted in an experiment according to the present application;

[0050] Figure 5 A feature visualization effect diagram of a third experimental object extracted in an experiment according to the present application;

[0051] Figure 6 A confusion matrix diagram of a 1DCNN-Softmax() model for optical fiber signal classification effect in an experiment according to the present application;

[0052] Figure 7 A confusion matrix diagram of a 1DCNN-CBAM-Softmax() model for optical fiber signal classification effect in an experiment according to the present application;

[0053] Figure 8 A confusion matrix diagram of a 1DCNN-CBAM-LSTM-Softmax() model for optical fiber signal classification effect in an experiment according to the present application;

[0054] Figure 9 A confusion matrix diagram of a 1DCNN-CBAM-LSTM-GWO-SVM model for optical fiber signal classification effect in an experiment according to the present application. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the scope of the present application.

[0056] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terminology used in the description of the present application herein only for the purpose of describing specific embodiments and is not intended to limit the present application.

[0057] As Figure 1As shown, this embodiment provides a method for identifying abnormal events in a perimeter security zone based on optical fiber signals, comprising the following steps:

[0058] First, we set up a deep learning model computing environment. We used a computer as a server, built a fiber vibration event database on the server, and created a data table to store the data. We then prepared a local Windows workstation, downloaded the data from the server, and used Python to build a deep learning model and classifier based on the data's dimensions and length.

[0059] Step 1: Collect a fiber vibration event dataset using a Φ-OTDR (Phase-Sensitive Optical Time Domain Reflectometer) distributed fiber optic sensing system, and use a preprocessing method corresponding to the fiber vibration event dataset collected by the Φ-OTDR distributed fiber optic sensing system.

[0060] Specifically, fiber optic vibration event data was collected, including six typical events: background noise, digging, knocking, watering, shaking fences, and walking. In the data set, each sample of each event consists of 10,000 time domain points and 12 adjacent spatial points. Therefore, the data format of each vibration event is a matrix with 10,000 rows and 12 columns. Due to the variation in signal amplitude, the signal is normalized to frame the data within a certain value range. Its main purpose is to eliminate the dimensional influence between the various evaluation indicators and the negative impact caused by singular sample data. The formula is as follows:

[0061]

[0062] in, Indicates that each dimension data is linearly normalized. Represents the original data of each dimension data, Indicates the maximum value in each dimension data, It represents the minimum value in each dimension data. After linear normalization, the data of each dimension is mapped to the unified metric space of [0,1].

[0063] Step 2. Input the training samples into the improved 1DCNN deep learning model. The improved 1DCNN deep learning model includes a four-layer convolution module, a long short-term memory network, a global mean pooling layer, a fully connected layer, and an SVM classifier optimized by the gray wolf optimization algorithm. Each layer of the convolution module includes a convolution layer, a batch normalization layer, a Relu activation function layer, and a maximum pooling layer, and an attention mechanism module is connected between the first layer of the convolution module and the second layer of the convolution module.

[0064] like Figure 2As shown in the figure, the normalized signals of the 12 nodes are first connected and input into the 1DCNN network to extract local features of all sensing points and perform dimensionality reduction. The improved 1DCNN consists of four convolutional modules, each of which includes a convolutional layer, a batch normalization layer, a maximum pooling layer, and a ReLU activation function layer. The convolutional layer is used to extract local features in the vibration signal and the characteristic structure hidden in the data. The formula is as follows:

[0065]

[0066] in, For the layer convolutional layer; For the The output of the convolutional layer, is the activation function; For the The convolutional layer corresponds to the The input of the convolution kernel; For the The convolutional layer The weight of the convolution kernel; For the The convolutional layer The bias of the convolution kernel; is the convolution operation; is the total number of convolution kernels.

