Φ-OTDR power optical cable perimeter security intrusion event recognition method based on AsymConv-Densenet network

Abstract

The Φ-OTDR power optical cable perimeter security intrusion event identification method based on the AsymConv-Densenet network relates to the field of Φ-OTDR event identification technology, and solves the problem that the signal features in the prior art for identifying Φ-OTDR intrusion event signals for power optical cable perimeter security need to be manually extracted, the recognition performance depends on pre-trained weights, and there is recognition confusion for typical intrusion events such as climbing and walking, shaking and knocking, etc., with similar time-frequency features. The present invention combines densely connected convolutional networks with asymmetric convolutional blocks and proposes an AsymConv-Densenet network. This network optimizes the Densenet network using asymmetric convolutional blocks, and re-extracts features from the result matrix of the conversion block in the Densenet network. In order to avoid adding additional computational burden to subsequent layers, the size and dimension of the result matrix obtained by the asymmetric convolution block are the same as the result matrix of the conversion block.

Description

Technical Field

[0001] The invention relates to the technical field of Φ-OTDR event recognition, and in particular to a Φ-OTDR power optical cable perimeter security intrusion event recognition method based on a densely connected convolutional network (AsymConv-Densenet network) of an asymmetric convolution module. Background Art

[0002] At present, power optical cable is one of the important components of modern smart grid, which is of great significance for realizing the informatization and intelligence of power grid. Therefore, the requirements for perimeter security of power optical cable are getting higher and higher. Common security measures include manual inspection and video monitoring. Manual inspection is often affected by factors such as time, region, and quality of staff, and often omissions and errors occur. Video monitoring is easily affected by bad weather, such as strong light, haze, frost, etc., and with the increase in the number of monitoring equipment, more staff are needed to monitor and monitor, and the economic cost will increase. Φ-OTDR has the advantages of high sensitivity, wide deployment range, long service life, and anti-electromagnetic interference. It can monitor the vibration signal generated by the perimeter of power optical cable in real time and reflect the characteristics of different vibration sources. Therefore, the power optical cable is fortified with fences, and then the optical fiber is laid around the fence and connected to the Φ-OTDR system. The generated Φ-OTDR signal is analyzed by algorithm to identify different intrusion events, which greatly reduces labor costs and economic costs. Therefore, Φ-OTDR is very suitable for perimeter security deployment of power optical cable.

[0003] In recent years, machine learning and deep learning models have been applied to Φ-OTDR event recognition. The machine learning method first extracts the distinguishable features of Φ-OTDR events by manual methods and uses classification algorithms for recognition, but the method of manually extracting features is easy to implement when the data set is limited and the event types are few. In 2019, Jia et al. screened out 9 artificially extracted features of Φ-OTDR events in the time domain and frequency domain by the k-nearest neighbor (kNN) method, and then used SVM to identify five interference events such as watering, climbing, knocking, pressing, and false interference. The total number of samples was 750, and the average recognition rate was 94.38%. If the event types are similar, the distinguishable features of the events are difficult to extract manually, and the accuracy of recognition is difficult to guarantee. In 2023, Cao et al. extracted 32 features of Φ-OTDR events in the time domain and frequency domain, and used the SVM algorithm to identify the vibration data of background noise, excavation, knocking, watering, shaking, and walking. The number of samples was 15,612, and the average recognition rate was only 82.60%. Among them, the excavation and knocking vibration data were intermittent peaks, and the background noise and watering vibration data were continuous and slowly changing. There were events with similar vibrations. Although the number of manual feature extraction and the number of samples were increased, the accuracy was low.

[0004] Deep learning models mainly extract features in an adaptive way. In 2019, Yi Shi et al. proposed an event recognition method that applied convolutional neural networks (CNNs) to the Φ-OTDR spatiotemporal data matrix, with an average accuracy of 96.67%. In 2020, H. Wu et al. used one-dimensional CNNs to extract the spatiotemporal features of signals. The bidirectional long short-term memory (BiLSTM) network was used to mine the spatial relationship between different signal nodes, with an average accuracy of 97.31%. In 2021, Tian et al. combined the attention mechanism with the convolutional neural network and proposed the ATCN-BiLSTM network, in which ATCN extracted the causal relationship of the time domain signal and BiLSTM extracted the bidirectional spatial relationship. The two were used to optimize the convolutional neural network, with an average accuracy of 99.60%. In 2023, Yi Shi et al. proposed a method of combining transformer networks with Φ-OTDR to identify vibration events in underground conditions, including background, bicycle passing, rain, tapping, and walking, with an average accuracy of 99.64%. Compared with convolutional neural networks, transformer networks have higher accuracy, but they are more dependent on pre-trained models. If the pre-trained model is not suitable, the performance will be lower than that of convolutional neural networks. The real-life scenarios and simulated experimental environments for the study of Φ-OTDR vibration data are mostly for underground burials, and rarely collect and simulate events related to perimeter security of power optical cables, making it difficult to find a suitable pre-trained model to train the transformer network.

