Perimeter security intrusion event recognition method

Abstract

A perimeter security intrusion event recognition method, relating to the technical field of perimeter security intrusion event monitoring. The method comprises the following steps: S1, collecting a background noise vibration signal and an intrusion event vibration signal, and acquiring a model training set, a model test set, and a model verification set; S2, constructing a perimeter security intrusion event recognition network; S3, training the perimeter security intrusion event recognition network to obtain a perimeter security intrusion event recognition network model; and S4, processing the collected intrusion event vibration signal by using step S1, inputting the processed intrusion event vibration signal into the perimeter security intrusion event recognition network model, and outputting predicted background noise and the classification accuracy of three intrusion events.

Description

A perimeter security intrusion event identification method TECHNICAL FIELD

[0001] The present application relates to the field of perimeter security intrusion event monitoring, and particularly relates to a perimeter security intrusion event identification method. BACKGROUND

[0002] With the development of the state and social economy, the demand for safety in various industries is increasingly strong, and security technology is increasingly concerned by researchers. In order to improve the identification and classification accuracy of perimeter security intrusion events, researchers mainly focus on three types of identification methods, including machine learning-based intrusion event identification method, artificial neural network-based intrusion event identification method and deep learning-based intrusion event identification method. Among them, the deep learning-based intrusion event identification method mainly uses convolutional neural network and long short-term memory network for intrusion event identification. Although the deep learning-based intrusion event identification method can effectively overcome the shortcomings of the machine learning-based intrusion event identification method and improve the learning ability of the neural network compared with the machine learning-based intrusion event identification method, the classification accuracy and running time of the deep learning-based intrusion event identification method in the prior art are still not ideal. Therefore, the present application provides a perimeter security intrusion event identification method. SUMMARY

[0003] The present application provides a perimeter security intrusion event identification method to make up for the shortcomings of the prior art.

[0004] The application is implemented through the following technical scheme: A perimeter security intrusion event recognition method, comprising the following steps: S1, collecting background noise vibration signals and intrusion event vibration signals, processing the collected background noise vibration signals and intrusion event vibration signals to obtain time domain vibration signal data and two-dimensional image data, and dividing the time domain vibration signal data and two-dimensional image data after normalization processing to obtain a model training set, a model test set and a model validation set; S2, constructing a perimeter security intrusion event recognition network; the perimeter security intrusion event recognition network comprises a 1DCNN network, an At-DenseNet network, a feature fusion module connected with the 1DCNN network and the At-DenseNet network, and a Softmax classification layer connected with the feature fusion module; the At-DenseNet network in the application is obtained by improving an existing DenseNet121 network; wherein the 1DCNN network is used to extract vibration signal time evolution information of one-dimensional sample data, the At-DenseNet network is used to extract time and frequency change rules of two-dimensional sample data, the feature fusion module is used to fuse the vibration signal time evolution information of one-dimensional sample data extracted by the 1DCNN network and the time and frequency change rules of two-dimensional sample data extracted by the At-DenseNet network to obtain a fusion feature map containing vibration signal time evolution information and time and frequency change rule information, and the Softmax classification layer is used to map the feature vector of the fusion feature map output by the feature fusion module to the [0, 1] interval and output the classification accuracy of the predicted background noise and three intrusion events; S3, training the perimeter security intrusion event recognition network by using the training set and a multi-classification cross-entropy loss function, and verifying the perimeter security intrusion event recognition network by using the validation set to obtain a perimeter security intrusion event recognition network model; S4, processing the collected intrusion event vibration signals by using the steps S1-1 to S1-4 to obtain one-dimensional sample data and two-dimensional sample data; then inputting the obtained one-dimensional sample data and two-dimensional sample data into the perimeter security intrusion event recognition network model, performing forward propagation once, and outputting the classification accuracy of the predicted background noise and three intrusion events.

[0005] Preferably, step S1 specifically comprises the following steps: S1-1, collecting background noise vibration signals and intrusion event vibration signals of the laid optical fiber area by using a distributed optical fiber vibration sensing system (DVS); wherein the background noise vibration signals refer to vibration signals of background noise generated by the natural environment without intrusion events, the types of intrusion events include artificial excavation events, personnel walking events and throwing stone events, the intrusion event vibration signals refer to vibration signals generated by artificial excavation events, personnel walking events and throwing stone events, the distributed optical fiber vibration sensing system used in the present application is a distributed optical fiber vibration sensing DVS integrated module sold by Guilin Guangyi Intelligent Technology Co., Ltd., and the access address is: https: / / b2b.baidu.com / slist / 7e1f7b750374702b06027d627119067c047aadfbab a; S1-2, performing data segmentation processing on the background noise vibration signals and the intrusion event vibration signals collected in step S1-1, the unit time of the data segmentation processing is 1s, and the data obtained within 1s after the data segmentation processing is a time domain vibration signal data, which is one-dimensional data; S1-3, converting the time domain vibration signal data into two-dimensional image data by using an existing data conversion method; S1-4, performing normalization processing on the time domain vibration signal data obtained in step S1-2 to obtain one-dimensional sample data; performing normalization processing on the two-dimensional image data obtained in step S1-3 to obtain two-dimensional sample data; S1-5, randomly dividing the one-dimensional sample data obtained in step S1-4 in a quantity ratio of 5:3:2 to obtain a sub-training set, a sub-validation set and a sub-test set of the one-dimensional sample data; randomly dividing the two-dimensional sample data obtained in step S1-4 in a quantity ratio of 5:3:2 to obtain a sub-training set, a sub-validation set and a sub-test set of the two-dimensional sample data; the sub-training set of the one-dimensional sample data and the sub-training set of the two-dimensional sample data constitute a model training set, the sub-validation set of the one-dimensional sample data and the sub-validation set of the two-dimensional sample data constitute a model validation set, and the sub-test set of the one-dimensional sample data and the sub-test set of the two-dimensional sample data constitute a model test set.