[0067] The Batch Normalization layer is an optimization method to address the internal covariate shift caused by input transformation in convolutional neural networks. By calculating the variance and standard deviation on a small batch training set, the output distribution of each layer is standardized, which speeds up the network training and enhances the generalization ability. The calculation process of batch normalization is as follows:

[0068]

[0069]

[0070]

[0071]

[0072] in, represents the size of the mini-batch, and They are respectively The input and output of observations, represents the mean of the mini-batch samples, is the standard deviation of the mini-batch samples. is a constant close to zero to ensure numerical stability, denotes the data after regularization, denotes the data after scaling and translation, denotes the data after scaling and translation, and are parameters, used to adjust the variance of the value distribution. At the beginning of model training, set , and then gradually adjust to the appropriate value through the learning process.

[0073] The activation layer is to process the operation results of the convolution layer, so as to obtain the corresponding output results. It is usually nonlinear, which can make the nonlinear relationship between the sample number and the label be fitted in the neural network. The Relu activation function maps negative numbers to 0 and keeps positive numbers unchanged, which can effectively reduce the problem of gradient disappearance, reduce the training difficulty of the network, and accelerate the convergence speed. The formula is as follows:

[0074]

[0075] In the formula, denotes the output of the Relu activation function, max denotes the maximum of two numbers, and denotes the input of the Relu activation function.

[0076] The pooling layer is also called the down-sampling layer, which can play the role of information screening. Through the sliding window, the feature map is processed by pooling, which reduces the dimension of the feature map and the number of parameters while retaining significant feature information. The maximum pooling layer extracts the maximum value of the data in the local region, and the formula is as follows:

[0077]

[0078] In the formula, is the th pooling result of the th feature matrix of the th layer; is the th feature matrix of the th layer; is the region processed by the pooling kernel; is the number of pooling operations; is the width of the pooling kernel.

[0079] The Convolutional Attention Module (CBAM) attention mechanism is inserted between convolutional modules. While 1DCNN assumes that each channel feature is equally important during convolution and pooling, the information carried by each channel feature varies in importance in reality, making it unreasonable to simply assume that all channel features are equal. The Convolutional Attention Module (CBAM) improves information flow within the network by understanding which information to emphasize or suppress, identifying more important features and focusing attention on these key areas to improve classification accuracy. By assigning different weights to information from different channels, it emphasizes important features and suppresses unnecessary ones.

[0080] Fiber vibration data is time series data, and 1DCNN cannot learn the time-related characteristics of the signal. The spatial local features extracted by 1DCNN are input into the long short-term memory network (LSTM), which extracts time-related features from it. The calculation formula is as follows:

[0081]

[0082] in, for The importance of the gate result at each moment; Calculate the sigmoid function; is the weight matrix of the input gate; is the offset matrix of the input gate; for The degree of retention of the results of the moment-forget gate; is the weight matrix of the forget gate; is the offset matrix of the forget gate; Selecting information for the memory cell's history; is the weight matrix of the memory cell; is the offset matrix of the memory cell; for Memory cells of the moment; for Memory cells of the moment; for The degree of information selection of the output gate results at each moment; is the weight matrix of the output gate; is the offset matrix of the output gate; for Output at each moment; for Output at the moment; is the tanh activation function, express The local spatial features of the input at each moment.

[0083] The application uses global mean pooling to replace the first several layers of the full connection layer in the original model, only retains the last full connection layer, inputs the space-time features extracted by the long short-term memory network (LSTM) into the global mean pooling layer, and the global mean pooling layer is used to solve too many parameters of the full connection network, and the formula is as follows:

[0084]

[0085] In the formula, is information after global mean pooling operation of the i-th channel in the output feature map; represents the i-th feature value in the j-th channel; is the dimension of the feature vector.

[0086] Further, the embodiment introduces an SVM classifier, and the SVM classifier is optimized by using a grey wolf optimization algorithm (GWO), so that the SVM classifier replaces a traditional Softmax classifier in the 1DCNN model, and features extracted by the deep learning model are input into the optimized SVM classifier for event recognition.