[0005] Therefore, it is reasonable to use a method combining convolutional neural networks and optimization modules to identify power optical cable perimeter security intrusion events. First, there is no need for pre-training models for initialization. Second, it can avoid the impact of manual labeling on subsequent feature extraction. Finally, features can be adaptively extracted through a large number of samples to solve the problem of similar events being difficult to distinguish. Summary of the invention

[0006] The present invention aims to solve the problems in the prior art of identifying Φ-OTDR intrusion event signals for power optical cable perimeter security, namely, that signal features need to be manually extracted, recognition performance depends on pre-trained weights, and there is recognition confusion for typical intrusion events such as climbing and walking, shaking and knocking, etc. The present invention provides a Φ-OTDR power optical cable perimeter security intrusion event recognition method based on an AsymConv-Densenet network.

[0007] A Φ-OTDR power optical cable perimeter security intrusion event recognition method based on AsymConv-Densenet network, the implementation process of this method is as follows:

[0008] Step 1: Convert the collected Φ-OTDR signals of different events into time-frequency diagrams to construct a Φ-OTDR signal vibration event image dataset;

[0009] Step 2: construct an AsymConv-Densenet network model, and train the network model to obtain a trained network model;

[0010] The network model consists of a convolutional layer, a dense connection module, a conversion module, a feature extraction module, a global pooling layer and a fully connected layer;

[0011] The dense connection module includes dense block 1, dense block 2 and dense block 3;

[0012] The conversion module includes conversion block 1, conversion block 2 and conversion block 3;

[0013] The feature extraction module includes an asymmetric convolution block 1, an asymmetric convolution block 2 and an asymmetric convolution block 3;

[0014] Step 3: Use the trained network model to identify the Φ-OTDR signal and realize the identification of the vibration event of the Φ-OTDR signal; the specific identification process is as follows:

[0015] The time-frequency diagram of the Φ-OTDR signal is input into the network model, and the frequency components, time variation characteristics and texture characteristics in the time-frequency diagram are extracted through the convolution layer to obtain a feature diagram; the feature diagram sequentially captures the global feature diagram of different vibration events through dense block 1, conversion block 1, asymmetric convolution block 1, dense block 2, conversion block 2, asymmetric convolution block 2, dense block 3, conversion block 3 and asymmetric convolution block 3, and the global feature diagram is averaged by the pixel values ​​of each global feature diagram through the global pooling layer to obtain a feature vector of fixed length; the feature vector is output by SoftMax after the fully connected layer to output the probability corresponding to each label in the Φ-OTDR data, and the vibration event label with the maximum probability is used as the output result, finally realizing the recognition of vibration events of the Φ-OTDR signal.

[0016] Beneficial effects of the present invention:

[0017] 1. In the method of the present invention, a densely connected convolutional network (Densenet network) and an asymmetric convolution block (AsymConvBlock) are combined to propose an AsymConv-Densenet network. This network uses an asymmetric convolution block to optimize the Densenet network, and re-extracts features from the result matrix of the conversion block in the Densenet network. It solves the problem of feature omission caused by the conversion block in the Densenet network due to the reduction of calculation amount and compression of feature size. In order to avoid adding additional computational burden to subsequent layers, the size and dimension of the result matrix obtained by the asymmetric convolution block are the same as the result matrix of the conversion block.

[0018] 2. In order to better verify the rationality of the design of Densenet network combined with asymmetric convolution blocks, this paper compares the Densenet network and the Densenet network combined with the attention mechanism module, such as SE-Densenet network and CBAM-Densenet network, which have performance advantages in precision, average recall rate and average F1 score. By comparing their respective confusion matrices, it is found that they have precision advantages in tapping, shaking, climbing and walking.