[0006] Preferably, in step S1-3, the existing data conversion method is an existing time-frequency analysis method and / or an existing image encoding method, the time-frequency analysis method is one or more of a short-time Fourier transform method, a continuous wavelet transform method and a Hilbert-Huang transform method; the image encoding method is one or more of a Gram angle field method, a Markov transition field method and a recurrence plot method; wherein the Gram angle field method refers to converting time-domain vibration signal data into two-dimensional image data through Gram angle field transformation, the Markov transition field method refers to converting time-domain vibration signal data into two-dimensional image data through Markov transition field transformation, and the recurrence plot method refers to converting time-domain vibration signal data into two-dimensional image data by using a recurrence plot technique; Preferably, in step S2, the 1DCNN network comprises three convolution groups connected in sequence, the third convolution group is connected with a Flatten flattening layer, the Flatten flattening layer is connected with five fully connected layers in sequence, and the fifth fully connected layer is connected with a feature fusion module; wherein the three convolution groups are all used for extracting vibration signal time evolution information in one-dimensional sample data, the Flatten flattening layer is used for flattening operation on the feature map with vibration signal time evolution information output by the third convolution group to obtain a feature map of high-dimensional one-dimensional data with vibration signal time evolution information, and the five fully connected layers are all used for integrating and mapping features in the input feature map to spaces of different dimensions, and the first fully connected layer to the fifth fully connected layer respectively map the feature map to spaces with dimension sizes of 128, 64, 32, 16 and 8.

[0007] Preferably, in step S2, the two one-dimensional convolution layers of the first convolution group each have 64 convolution kernels with a size of 3, the two one-dimensional convolution layers in the second convolution group each have 128 convolution kernels with a size of 3, and the two one-dimensional convolution layers in the third convolution group each have 256 convolution kernels with a size of 3.

[0008] Preferably, in step S2, the At-DenseNet network in the application is obtained by improving the existing DenseNet121 network as follows: the initial convolution group in the existing DenseNet121 network is replaced by the initial convolution block in the application, the initial convolution block in the application includes three convolution units connected in turn, and the convolution unit includes a 3x3 convolution layer, a batch normalization layer and a 2D maximum pooling layer connected in turn; the above setting of the application effectively reduces the complexity of the model and improves the running speed, and since the convolution kernel size of the three convolution units is 3x3, the initial convolution block in the application can better capture the local features in the input two-dimensional sample data, thereby extracting more detailed information of the frequency information; the dense convolution block in the DenceBlock in the existing DenseNet121 network is replaced by the dense convolution group in the application, the dense convolution group in the application includes a batch normalization layer, a convolution kernel size of 3x1 convolution layer and a convolution kernel size of 1x3 convolution layer connected in turn; the number of dense convolution groups is set to six, and the bottleneck block and the six dense convolution groups are connected in a dense connection manner to obtain the first dense module; the number of dense convolution groups is set to eight, and the bottleneck block and the eight dense convolution groups are connected in a dense connection manner to obtain the second dense module, the third dense module and the fourth dense module, respectively; the structure of the bottleneck block in the application is the same as that of the bottleneck block in the DenceBlock in the existing DenseNet121 network, and the functions are the same. The batch normalization layer in the dense convolution block in the application is used to alleviate overfitting, the convolution kernel size of 3x1 convolution layer is used for feature extraction operation on the input feature map, and the convolution kernel size of 1x3 convolution layer is used for feature extraction operation on the feature map output by the convolution kernel size of 3x1 convolution layer; the convolution kernel size of 3x1 convolution layer and the convolution kernel size of 1x3 convolution layer in the dense convolution group in the application are used for convolution operation in turn, which is an asymmetric convolution design. Since the convolution kernel size of 3x1 convolution layer and the convolution kernel size of 1x3 convolution layer can capture the features in different directions of the feature map, respectively, the dense convolution block designed in the application can extract more features of the time and frequency variation law in the two-dimensional sample data. In addition, the setting of the convolution kernel size of 3x1 convolution layer and the convolution kernel size of 1x3 convolution layer can also effectively reduce the calculation amount due to the reduction of parameters.

[0009] A channel attention module is added after each transition layer in the existing DenseNet121 network, the output end of the first channel attention module is connected with the second dense module, the output end of the second channel attention module is connected with the third dense module, and the output end of the third channel attention module is connected with the fourth dense module. The channel attention module in the application is used for spatial compression of the feature map output by the transition layer, and learning of the spatial compressed feature map in the channel dimension to obtain a feature map with rich channel dimension features.

[0010] Preferably, in step S2, the feature fusion module comprises three fully connected layers, a Concat layer and a layer normalization layer, wherein two fully connected layers are connected with the Concat layer, the Concat layer, the layer normalization layer and the other fully connected layer are connected in sequence, and the input ends of the two fully connected layers connected with the Concat layer are respectively connected with the output end of the fifth fully connected layer in the 1DCNN network and the output end of the 2D global average pooling layer in the At-DenseNet network; in the application, the two fully connected layers connected with the output end of the fifth fully connected layer in the 1DCNN network and the output end of the 2D global average pooling layer in the At-DenseNet network are both used for dimension reduction operation, so that the feature map output by the 1DCNN network and the feature map output by the At-DenseNet network have the same dimension, to facilitate subsequent feature fusion operation; the Concat layer is used for splicing the two feature maps in the feature dimension to obtain a preliminary fusion feature map; the layer normalization layer is used for normalizing the preliminary fusion feature map input therein, effectively reducing the difference between samples, so that the feature fusion module pays more attention to the distribution of the feature itself rather than the difference between samples, which helps to improve the generalization ability of the feature fusion module and alleviate the overfitting situation; and the fully connected layer connected with the layer normalization layer is used for dimension reduction operation on the feature map output by the layer normalization layer to obtain the final fusion feature map.