[0087] Specifically, first, a support vector machine (SVM) classifier is constructed, an input sample set is mapped to a high-dimensional feature space, and an optimal hyperplane that can be regarded as a decision boundary is constructed in the space, the optimal hyperplane is used to effectively separate the sample set and find support vectors on both sides of the hyperplane, the support vectors are sample points closest to the hyperplane, and the support vectors have good generalization ability and robustness, and the expression of the input sample set is as follows:

[0088]

[0089] In the formula, is the input sample set, is the i-th input variable; is the i-th output variable; is the vector dimension, indicates the number of samples in the sample set. The optimal hyperplane for division can be expressed as:

[0090]

[0091] In the formula, is the normal vector of the optimal hyperplane,

[0092] is the distance between the hyperplane and the origin, and and the following problem needs to be solved: ​​​​​

[0093]

[0094] wherein, is a penalty factor; is an insensitive factor; and is a relaxation factor.

[0095] To solve some sample sets that cannot be divided in linear space, the samples are mapped from the input space to a higher dimensional feature space, so that the samples are linearly separable in the new space, and the kernel function is introduced and the high-dimensional space classification hyperplane is solved by Lagrange equation:

[0096]

[0097] wherein, is a Lagrange multiplier, represents the number of support vectors, is the th data point, is the th data point.

[0098] The Gaussian radial basis kernel function is selected as the kernel function, and the formula is as follows:

[0099]

[0100] In the formula, is the influence factor of the training sample, that is, the kernel function parameter.

[0101] Then the features output by the full connection layer of the improved 1DCNN deep learning model are input into the constructed SVM classifier, and the grey wolf optimization algorithm GWO is used to optimize the SVM classifier to find the optimal parameters of the SVM classifier. The specific steps include:

[0102] Step A: first, normalize the extracted features, set the wolf population size and the maximum number of iterations, and set the parameters and in the SVM classifier as the two-dimensional coordinates of the wolf individual position, and randomly initialize and ;

[0103] Step B: train the SVM classifier with the training set samples to calculate the fitness value: classify the wolves according to the fitness value (wolf, wolf, wolf, wolf and wolf), and the formula for calculating the fitness value is as follows:

[0104]

[0105] wherein, is the fitness value, is the number of correct classifications, is the number of incorrect classifications.

[0106] The mathematical model of the gray wolf hunting is as follows:

[0107]

[0108]

[0109] wherein, is the distance between the gray wolf and the prey, is the current iteration number, is the current position of the prey, represents the current search wolf position, is the updated position of the search wolf, is the first coefficient, is the second coefficient. It can be calculated by the following formula:

[0110]

[0111]

[0112] wherein, is the convergence factor, which is updated from 2 linearly to 0 as the iteration number increases; and are random variables, ranging from [0, 1].

[0113] The distance between the wolf and the wolf, the wolf, the wolf is as follows:

[0114]

[0115] The mathematical model of the step size and direction of the wolf is as follows:

[0116]

[0117] wherein, , , are the positions of the wolf, the wolf, the wolf, respectively; , , respectively wolf, wolf, wolf and distance of wolf.

[0118] The mathematical formula for wolf to update the position is:

[0119]

[0120] Step C: Calculate the fitness value of the updated individual, and record the optimal fitness value of the current generation as If (wolf's fitness), then is updated to , and the corresponding position is recorded. If , then is assigned to wolf, and the corresponding position is updated to wolf's position. If , then is assigned to wolf, and the corresponding position is also updated to wolf's position. Obviously, when the iteration reaches the current generation, wolf's position is the optimal position of the population.

[0121] Step D: Determine whether the number of iterations has reached the maximum number of iterations. If it has, the loop is terminated, and the best parameters and are obtained; if not, return to step B to continue iteration.

[0122] Step E: Use the best parameters and to establish the optimized SVM classifier.

[0123] Step 3, based on the output of the full connection layer in the improved 1DCNN deep learning model and the label of the training sample, calculate the first loss function, and update the weight parameters in the improved 1DCNN deep learning model before the SVM classifier optimized by the grey wolf optimization algorithm, repeat the calculation of the first loss function based on the training sample and the update until the training is completed. Wherein the first loss function takes the cross-entropy loss function.