[0019] 3. Experiments are used to verify that the AsymConv-Densenet network has higher performance advantages for typical power optical cable perimeter security intrusion events such as knocking, shaking, climbing, and walking. The proposed algorithm has been proven to be superior to the existing Alexnet network, Resnet network-18, Googlenet network, and SwinTransformer network. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 It is a structural diagram of the Densenet network;

[0021] Figure 2 It is a structural diagram of the AsymConv-Densenet network model in the Φ-OTDR power optical cable perimeter security intrusion event identification method based on the AsymConv-Densenet network described in the present invention;

[0022] Figure 3 Schematic diagram of the structure of the asymmetric convolution block;

[0023] Figure 4 The following are the effects of eight Φ-OTDR event types: (a) the effect of heavy objects falling; (b) the effect of friction; (c) the effect of noise; (d) the effect of piezoelectric ceramic vibration; (e) the effect of knocking; (f) the effect of shaking; (g) the effect of climbing; (h) the effect of walking;

[0024] Figure 5Schematic diagram of confusion matrix: (a) is the confusion matrix of CBAM-Densenet network; (b) is the confusion matrix of SE-Densenet network; (c) is the confusion matrix of Densenet network; (d) is the confusion matrix of AsymConv-Densenet network.

[0025] Figure 6 It is a comparison effect diagram of the accuracy curves of the network in the method of the present invention and four networks. DETAILED DESCRIPTION

[0026] Specific implementation method 1. Combination Figures 1 to 6 This embodiment describes a Φ-OTDR power optical cable perimeter security intrusion event identification method based on the AsymConv-Densenet network. The method selects a backbone network by analyzing the advantages of the Densenet network structure, and on this basis, designs an AsymConv-Densenet network model. The specific method is implemented by the following steps:

[0027] First, the collected Φ-OTDR signals of different events are converted into time-frequency diagrams through pre-emphasis, windowing and framing, spectral subtraction and short-time Fourier transform, and then constructed into a Φ-OTDR signal vibration event data set;

[0028] 2. Construct the AsymConv-Densenet network model. The core modules include dense connection module, conversion module and feature extraction module. The dense connection module includes dense block 1, dense block 2 and dense block 3. The conversion module includes conversion block 1, conversion block 2 and conversion block 3. The feature extraction module includes asymmetric convolution block 1, asymmetric convolution block 2 and asymmetric convolution block 3.

[0029] The AsymConv-Densenet network model is trained using the Φ-OTDR signal vibration event data set; and the trained AsymConv-Densenet network model is obtained;

[0030] 3. Using the trained AsymConv-Densenet network model to perform Φ-OTDR signal recognition, the specific process is as follows:

[0031] After an initial convolutional layer, the frequency components, time-varying features, and texture features in the time-frequency graph are extracted;

[0032] After dense block 1, through dense connections, each layer can be regarded as an integration and enhancement of the features of all previous layers. The feature map learned by the new layer is spliced ​​with the previous feature map to form a richer feature representation for different vibration events;

[0033] After conversion block 1, the feature dimension and spatial size are reduced, the amount of calculation is reduced, and overfitting is avoided. This helps to control the complexity of the model and the demand for computing resources.

[0034] In the training phase, after the asymmetric convolution block 1, the result obtained by the conversion block 1 will be analyzed at multiple scales through 1×3 convolution kernel, 3×1 convolution kernel, and 3×3 convolution kernel, further extracting more detailed information to distinguish vibration events, and merging the features through splicing and batch normalization. In this phase, the weight parameters of the trained asymmetric convolution block 1 will no longer change. The 1×3 convolution kernel, 3×1 convolution kernel, and 3×3 convolution kernel can be added together, and the obtained mixed 3×3 convolution can be used for feature extraction and batch normalization. At the same time, the asymmetric convolution block 1 does not change the size and dimension of the result of the conversion block 1, and does not increase the computational burden of subsequent layers.

[0035] Based on the above steps, dense block 2, conversion block 2, asymmetric convolution block 2, dense block 3, conversion block 3, asymmetric convolution block 3 are sequentially passed to capture the basic features of different vibration events as well as more abstract and global feature representations, thus ensuring the high efficiency and generalization ability of the model;

[0036] After global average pooling, the pixel values ​​of each feature map are averaged to obtain a feature vector of fixed length;

[0037] After a fully connected layer, SoftMax is used to output the probabilities corresponding to each label in the Φ-OTDR data, and the event label with the maximum probability is used as the output result. Finally, the vibration event recognition of the Φ-OTDR signal is realized.