[0011] Preferably, step S3 comprises the following specific steps: inputting the one-dimensional sample data and the two-dimensional sample data in the training set into the perimeter security intrusion event recognition network, then calculating the loss value of the perimeter security intrusion event recognition network by using a multi-classification cross-entropy loss function, and optimizing the loss gradient by using an Adam optimizer and backpropagation, saving the parameters and updating the model parameters of the perimeter security intrusion event recognition network, while outputting the classification accuracy, to complete one training process; then, verifying the classification accuracy of the perimeter security intrusion event recognition network with updated model parameters by using the verification set; when the training times epoch reach the preset 50 times, the parameters reserved in the training segment epoch with the maximum classification accuracy obtained in the process of verifying the perimeter security intrusion event recognition network by using the verification set for 50 times are taken as the final model parameters of the perimeter security intrusion event recognition network, to obtain the perimeter security intrusion event recognition network model; the multi-classification cross-entropy loss function used in the present application is the multi-classification cross-entropy loss function disclosed in https: / / blog.csdn.net / BlackValley / article / details / 118006249.

[0012] Compared with the prior art, the beneficial effects of the present application are: the 1DCNN network in the present application can effectively extract the vibration signal time evolution information of the one-dimensional sample data, the At-DenseNet network can effectively extract the features of the time and frequency variation law in the two-dimensional sample data, the feature fusion module can fuse the vibration signal time evolution information features of the one-dimensional sample data extracted by the 1DCNN network and the features of the time and frequency variation law in the two-dimensional sample data extracted by the At-DenseNet network, to obtain a fusion feature map containing vibration signal time evolution information and time and frequency variation law information, and the Softmax classification layer can map the feature vector of the fusion feature map output by the feature fusion module to the [0, 1] interval and output the predicted classification accuracy of the background noise and the three intrusion events. Through testing, it is found that the classification accuracy of the background noise and the above three intrusion events can reach 99.5%-100% by using the present application, and the running time of the identification method described in the present application is 3.91S, and the parameter amount of the model of the perimeter security intrusion event recognition network is 4.97M, which shows that the identification method described in the present application has a shorter running time and lower model complexity, and is very suitable for monitoring perimeter security intrusion events. BRIEF DESCRIPTION OF DRAWINGS

[0013] Figure 1 is a schematic diagram of a perimeter security intrusion event recognition network structure in the present application; Figure 2 is a schematic diagram of a convolution group in the 1DCNN in Figure 1; Figure 3 is a schematic diagram of an initial convolution block in the At-DenseNet in Figure 1; Figure 4 is a schematic diagram of the first dense module in Figure 1, which is the leftmost dense module in Figure 1; Figure 5 is a schematic diagram of a feature fusion module; Figure 6 is a confusion matrix obtained by testing based on a test set; in Figure 6, label 0 represents background noise, label 1 represents a manual digging event, label 2 represents a personnel walking event, and label 3 represents a stone throwing event; (a) represents a DenseNet121 recognition method, (b) represents a 1DCNN+DenseNet121 recognition method, and (c) represents the recognition method described in the present application. DETAILED DESCRIPTION

[0014] A perimeter security intrusion event identification method, a flowchart of the perimeter security intrusion event identification method is shown in Figure 1, the perimeter security intrusion event identification method comprises the following steps: S1, collecting background noise vibration signals and intrusion event vibration signals, processing the collected background noise vibration signals and intrusion event vibration signals to obtain time domain vibration signal data and two-dimensional image data, and dividing the time domain vibration signal data and two-dimensional image data after normalization processing to obtain a model training set, a model test set and a model validation set; specifically comprising the following steps: S1-1, using a distributed optical fiber vibration sensing system (DVS) to collect background noise vibration signals and intrusion event vibration signals in the laid optical fiber area; wherein the background noise vibration signal refers to the vibration signal of the background noise generated by the natural environment without intrusion events, the types of intrusion events include artificial excavation events, personnel walking events and throwing stone events, the intrusion event vibration signal refers to the vibration signal generated by artificial excavation events, personnel walking events and throwing stone events, the distributed optical fiber vibration sensing system used in the present application is a distributed optical fiber vibration sensing DVS integrated module sold by Guilin Guangyi Intelligent Technology Co., Ltd., and the access address is: https: / / b2b.baidu.com / slist / 7e1f7b750374702b06027d627119067c047aadbfba; S1-2, performing data segmentation processing on the background noise vibration signals and intrusion event vibration signals collected in step S1-1, the unit time of data segmentation processing is 1s, and the data obtained after data segmentation processing within 1s is a time domain vibration signal data, and the time domain vibration signal data is one-dimensional data; S1-3, converting the time domain vibration signal data into two-dimensional image data using an existing data conversion method; wherein the existing data conversion method used in this embodiment is a continuous wavelet transform method, which is an existing time-frequency analysis method; S1-4, performing normalization processing on the time domain vibration signal data obtained in step S1-2 to obtain one-dimensional sample data; performing normalization processing on the two-dimensional image data obtained in step S1-3 to obtain two-dimensional sample data, and the pixel size of the two-dimensional sample data in this embodiment is 224x224.