[0124] ​Step 4, fix the weight parameters before the SVM classifier optimized by the grey wolf optimization algorithm in the improved 1DCNN deep learning model, calculate the second loss function according to the output of the SVM classifier optimized by the grey wolf optimization algorithm and the label of the training sample, and adjust the parameters of the SVM classifier optimized by the grey wolf optimization algorithm according to the second loss function, and repeat the calculation of the second loss function based on the training sample until the training is completed. The second loss function is the loss function provided by the SVM classifier.

[0125] Step 5, take the improved 1DCNN deep learning model trained, and output the abnormal event recognition result for the optical fiber vibration event data to be identified. In the identification process, the data of the time domain points and the space points are also collected and normalized, and the data of all space points are connected and input into the improved 1DCNN deep learning model to obtain the abnormal event recognition result.

[0126] In order to verify the classification and identification performance of the model, the experiment is carried out on an 8GB memory, Inter(R)Core(TM)i7-14700HX, RTX4070GPU computer, the model is trained on python3.9, and the authoritative data set is used for verification.

[0127] The experimental data has a total of 15612 samples, including background noise, digging, knocking, watering, shaking fence, walking 6 typical events, wherein the background noise sample amount is 3094, numbered as event 0; the digging sample amount is 2512, numbered as event 1; the knocking sample amount is 2530, numbered as event 2; the watering sample amount is 2298, numbered as event 3; the shaking fence sample amount is 2728, numbered as event 4; the walking sample amount is 2450, numbered as event 5. Randomly extract 80% as a training set and 20% as a test set. The data is input into the model to evaluate the performance of the model.

[0128] The classification effect depends largely on the feature extraction. In order to prove the improvement of the feature extraction ability of the 1DCNN model by adding the CBAM attention mechanism module and the long short-term memory network LSTM, the feature distinguishability is compared. The final feature vector obtained by the trained model is mapped to a three-dimensional space vector by linear discriminant analysis (Linear Discriminant Analysis, LDA) algorithm. As shown in the following table, the feature vectors of the 1DCNN model without the CBAM attention mechanism module and the long short-term memory network LSTM are not well distinguishable, and the feature vectors of the 1DCNN model with the CBAM attention mechanism module and the long short-term memory network LSTM are well distinguishable. Figures 3-5As shown, each color corresponds to a vibration event. The stronger the discriminability of the model feature extraction, the stronger the discreteness between events. It can be seen that 1DCNN-CBAM (denoted as the second experimental object, which is connected with the CBAM attention mechanism module between the first layer convolution module and the second layer convolution module on the basis of the first experimental object) has stronger feature extraction capability than 1DCNN (denoted as the first experimental object, which is part of the present application including four convolution modules, a global mean pooling layer and a fully connected layer, and the first layer convolution module and the second layer convolution module are not connected with the attention mechanism module). 1DCNN-CBAM-LSTM (denoted as the third experimental object, which increases the long short-term memory network LSTM between the last layer convolution module and the global mean pooling layer on the basis of the second experimental object) further improves the feature extraction capability due to the addition of LSTM, which captures the time correlation between features.