[0038] In this implementation, the backbone network Densenet is first selected;

[0039] A single-channel image is used, i.e., images in the existing public Φ-OTDR dataset; a three-channel image is used, i.e., images in the existing public CIFAR-100 dataset and Imagenet dataset. Compared with images in the CIFAR-100 dataset and the Imagenet dataset, the images in the Φ-OTDR dataset contain less information. Moreover, the pixel values ​​of the two types of events, noise and walking, are reduced overall, and the extracted features are not obvious. However, the feature extraction capability of the convolutional neural network is required to be high. Compared with the Alexnet network, the Googlenet network, and the Resnet network, the densely connected module in the Densenet network concatenates the outputs of all previous layers as the input of the current layer, which enhances the propagation of features between networks, realizes feature reuse, and can effectively improve the accuracy of vibration event recognition. The number of output channels of each convolution block in the Densenet network is a fixed small value, and the amount of calculation is smaller. This embodiment uses the Densenet network as the backbone network. The Densenet network architecture is as follows: Figure 1 shown.

[0040] The Densenet network includes a convolution layer, a dense connection module, a conversion module, a global pooling layer and a fully connected layer; the dense connection module includes dense block 1, dense block 2, dense block 3 and dense block 4; the conversion module includes conversion block 1, conversion block 2 and conversion block 3;

[0041] After the input image is feature extracted by the convolution layer, it passes through dense block 1, conversion block 1, dense block 2, conversion block 2, dense block 3, conversion block 3, dense block 4 in sequence, and then performs operations of the global pooling layer and the fully connected layer before being output.

[0042] based on Figure 1 The Densenet network structure is designed in this implementation. The AsymConv-Densenet network is designed. Since the conversion module will compress the dimension and size of the feature matrix, some detail information will be lost. In the Φ-OTDR image data set, the features in the images of climbing and walking event types are scattered and irregular. If only the Densenet network is used for identification, the performance improvement of intrusion event identification is limited. In order to enable the Densenet network to extract more distinguishing features of similar intrusion events such as Φ-OTDR climbing and walking, an asymmetric convolution block is added after each conversion block of the Densenet network, and the dense fast 4 is removed. The Densenet network and the feature extraction module are merged to form an AsymConv-Densenet network, whose structure is as follows: Figure 2 shown.

[0043] In this embodiment, the AsymConv-Densenet network consists of a convolutional layer, three dense blocks, three conversion blocks, three asymmetric convolutional blocks, and a fully connected layer. Each layer in the dense block is a nonlinear combination function, which consists of a normalization layer (BN), a rectified linear unit (ReLU), and a 3×3 convolutional layer. In order to achieve a higher degree of feature reuse, each layer in the dense block can process all the features obtained by the previous layer. For the kth layer, its forward layer has x0, x1, x2, ..., x k-1 Combination function, the formula for dense connection is:

[0044] X k =H k ([x0,x1,x2,...,x k-1 ])(1)

[0045] where [x0,x1,x2,...,x k-1 ], indicating that these feature layers are concatenated in the channel dimension. k () represents a combined operation, namely: the result of a normalization layer (BN), a rectified linear unit (ReLU), and a 3*3 convolution layer. The conversion block is composed of a 1×1 convolution layer and a 2×2 average pooling layer.

[0046] In this embodiment, the time-frequency diagram of the Φ-OTDR signal is sequentially passed through the convolution layer, dense block 1, conversion block 1 and asymmetric convolution block 1, dense block 2, conversion block 2 and asymmetric convolution block 2, dense block 3, conversion block 3, asymmetric convolution block 3 and a global pooling layer. Finally, the feature learning of the image data of Φ-OTDR is completed through the fully connected layer. Finally, this layer uses SoftMax to output the probability corresponding to each label in the Φ-OTDR data, and uses the event label with the maximum probability as the output result to complete the recognition of the vibration event of the Φ-OTDR signal.

[0047] like Figure 3 As shown, in this implementation, the asymmetric convolution block is responsible for feature extraction of the result matrix obtained by the conversion block, and the core idea of ​​the asymmetric convolution block is to add an asymmetric convolution block to enhance the kernel skeleton of the convolution layer. The specific operation is to replace the three parallel d×d convolution kernels with 1×d convolution kernel, d×1 convolution kernel, and d×d convolution kernel. The advantage of this module is that it extracts more detailed information to distinguish vibration events without introducing additional hyperparameters. After the core weights of the validation set have been determined, the asymmetric convolution block can be equivalently converted into a standard layer to reduce the amount of computation in the network verification stage. For those who do not want additional computational burden while improving the performance of intrusion event recognition, the asymmetric convolution block is an ideal network feature extraction module.