[0015] S1-5, the one-dimensional sample data obtained in step S1-4 is randomly divided in a quantity ratio of 5:3:2 to obtain a sub-training set, a sub-validation set and a sub-test set of the one-dimensional sample data; the two-dimensional sample data obtained in step S1-4 is randomly divided in a quantity ratio of 5:3:2 to obtain a sub-training set, a sub-validation set and a sub-test set of the two-dimensional sample data; the sub-training set of the one-dimensional sample data and the sub-training set of the two-dimensional sample data constitute a model training set, the sub-validation set of the one-dimensional sample data and the sub-validation set of the two-dimensional sample data constitute a model validation set, and the sub-test set of the one-dimensional sample data and the sub-test set of the two-dimensional sample data constitute a model test set; S2, a perimeter security intrusion event recognition network is constructed: the perimeter security intrusion event recognition network comprises a 1DCNN network, an At-DenseNet network, a feature fusion module connected with the 1DCNN network and the At-DenseNet network, and a Softmax classification layer connected with the feature fusion module; in the present application, the At-DenseNet network is obtained by improving an existing DenseNet121 network (from the website: https: / / gitee.com / futureflsl / DenseNet_TensorFlow2 / blob / master / densenet.py); in the present application, the 1DCNN network is used to extract vibration signal time evolution information of the one-dimensional sample data, the At-DenseNet network is used to extract time and frequency variation characteristics of the two-dimensional sample data, the feature fusion module is used to fuse the vibration signal time evolution information of the one-dimensional sample data extracted by the 1DCNN network and the time and frequency variation characteristics of the two-dimensional sample data extracted by the At-DenseNet network to obtain a fusion feature map containing vibration signal time evolution information and time and frequency variation information, and the Softmax classification layer is used to map the feature vector of the fusion feature map output by the feature fusion module to the interval [0, 1] and output the classification accuracy of the predicted background noise and three intrusion events.

[0016] The 1DCNN network in the application comprises three convolution groups connected in sequence, the third convolution group is connected with a Flatten flattening layer, the Flatten flattening layer is connected with five fully connected layers in sequence, and the fifth fully connected layer is connected with a feature fusion module; wherein the three convolution groups are the same in structure and function, and the three convolution groups are all used for extracting vibration signal time evolution information in one-dimensional sample data, the Flatten flattening layer is used for flattening operation on the feature map with vibration signal time evolution information output by the third convolution group, to obtain a feature map of high-dimensional one-dimensional data with vibration signal time evolution information, the five fully connected layers in the 1DCNN network of the application are the same in structure and function, and the five fully connected layers are all used for integrating and mapping the features in the input feature map to different dimensional spaces. The first fully connected layer to the fifth fully connected layer respectively map the feature map to spaces with dimensions of 128, 64, 32, 16 and 8.

[0017] The three convolution groups of the 1DCNN network in the application are the same in structure, and the structure of the convolution group in the application is shown in FIG. 2, which comprises two one-dimensional convolution layers and one maximum pooling layer, and the first one-dimensional convolution layer, the second one-dimensional convolution layer and the maximum pooling layer are connected in sequence. Among them, the one-dimensional convolution layer in the convolution group is used for feature extraction of vibration signal time evolution information of one-dimensional sample data, and the maximum pooling layer in the convolution group is used for down-sampling dimension reduction operation on the feature map output by the second convolution layer, so as to reduce the calculation amount; the two one-dimensional convolution layers of the first convolution group each have 64 convolution kernels with a size of 3, and the two one-dimensional convolution layers of the first convolution group are mainly used for extracting local low-level features of vibration signal time evolution information of one-dimensional sample data, and then the maximum pooling layer performs down-sampling dimension reduction operation on the feature map with local low-level features output by the second one-dimensional convolution layer; the two one-dimensional convolution layers in the second convolution group each have 128 convolution kernels with a size of 3, and the two one-dimensional convolution layers in the second convolution group are mainly used for extracting middle-level features in the feature map output by the first convolution group, wherein the middle-level features refer to more complex local information or higher-level features, and the maximum pooling layer is used for down-sampling dimension reduction operation on the feature map with middle-level features output by the second one-dimensional convolution layer; the two one-dimensional convolution layers in the third convolution group each have 256 convolution kernels with a size of 3, and the two one-dimensional convolution layers in the third convolution group are mainly used for extracting high-level features in the feature map output by the second convolution group, and the maximum pooling layer is used for down-sampling dimension reduction operation on the feature map with high-level features output by the second one-dimensional convolution layer.

[0018] The setting of the three convolution groups in the 1DCNN network can extract the vibration signal time evolution information of the one-dimensional sample data at different scales, effectively improve the receptive field of the 1DCNN network, and thus enable the 1DCNN network to extract more abundant vibration signal time evolution information, and enable the 1DCNN network to better understand the time evolution information of the one-dimensional sample data.