[0129] A confusion matrix is an error matrix that is often used to visually assess the performance of a classification model. The main diagonal line represents the number of correctly classified samples, while the other cells represent the number of incorrectly classified samples. The darker the color, the larger the number of samples. The confusion matrix obtained after inputting the data into the model is as follows Figures 6-9As shown, the accuracy, precision, recall, and F1 score evaluation criteria of each model are calculated according to the confusion matrix, as shown in Tables 1-4. From the confusion matrix chart and the evaluation criteria table, it can be seen that different models have different performances in identifying fiber vibration events. The 1DCNN-Softmax() model (as the first experimental object combined with a Softmax classifier) has low precision on events 1 and 2 and event 5, low recall on events 0 and 3, and an accuracy of only 95%, which may be due to its insufficient ability to extract local features. The simple 1DCNN is difficult to undertake the fiber vibration event classification task. The addition of the CBAM mixed attention mechanism module enables the model to emphasize important features and suppress unnecessary feature information, enhancing the feature extraction capability of the 1DCNN, and the evaluation indicators of each event are significantly improved, but the precision of event 5 and the recall of event 3 are still not high, which may be that the 1DCNN-CBAM-Softmax() model (as the second experimental object combined with a Softmax classifier) cannot learn the time-dependent characteristics in the signal. After adding the LSTM neural network, the 1DCNN-CBAM-LSTM-Softmax() model (as the third experimental object combined with a Softmax classifier) fully extracts the local and temporal features of the signal, and the evaluation indicators of each event exceed 98% except for the recall and F1 score of event 5, and the final accuracy reaches 98.7%, indicating that the model has excellent feature extraction capability. In the last two groups of experiments, the classification effects of the Softmax() classifier and the GWO-SVM classifier on the fiber vibration events are compared, and it can be seen that the GWO-SVM classifier has improved in almost every evaluation indicator compared to the Softmax() classifier, especially the recall of event 5, which has increased from 96.7% to 99.0%, and the evaluation indicators of each event all exceed 98%, with an accuracy of 99.3%, indicating that the GWO-SVM classifier achieves better classification effect than the Softmax classifier and is more suitable for the classification and identification of fiber vibration events. Among them, the 1DCNN-CBAM-LSTM-GWO-SVM model is an improved 1DCNN deep learning model of the application.

[0130] Table 1 Performance indicators of 1DCNN-Softmax() model

[0131] Event type Precision Recall F1 score Accuracy 0 0.973 0.931 0.951 0.950 1 0.933 0.950 0.942 0.950 2 0.913 0.996 0.953 0.950 3 0.981 0.898 0.938 0.950 4 0.961 0.982 0.971 0.950 5 0.939 0.939 0.939 0.950

[0132] Table 2 Performance indicators of 1DCNN-CBAM-Softmax() model

[0133] Event type Precision Recall F1 score Accuracy 0 0.985 0.981 0.983 0.966 1 0.966 0.962 0.964 0.966 2 0.955 1 0.977 0.966 3 0.980 0.917 0.947 0.966 4 0.967 0.978 0.973 0.966 5 0.943 0.953 0.948 0.966

[0134] Table 3 Performance indicators of 1DCNN-CBAM-LSTM-Softmax() model

[0135] Event type Precision Recall F1 score Accuracy 0 0.994 0.989 0.991 0.987 1 0.982 0.984 0.983 0.987 2 0.986 1 0.993 0.987 3 0.989 0.985 0.987 0.987 4 0.985 0.993 0.989 0.987 5 0.981 0.967 0.974 0.987

[0136] Table 4 Performance indicators of 1DCNN-CBAM-LSTM-GWO-SVM model

[0137] Event type Precision Recall F1 score Accuracy 0 0.997 0.997 0.997 0.993 1 0.994 0.986 0.990 0.993 2 0.996 0.998 0.997 0.993 3 0.989 0.996 0.993 0.993 4 0.996 0.993 0.994 0.993 5 0.986 0.990 0.988 0.993

[0138] Wherein the accuracy is the total accuracy of the model for 6 types of events.

[0139] The technical features of the above-described embodiments can be combined in any manner. To make the description concise, all possible combinations of the technical features in the above-described embodiments are not described, but as long as the combinations of the technical features do not contradict, they should be considered within the scope of the present disclosure.