[0048] The AsymConv-Densenet network uses asymmetric convolution blocks to extract the feature matrix obtained by the conversion block in the Densenet network. The asymmetric convolution block does not change the size and dimension of the feature matrix, nor does it increase the computational burden of subsequent layers. It can also extract more feature information, which can further improve the ability to identify intrusion events in Φ-OTDR data.

[0049] In this implementation, the 1×d convolution kernel, d×1 convolution kernel, and d×d convolution kernel in the asymmetric convolution block are replaced by 1×3, 3×1, and 3×3 convolution kernels. According to formula (2), 1×3, 3×1, and 3×3 convolution kernels are compatible and can be used as the convolution kernel sizes in the asymmetric convolution block.

[0050] M (q) =M (p) =M (r) ,Y q ≤Y p ,Y r ≤Y p , X q ≤X p ,X r ≤X p ,C q =C p =C r (2)

[0051] Where: M (q) , M (p) and M (r) are the inputs of the qth, pth and rth layers respectively, Y and X are the dimensions of the kernel, and C is the number of channels of the convolutional layer.

[0052] In this implementation, the structure of each asymmetric convolutional block in the AsymConv-Densenet network is as follows: Figure 3 As shown in the figure, in the training phase, the asymmetric convolution block contains 3 parallel layers, and the convolution kernel sizes are 3×3, 1×3 and 3×1 respectively. In order to ensure that the output feature matrix size and dimension are the same as the feature matrix obtained by the conversion block, in the 3×3, 1×3 and 3×1 convolution kernel operations, the asymmetric convolution block sets appropriate padding and stride for the feature matrix obtained by the conversion block. Then batch normalization is performed in sequence, and the outputs of the three branches are element-wise added. In the verification phase, the 3×3, 1×3 and 3×1 convolution kernels are normalized respectively, and the three convolution kernels are added together, and then the feature matrix obtained by the conversion block is operated with the added convolution kernel. The output results obtained in the training phase and the verification phase are the same and have the same size and dimension as the conversion block. The proof process is as follows:

[0053] According to formula (3), the 1×3, 3×1 and 3×3 convolution kernels conform to the addition principle in the asymmetric convolution block.

[0054] I*F(1)+I*F(2)+I*F(3)=I*(F(1)+F(2)+F(3))(3)

[0055] Where * denotes a two-dimensional convolution operation, I is a matrix, F(1), F(2), and F(3) are three two-dimensional kernels with compatible sizes, and + is the element-wise addition of the kernel parameters at corresponding positions.

[0056] In this embodiment, I∈R H×W×C Represents a convolution kernel with spatial resolution H×W and C channels, M∈R U×V×C Represents a feature matrix with spatial resolution U×V and C channels, O∈R R×T×D Represents the output matrix with spatial resolution R×T and D after the convolution operation. In order to reduce overfitting and speed up the training process, the input feature matrix after the convolution operation is normalized, and the formula is as follows:

[0057]

[0058] Formula (4) can be transformed into formula (5), which is as follows:

[0059]

[0060] Where * represents a two-dimensional convolution operation, O :,:,j It represents the result matrix obtained by convolution operation on the input feature matrix M using convolution kernel j. Where * represents two-dimensional convolution operation, M :,:,k Represents the k-th channel U×V matrix, Indicates F (j) The k-th channel H×W 2D convolution kernel. j and σ j is the channel mean and standard deviation of batch normalization, γ j and β j are the learned scale factor and bias term respectively.

[0061] In the verification stage, according to formula (5), the 3×3, 1×3 and 3×1 convolution kernels are batch normalized, and the three convolution kernels obtained are and The constant terms are and According to the principle of formula (3), we can get formula (6). Add the constant terms corresponding to the 3×3, 1×3 and 3×1 convolution kernels to get formula (7), which is as follows:

[0062]

[0063] In the formula, F′ (j) is the fused 3×3 convolution kernel, b j is the fused bias term.