[0019] The At-DenseNet network in the present application is obtained by improving the existing DenseNet121 network (from the website: https: / / gitee.com / futureflsl / DenseNet_TensorFlow2 / blob / master / densenet.py) as follows: improvement of the initial convolution group in the existing DenseNet121 network: the initial convolution group in the existing DenseNet121 network includes a 7x7 convolution layer, a batch normalization layer and a 2D maximum pooling layer connected in turn; the present application replaces the initial convolution group in the existing DenseNet121 network with the initial convolution block in the present application, the structure of the initial convolution block in the present application is shown in Figure 3, which includes three convolution units connected in turn, the convolution unit includes a 3x3 convolution layer, a batch normalization layer and a 2D maximum pooling layer connected in turn; the above setting effectively reduces the complexity of the model and improves the running speed, and since the convolution kernel size of the convolution layer used in the three convolution units is 3x3, the initial convolution block in the present application can better capture the local features in the two-dimensional sample data input, thereby extracting more detailed information of the frequency information; improvement of the DenceBlock in the existing DenseNet121 network: replace the dense convolution block (the dense convolution block is a convolution layer with a convolution kernel size of 3x3) in the DenceBlock in the existing DenseNet121 network with the dense convolution group in the present application, the structure of the dense convolution group is shown in Figure 4(b), the dense convolution group includes a batch normalization layer, a convolution layer with a convolution kernel size of 3x1 and a convolution layer with a convolution kernel size of 1x3 connected in turn; the number of dense convolution groups is set to six, and the bottleneck block and the six dense convolution groups are connected in a dense connection manner to obtain the first dense module; the number of dense convolution groups is set to eight, and the bottleneck block and the eight dense convolution groups are connected in a dense connection manner to obtain the second dense module, the third dense module and the fourth dense module respectively; the structure of the bottleneck block in the present application is the same as that of the bottleneck block in the DenceBlock in the existing DenseNet121 network, and has the same function, the structure of the bottleneck block in the present application is shown in Figure 4(a).The batch normalization layer in the dense convolution block in the application is used to alleviate overfitting, the convolution layer with a kernel size of 3x1 is used to perform feature extraction operation on the feature map in the input, and the convolution layer with a kernel size of 1x3 is used to perform feature extraction operation on the feature map output by the convolution layer with a kernel size of 3x1; in the application, the convolution layer with a kernel size of 3x1 and the convolution layer with a kernel size of 1x3 are used to perform convolution operation in sequence in the dense convolution group, which is an asymmetric convolution design, because the convolution layer with a kernel size of 3x1 and the convolution layer with a kernel size of 1x3 can capture features in different directions of the feature map respectively, therefore, the dense convolution block designed in the application can extract more features of time and frequency variation law in two-dimensional sample data, in addition, the setting of the convolution layer with a kernel size of 3x1 and the convolution layer with a kernel size of 1x3 can also effectively reduce the calculation amount due to the reduction of parameters.

[0020] In the existing DenseNet121 network, a channel attention module is added after each transition layer, the output end of the first channel attention module is connected with the second dense module, the output end of the second channel attention module is connected with the third dense module, and the output end of the third channel attention module is connected with the fourth dense module. The channel attention module in the application is prior art, which is derived from the paper Squeeze-and-Excitation Networks. The channel attention module in the application is used to perform spatial compression on the feature map output by the transition layer, and learn the spatial compressed feature map in the channel dimension to obtain a feature map with rich channel dimension features.

[0021] The feature fusion module in the application has a structure as shown in FIG. 5, including three fully connected layers, a Concat layer, and a layer normalization layer, wherein two fully connected layers are connected to the Concat layer, the Concat layer, the layer normalization layer, and another fully connected layer are sequentially connected, and the input ends of the two fully connected layers connected to the Concat layer are respectively connected to the output end of the fifth fully connected layer in the 1DCNN network and the output end of the 2D global average pooling layer in the At-DenseNet network; in the application, the two fully connected layers connected to the output end of the fifth fully connected layer in the 1DCNN network and the output end of the 2D global average pooling layer in the At-DenseNet network are both used for dimension reduction operation, so that the feature maps output by the 1DCNN network and the feature maps output by the At-DenseNet network have the same dimension, to facilitate subsequent feature fusion operation; the Concat layer is used for splicing two feature maps input therein in the feature dimension to obtain a preliminary fusion feature map; the layer normalization layer is used for normalizing the preliminary fusion feature map input therein, effectively reducing the difference between samples, so that the feature fusion module pays more attention to the distribution of the feature itself rather than the difference between samples, which helps to improve the generalization ability of the feature fusion module and relieve overfitting; and the fully connected layer connected to the layer normalization layer is used for dimension reduction operation on the feature map output by the layer normalization layer to obtain a final fusion feature map.

[0022] S3, training the perimeter security intrusion event recognition network by using the training set and a multi-classification cross-entropy loss function, and verifying the perimeter security intrusion event recognition network by using the verification set to obtain a perimeter security intrusion event recognition network model; the specific steps are as follows: inputting one-dimensional sample data and two-dimensional sample data in the training set into the perimeter security intrusion event recognition network, then calculating the loss value of the perimeter security intrusion event recognition network by using a multi-classification cross-entropy loss function, and optimizing the loss gradient by using an Adam optimizer and performing reverse propagation, saving parameters and updating the model parameters of the perimeter security intrusion event recognition network, and outputting the classification accuracy, to complete a training process; then, verifying the classification accuracy of the perimeter security intrusion event recognition network with updated model parameters by using the verification set, specifically: inputting one-dimensional sample data and two-dimensional sample data in the verification set into the perimeter security intrusion event recognition network with updated model parameters and performing forward propagation once, and outputting the classification accuracy. The present application verifies the perimeter security intrusion event recognition network once by using the verification set after training the perimeter security intrusion event recognition network once by using the training set, and when the training times epoch reach a preset 50 times, the parameters reserved in the largest training segment epoch in the classification accuracy obtained in the process of verifying the perimeter security intrusion event recognition network 50 times by using the verification set (that is, the process of verifying the perimeter security intrusion event recognition network 50 times by using the verification set) are reserved as the final model parameters of the perimeter security intrusion event recognition network, and the perimeter security intrusion event recognition network is the perimeter security intrusion event recognition network model. The multi-classification cross-entropy loss function used in the present application is the multi-classification cross-entropy loss function disclosed in https: / / blog.csdn.net / BlackValley / article / details / 118006249.