[0140] The above-described embodiments only express several implementation manners of the present application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the scope of the present application. It should be pointed out that, for those skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are all within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A method for identifying abnormal events in perimeter security zones based on optical fiber signals, characterized in that: The method for identifying abnormal events in a perimeter security zone based on optical fiber signals includes: A fiber optic vibration event data set is collected by phase-sensitive optical time domain reflectometry and preprocessed to form a training set, wherein the training samples in the training set are the preprocessed fiber optic vibration event data; Input the training samples into the improved 1DCNN deep learning model, wherein the improved 1DCNN deep learning model includes a four-layer convolution module, a long short-term memory network, a global mean pooling layer, a fully connected layer, and an SVM classifier optimized by the gray wolf optimization algorithm, wherein each layer of the convolution module includes a convolution layer, a batch normalization layer, a Relu activation function layer, and a maximum pooling layer, and an attention mechanism module is connected between the first layer of the convolution module and the second layer of the convolution module; Calculate a first loss function based on the output of the fully connected layer in the improved 1DCNN deep learning model and the labels of the training samples, and update the weight parameters of the improved 1DCNN deep learning model before the SVM classifier optimized by the gray wolf optimization algorithm according to the first loss function until the training is completed; Fixing the weight parameters before the SVM classifier optimized by the Gray Wolf Optimization Algorithm in the improved 1DCNN deep learning model, calculating the second loss function based on the output of the SVM classifier optimized by the Gray Wolf Optimization Algorithm and the labels of the training samples, and adjusting the parameters of the SVM classifier optimized by the Gray Wolf Optimization Algorithm based on the second loss function until the training is completed; Take the trained improved 1DCNN deep learning model and output the abnormal event recognition results for the optical fiber vibration event data to be identified.

2. The method for identifying abnormal events in perimeter security zones based on optical fiber signals according to claim 1, characterized in that: The optical fiber vibration event data includes multiple abnormal event categories. Each training sample under each abnormal event category consists of m time domain points and n adjacent space points, that is, each training sample is a matrix with m rows and n columns.

3. The method for identifying abnormal events in perimeter security zones based on optical fiber signals according to claim 2, characterized in that: The step of inputting the training samples into the improved 1DCNN deep learning model includes: The data points in the matrix of m rows and n columns are spliced ​​row by row to obtain a spliced ​​signal; The spliced ​​signal is input into the improved 1DCNN deep learning model.

4. The method for identifying abnormal events in perimeter security zones based on optical fiber signals according to claim 1, characterized in that: The process of optimizing the SVM classifier using the Gray Wolf Optimization Algorithm is as follows: Set the wolf pack size and the maximum number of iterations, and set the penalty parameter in the SVM classifier and kernel function parameters Set as the two-dimensional coordinates of the individual positions of the wolf pack, and randomly initialize the penalty parameters and kernel function parameters ; The training samples of the training set are input into the SVM classifier, the fitness value is calculated according to the output of the SVM classifier, and the wolf levels are classified according to the fitness value. The wolf levels include Wolf, Wolf, Wolf and Wolf, the formula for calculating the fitness value is as follows: Where, is the fitness value, is the number of correct classifications, is the number of misclassifications; The mathematical model of gray wolf hunting is as follows: Where, is the distance between the gray wolf and its prey, is the number of iterations, is the current location of the prey, Indicates the current location of the wolf being searched. To search for the updated location of the wolf, is the first coefficient, is the second coefficient; it is calculated using the following formula: Where, is the convergence factor, as the number of iterations Update, linearly decrease from 2 to 0; and is a random variable with a range of [0,1]; Wolf and Wolf, Wolf, The distances between wolves are shown below: The mathematical model of the wolf's stride length and direction is as follows: Where, 、 、 They are Wolf, Wolf, Wolf and The distance of the wolf, 、 、 They are Wolf, Wolf, The first coefficient corresponding to the wolf, 、 、 They are Wolf, Wolf, The wolf's current location, 、 、 For Wolf, Wolf, The second coefficient corresponding to the wolf, 、 、 They are Wolf Target Wolf, Wolf, The updated position of the wolf; The mathematical formula for the wolf's final updated position is: Calculate the fitness value of the updated individual and record the optimal fitness value of the current iteration as ,if , for The wolf's fitness value is Updated to , and record the optimal position as The wolf's position; if , for The wolf's fitness value will be Assign to Wolf, and record the optimal position as The wolf's position; if , for The wolf's fitness value will be Assign to Wolf, and record the optimal position as The location of the wolf; Determine whether the number of iterations reaches the maximum number of iterations. If so, the loop terminates and the optimal penalty parameter is obtained. and the optimal kernel function parameters , and use the optimal penalty parameter and the optimal kernel function parameters Get the SVM classifier optimized by the gray wolf optimization algorithm; if it is not reached, return to continue iteration.

Images