[0064] Combine the feature matrix M with the fused 3×3 convolution kernel F′ (j) Perform convolution operation and add the fused bias term b j According to formula (5), formula (6) and formula (7), we have the following formula:

[0065]

[0066] In the formula, O :,:,j It represents the result matrix obtained by convolving the feature matrix M with the 3×3 convolution kernel in filter j. It represents the result matrix obtained by convolving the feature matrix M with the 1×3 convolution kernel in filter j. Represents the result matrix obtained by convolving the feature matrix M with the 3×1 convolution kernel in filter j.

[0067] In summary, the result matrix obtained in the training phase and the validation phase is the same. Therefore, the asymmetric convolution block is equivalently converted to a standard layer, which reduces the amount of computation in the network validation phase.

[0068] Specific implementation method 2: Combination Figures 4 to 6 This embodiment is described as an example of the method for identifying Φ-OTDR power optical cable perimeter security intrusion events based on the AsymConv-Densenet network described in the first embodiment.

[0069] 1. In this implementation, the Φ-OTDR data set of Zhejiang University is selected. The laboratory performs pre-emphasis, windowing, spectral subtraction, and short-time Fourier transform on the signal data collected by the Φ-OTDR device, and converts it into a time-frequency diagram with 200 frames, a frequency range of 0 to 2500HZ, 128 frequency components, and a relative amplitude size of 0 to 255. The horizontal direction of each image represents time, and the vertical direction represents frequency. The high and low pixel values ​​of the image represent the amplitude corresponding to the frequency. The Φ-OTDR data set contains a total of 33,600 time-frequency diagrams. There are 8 types of events, including knocking, shaking, friction, climbing, walking, falling heavy objects, piezoelectric ceramic vibration, and no vibration. Each event contains 4,200 images. Randomly select one image of each of the 8 types of data, such as Figure 4 shown. Figure 4Among them, (a) is a diagram showing the effect of a heavy object falling; (b) is a diagram showing the effect of friction; (c) is a diagram showing the effect of noise; (d) is a diagram showing the effect of piezoelectric ceramic vibration; (e) is a diagram showing the effect of knocking; (f) is a diagram showing the effect of shaking; (g) is a diagram showing the effect of climbing; and (h) is a diagram showing the effect of walking. This embodiment divides Φ-OTDR events into two categories: intrusion events and non-intrusion events. Intrusion events include knocking, shaking, climbing, walking, and heavy object falling. Knocking and shaking present intermittent changes in amplitude over time. Knocking lasts for a short time when it occurs, while shaking lasts for a long time. At the high-frequency end, the amplitude distributions of climbing and walking are similar. At the low-frequency end, the amplitude of walking is significantly lower than that of climbing. The amplitude span of a heavy object falling over time is obvious, and the duration of the event is relatively short. Non-intrusive events include friction, noise, and piezoelectric ceramic vibration. Noise and friction events exist over time, and their amplitudes are significantly lower than those of other events. Through Figure 5 It can be seen that the changes in time and frequency of non-intrusive events have obvious distinguishing features, while the distinguishing features of intrusive events are not obvious enough. The characteristics of knocking and shaking are similar, and the characteristics of walking and climbing are similar. It is difficult to manually extract features and manually label Φ-OTDR events. In order to solve this problem, it is necessary to use convolutional neural networks to adaptively extract features.

[0070] 2. Perform data preprocessing;

[0071] The set of time-frequency graphs of each category of the public dataset obtained above is randomly divided into training set and validation set in a ratio of 7:3. The convolutional neural network used is designed and trained for images of size 224×224. In order to avoid incompatibility issues, the images are scaled to a smaller size at the same time, which can reduce the amount of calculation, training time and memory usage. In addition, images of size 224×224 are also sufficient to retain the key information of the image in order to train an effective classifier. Therefore, the size of each time-frequency graph is preset to 224×224. In order to better verify the superiority of the AsymConv-Densenet network, all networks in this experiment use the same dataset division ratio and the same data preprocessing process.

[0072] 3. Choose pytorch1.11.0 as the deep learning framework. All experiments are conducted on a workstation equipped with an NVIDIA GeForce RTX 2080 GPU with 64GB memory and CUDA version 11.3. For all networks in this experiment, the learning rate is set to 0.0001, the batch size of the network is set to 32, and the epoch is set to 30. The composition of each layer and the input size of the AsymConv-Densenet network structure are shown in the table.