[0023] S4, processing the collected intrusion event vibration signals by using the steps S1-1 to S1-4 to obtain one-dimensional sample data and two-dimensional sample data; then inputting the obtained one-dimensional sample data and two-dimensional sample data into the perimeter security intrusion event recognition network model, performing forward propagation once, and outputting the predicted background noise and the classification accuracy of the three intrusion events.

[0024] Test: In order to verify that the perimeter security intrusion event recognition method described in the application has better performance in classification accuracy and prediction time compared with the existing intrusion event recognition method, the perimeter security intrusion event recognition method described in the application and the existing intrusion event recognition method, i.e. DenseNet121 recognition method (from https: / / gitee.com / futureflsl / DenseNet_TensorFlow2 / blob / master / densenet.p) and 1DCNN+DenseNet121 recognition method (which refers to replacing the At-DenseNet network in the perimeter security intrusion event recognition network described in the application with the Dencenet121 network in the prior art). Among them, the intrusion event recognition method described in the application and the above-mentioned 1DCNN+DenseNet121 recognition method use the same test strategy for testing. Specifically, the input data of the recognition method described in the application and the 1DCNN+DenseNet121 recognition method is the one-dimensional sample data and the two-dimensional sample data in the test set of the application, and before testing, the network model in the recognition method described in the application and the 1DCNN+DenseNet121 recognition method is trained for 50 epochs using the training set of the application, and after training the network model once using the training set, it is verified once using the validation set. When the training times epoch reach the preset 50 times, the parameters retained in the largest training segment epoch in the classification accuracy obtained during the process of verifying the perimeter security intrusion event recognition network using the validation set for 50 times are used as the network model parameters after training, and then the test set is tested. During the process of training the network model in the recognition method described in the application and the 1DCNN+DenseNet121 recognition method using the training set for 50 epochs, the network model in the recognition method described in the application and the 1DCNN+DenseNet121 recognition method will output the training set classification accuracy. During the process of verifying the network model in the recognition method described in the application and the 1DCNN+DenseNet121 recognition method using the validation set, the network model in the recognition method described in the application and the 1DCNN+DenseNet121 recognition method will output the validation set classification accuracy. During the process of testing the network model in the recognition method described in the application and the 1DCNN+DenseNet121 recognition method using the test set, the network model in the recognition method described in the application and the 1DCNN+DenseNet121 recognition method will output the test set classification accuracy.

[0025] Since the above DenseNet121 recognition method cannot process one-dimensional sample data, the two-dimensional data in the training set can only be used to train the network model in the DenseNet121 recognition method, the two-dimensional data in the validation set can only be used to verify the network model in the DenseNet121 recognition method, and the two-dimensional data in the test set can only be used to test the network model in the DenseNet121 recognition method, wherein the training, verification and testing methods are the same as the training, verification and testing methods of the intrusion event recognition method described in the application. The test results are shown in Table 1. And the confusion matrix obtained by testing based on the test set is shown in Figure 6.

[0026] Table 1 In Table 1, the training set classification accuracy and the validation set classification accuracy are used to show the overfitting of the model, the test classification accuracy is used to evaluate the generalization performance and prediction accuracy of the model; the prediction time in the application refers to the time between the input of the test set into the perimeter security intrusion event recognition network model and the output of the classification accuracy of the perimeter security intrusion event recognition network model.

[0027] From Table 1, it can be seen that the training set, validation set and test classification accuracy of the perimeter security intrusion event recognition method described in the application are the highest. Specifically: 1) the classification accuracy based on the training set and the validation set of the recognition method described in the application is the closest, which means that the recognition method described in the application has the best effect of suppressing overfitting.

[0028] 2) The classification accuracy based on the test set of the recognition method described in the application is the highest, which means that compared with the DenseNet121 recognition method and the 1DCNN+DenseNet121 recognition method, the generalization performance of the recognition method described in the application is the best.

[0029] 3) Compared with the existing DenseNet121 recognition method, since the existing DenseNet121 recognition method can only input single-modal data of two-dimensional images, the representation information is not comprehensive, and the network is complex and easy to cause overfitting phenomenon, the classification accuracy of the test set of the recognition method described in the application is improved by (99.36%-84.01%) / 84.01%=18.27% compared with the test classification accuracy of the existing DenseNet121 recognition method, which means that it is feasible to use multi-modal data for classification and recognition by the recognition method described in the application, and the classification accuracy is higher.

[0030] 3) Compared with the 1DCNN+DenseNet121 recognition method, the classification accuracy of the recognition method described in the application based on the test set is obviously higher than that of the 1DCNN+DenseNet121 recognition method, and the running time of the recognition method described in the application is 3.91S, which is reduced by 86.90% compared with the running time 29.84S of the 1DCNN+DenseNet121 recognition method, which shows that the At-DenseNet network based on the improvement of the existing DenseNet121 network not only effectively improves the classification accuracy but also significantly improves the running speed.

[0031] The confusion matrix obtained by testing based on the test set in the application is also called a possibility matrix or an error matrix, which is a way to evaluate the performance of a classification model. The accuracy of the classification result is displayed in a confusion matrix, which is used to show the relationship between the classification result and the true class of the network model in different classes, and is mainly used to compare the classification result and the actual measured value. Each column of the confusion matrix represents the predicted class (PredictedLabel), and each row represents the true class (TrueLabel). In the confusion matrix, the elements on the diagonal line represent the classification accuracy, and the non-diagonal line represents the error classification rate.

[0032] As can be seen from FIG. 6, the classification accuracy of the recognition method described in the application for background noise and the above three intrusion events can reach 99.5%-100%; for background noise data (label 0): since background noise is the case of no event occurrence in a natural environment, the vibration signal can be easily distinguished, so the classification accuracy of the three methods for background noise reaches 100%; for artificial excavation events: the prediction accuracy of the recognition method described in the application reaches 99.5%, which is improved by (99.5%-92%) / 92%=8.15% compared with the existing DenseNet121 recognition method; for personnel walking events: the prediction accuracy of the recognition method described in the application reaches 100%, which is improved by (100%-82.67%) / 82.67%=20.96% compared with the existing DenseNet121 recognition method; for throwing stone events: the prediction accuracy of the recognition method described in the application reaches 100%, which is improved by (100%-59.64%) / 59.64%=67.67% compared with the existing DenseNet121 recognition method.

[0033] In addition, in order to more objectively evaluate the effectiveness of the recognition method described in the application and the performance of the model, the application also uses Precision, Recall and F1-Score three evaluation indexes to evaluate the classification result, and the Precision, Recall and F1-Score three evaluation indexes are shown in Table 2.

[0034] Table 2 In Table 2, Precision is the precision, indicating the proportion of samples that are actually true in all samples that are predicted to be true, used to evaluate the quality of the model in predicting positive samples; Recall is the recall rate, indicating the proportion of samples that are predicted to be true in all samples that are actually true, used to evaluate the ability of the model to find positive samples; F1-Score is the F1 score, which is the harmonic mean of precision and recall, and in the case of equal importance of precision and recall, the F1 score is a useful indicator.

[0035] As can be seen from Table 2, the recognition method described in the present application achieves good results in the three evaluation indicators of Precision, Recall and F1-Score, as follows: Precision index comparison: the recognition method described in the present application achieves 0.9975 in the Precision index, which is improved by (99.75%-87.50%) / 87.50%=14% compared with the existing DenseNet121 method; Recall index comparison: the recognition method described in the present application achieves 0.9975 in the Recall index, which is improved by (99.75%-83.75%) / 83.75%=19.1% compared with the existing DenseNet121 method; F1-Score index comparison: the recognition method described in the present application achieves 1 in the F1-Score index, which is improved by (100%-83.75%) / 83.75%=19.4% compared with the existing DenseNet121 method.

[0036] In addition, the present application also compares the parameter amount of the perimeter security intrusion event recognition network, the At-DenseNet network and the existing DenseNet121 network in the present application, as shown in Table 3: Table 3 The parameter amount of the model is one of the evaluation indicators of the model complexity.

[0037] As can be seen from Table 3, the model parameter amount of the At-DenseNet network in the present application is only 6.3% of the model parameter amount of the existing DenseNet121 network, and the At-DenseNet network and the existing DenseNet121 network can only process two-dimensional data; while the perimeter security intrusion event identification network in the present application can process one-dimensional data and two-dimensional data at the same time, and improve the identification accuracy based on the ability to process the above-mentioned multi-modal data. It is worth noting that although the present application can process the above-mentioned multi-modal data to improve the identification accuracy, the model parameter amount is not increased. Through testing, it is found that the model parameter amount of the perimeter security intrusion event identification network in the present application is 4.97M, which is only 66.28% of the model parameter amount of the existing DenseNet121 network. Obviously, the identification method described in the present application not only improves the identification accuracy, but also reduces the model complexity and improves the running speed.

Claims

1. A perimeter security intrusion event recognition method, characterized by: Comprise the following steps: S1, collect background noise vibration signal and intrusion event vibration signal, and then process to obtain time domain vibration signal data and two-dimensional image data, and then divide after normalization processing, obtain model training set, model test set and model validation set; S2, construct perimeter security intrusion event recognition network; the perimeter security intrusion event recognition network comprises a 1DCNN network, an At-DenseNet network, a feature fusion module and a Softmax classification layer; the 1DCNN network is used to extract the vibration signal time evolution information of one-dimensional sample data, the At-DenseNet network is used to extract the time and frequency variation characteristics of two-dimensional sample data, the feature fusion module is used to fuse the vibration signal time evolution information characteristics of one-dimensional sample data extracted by the 1DCNN network and the time and frequency variation characteristics of two-dimensional sample data extracted by the At-DenseNet network, and the Softmax classification layer is used to map the feature vector of the fusion feature map output by the feature fusion module to the interval [0, 1] and output the predicted background noise and three intrusion event classification accuracies; S3, use the training set and the multi-classification cross-entropy loss function to train the perimeter security intrusion event recognition network, and use the validation set to verify the perimeter security intrusion event recognition network, and obtain the perimeter security intrusion event recognition network model; S4, input the collected intrusion event vibration signal into the perimeter security intrusion event recognition network model after processing by step S1, forward propagate once, and output the predicted background noise and three intrusion event classification accuracies.

2. The perimeter security intrusion event recognition method of claim 1, wherein: Step S1 specifically comprises the following steps: S1-1, collect the background noise vibration signal and the intrusion event vibration signal of the laid optical fiber area by using a distributed optical fiber vibration sensing system; the intrusion event vibration signal refers to the vibration signal generated by artificial excavation events, personnel walking events and throwing stone events; S1-2, perform data segmentation processing on the background noise vibration signal and the intrusion event vibration signal collected in step S1-1, the unit time of data segmentation processing is 1s, and the data obtained after data segmentation processing within 1s is a time domain vibration signal data, which is one-dimensional data; S1-3, convert the time domain vibration signal data into two-dimensional image data by using an existing data conversion method; S1-4, normalize the time domain vibration signal data obtained in step S1-2 to obtain one-dimensional sample data; normalize the two-dimensional image data obtained in step S1-3 to obtain two-dimensional sample data; S1-5, the one-dimensional sample data obtained in step S1-4 is randomly divided in a quantity ratio of 5:3:2 to obtain a sub-training set, a sub-validation set and a sub-test set of the one-dimensional sample data; the two-dimensional sample data obtained in step S1-4 is randomly divided in a quantity ratio of 5:3:2 to obtain a sub-training set, a sub-validation set and a sub-test set of the two-dimensional sample data; the sub-training set of the one-dimensional sample data and the sub-training set of the two-dimensional sample data constitute a model training set, the sub-validation set of the one-dimensional sample data and the sub-validation set of the two-dimensional sample data constitute a model validation set, and the sub-test set of the one-dimensional sample data and the sub-test set of the two-dimensional sample data constitute a model test set.

3. The perimeter security intrusion event recognition method of claim 2, wherein: In step S1-3, the existing data conversion method is an existing time-frequency analysis method and / or an existing image encoding method; the time-frequency analysis method is one or more of a short-time Fourier transform method, a continuous wavelet transform method and a Hilbert-Huang transform method; and the image encoding method is one or more of a Gram angle field method, a Markov transition field method and a recursive graph method.

4. The perimeter security intrusion event recognition method of claim 1, wherein: In step S2, the 1DCNN network comprises three convolution groups connected in sequence, the third convolution group is connected with a Flatten flattening layer, the Flatten flattening layer is connected with five fully connected layers in sequence, and the fifth fully connected layer is connected with the feature fusion module; wherein the three convolution groups are used for extracting the vibration signal time evolution information in the one-dimensional sample data, the Flatten flattening layer is used for flattening operation on the feature map output by the third convolution group to obtain a feature map of high-dimensional one-dimensional data with vibration signal time evolution information, and the five fully connected layers are used for integrating and mapping the features in the input feature map to different dimensional spaces, and the first fully connected layer to the fifth fully connected layer respectively map the feature map to spaces with dimensions of 128, 64, 32, 16 and 8.

5. The perimeter security intrusion event recognition method of claim 4, wherein: In step S2, the two one-dimensional convolution layers of the first convolution group each have 64 convolution kernels with a size of 3, the two one-dimensional convolution layers in the second convolution group each have 128 convolution kernels with a size of 3, and the two one-dimensional convolution layers in the third convolution group each have 256 convolution kernels with a size of 3.

6. The perimeter security intrusion event recognition method of claim 1, wherein: In step S2, the At-DenseNet network is obtained by improving the existing DenseNet121 network as follows: the initial convolution group in the existing DenseNet121 network is replaced by an initial convolution block, and the initial convolution block comprises three convolution units connected in sequence, and each convolution unit comprises a 3*3 convolution layer, a batch normalization layer and a 2D maximum pooling layer connected in sequence; The dense convolution block in the Dence Block in the existing DenseNet121 network is replaced by a dense convolution group, the dense convolution group comprising a batch normalization layer, a convolution layer with a convolution kernel size of 3x1 and a convolution layer with a convolution kernel size of 1x3 connected in sequence; the number of dense convolution groups is set to six, and the bottleneck block and the six dense convolution groups are connected in a dense connection manner to obtain a first dense module; the number of dense convolution groups is set to eight, and the bottleneck block and the eight dense convolution groups are connected in a dense connection manner to obtain a second dense module, a third dense module and a fourth dense module respectively. A channel attention module is added after each transition layer in the existing DenseNet121 network, the output end of the first channel attention module is connected with the second dense module, the output end of the second channel attention module is connected with the third dense module, and the output end of the third channel attention module is connected with the fourth dense module.

7. The perimeter security intrusion event recognition method of claim 1, wherein: In step S2, the feature fusion module comprises three fully connected layers, a Concat layer and a layer normalization layer, wherein two fully connected layers are connected with the Concat layer, the Concat layer, the layer normalization layer and the other fully connected layer are connected in sequence, and the input ends of the two fully connected layers connected with the Concat layer are connected with the output end of the fifth fully connected layer in the 1DCNN network and the output end of the 2D global average pooling layer in the At-DenseNet network respectively; wherein the two fully connected layers connected with the output end of the fifth fully connected layer in the 1DCNN network and the output end of the 2D global average pooling layer in the At-DenseNet network respectively are used for dimension reduction operation, so that the feature maps output by the 1DCNN network and the feature maps output by the At-DenseNet network have the same dimension; the Concat layer is used for splicing two feature maps in the input in the feature dimension to obtain a preliminary fusion feature map; the layer normalization layer is used for normalizing the preliminary fusion feature map in the input; and the fully connected layer connected with the layer normalization layer is used for dimension reduction operation on the feature map output by the layer normalization layer to obtain a final fusion feature map.

8. The perimeter security intrusion event recognition method of claim 1, wherein: The step S3 comprises the following specific steps: inputting the one-dimensional sample data and the two-dimensional sample data in the training set into the perimeter security intrusion event identification network, then calculating the loss value of the perimeter security intrusion event identification network by using a multi-classification cross-entropy loss function, and optimizing the loss gradient and back-propagating by using an Adam optimizer, saving the parameters and updating the model parameters of the perimeter security intrusion event identification network, and outputting the classification accuracy, thereby completing a training process; then, verifying the classification accuracy of the perimeter security intrusion event identification network with the updated model parameters by using a verification set; when the training times epoch reach the preset 50 times, the parameters reserved in the training segment epoch with the maximum classification accuracy obtained in the process of verifying the perimeter security intrusion event identification network by using the verification set for 50 times are taken as the final model parameters of the perimeter security intrusion event identification network, and the perimeter security intrusion event identification network model is obtained.

Images