[0073] Table 1

[0074]

[0075]

[0076] 4. In this implementation, three standard indicators, namely, accuracy, precision, recall and F1 score, are used to evaluate the performance of the results of the intrusion event identification method. They are calculated based on TP (true positive), FP (false positive), TN (true negative) and FN (false negative) in the confusion matrix, as shown below:

[0077]

[0078] Taking knock events as an example, the precision rate is the proportion of actual knock events in the data set predicted as knock events; the recall rate is the proportion of events predicted as knock events in the data set that are actually knock events; the F1 score (the weighted harmonic mean of the precision rate and the recall rate) is the combination of the precision rate and the recall rate; the accuracy is the proportion of correctly predicted events in the entire data set.

[0079] 5. Effectiveness analysis of asymmetric convolution blocks;

[0080] To prove that adding asymmetric convolution blocks to the Densenet network can better improve the overall performance of the network, this experiment compares the average precision, average recall and average F1 score of the AsymConv-Densenet network, Densenet network, SE-Densenet network and CBAM-Densenet network at maximum accuracy. The Densenet network is the backbone network, the SE-Densenet network adds squeeze-and-excitationblock to each conversion block, and the CBAM-Densenet network adds Convolutional BlockAttentionModule to each dense block.

[0081] The average precision, average recall and average F1 score of the four networks at maximum average precision are shown in Table 2.

[0082] Table 2

[0083] Serial number method Maximum accuracy Average precision Average recall Average F1 score 1 Densenet Network 0.9546 0.9553 0.9546 0.9549 2 SE-Densenet Network 0.9581 0.9579 0.9581 0.9579 3 CBAM-Densenet Network 0.9594 0.9595 0.9594 0.9593 4 AsymConv-Densenet network 0.9652 0.9653 0.9653 0.9653

[0084] The SE-Densenet network, CBAM-Densenet network, and AsymConv-Densenet network all improve the maximum accuracy, average precision, average recall, and average F1 score of the Densenet network by adding optimization modules. The squeeze-and-excitation block in the SE-Densenet network only considers the attention in the channel dimension and ignores the attention in the spatial dimension. Therefore, the squeeze-and-excitation block has limited improvement on the overall performance of the Densenet network. Compared with the Densenet network, the maximum accuracy of the SE-Densenet network is improved by 1.06%, the average precision is improved by 1.00%, the average recall is improved by 1.07%, and the average F1 score is improved by 1.04%. In the CBAM-Densenet network, the Convolutional BlockAttentionModule integrates the channel attention mechanism and the spatial attention mechanism, which makes up for the problem that the squeeze-and-excitation block cannot capture the attention in the spatial dimension, but the computational cost is high and additional attention parameters need to be introduced, resulting in overfitting problems. Compared with SE-Densenet, the maximum accuracy of CBAM-Densenet is improved by 0.13%, the average precision is improved by 0.16%, the average recall is improved by 0.13%, and the average F1 score is improved by 0.14%. In AsymConv-Densenet, although the asymmetric convolution block only considers the features of the spatial dimension, it captures the changing features of the image in different directions through convolution kernels of different sizes, and compensates for the overfitting problem of Convolutional Block AttentionModul through batch normalization. Compared with CBAM-Densenet, the maximum accuracy of AsymConv-Densenet is improved by 0.58%, the average precision is improved by 0.58%, the average recall is improved by 0.59%, and the average F1 score is improved by 0.60%. Compared with squeeze-and-excitation block and Convolutional BlockAttentionModul, asymmetric convolution blocks have the best overall performance improvement on Densenet.

[0085] like Figure 5 As shown in the figure, in order to verify that the asymmetric convolution block can better improve the performance of the Densenet network in identifying intrusion events and reduce missed reports, this experiment analyzes the confusion matrix of the four networks with the highest accuracy. Figure 5In the figure, (a) is the confusion matrix of the CBAM-Densenet network; (b) is the confusion matrix of the SE-Densenet network; (c) is the confusion matrix of the Densenet network; (d) is the confusion matrix of the AsymConv-Densenet network. In the knocking event, the recall rate of the AsymConv-Densenet network is 99.84%. Compared with the Densenet network, the SE-Densenet network, the CBAM-Densenet network, and the AsymConv-Densenet network, the recall rates are increased by 0.96%, 0.88%, and 0.16%, respectively. In the shaking event, the recall rate of the AsymConv-Densenet network is 90.15%. Compared with the Densenet network, the SE-Densenet network, the CBAM-Densenet network, and the AsymConv-Densenet network, the recall rates are increased by 4.20%, 0.87%, and 2.38%, respectively. In the climbing event, the recall rate of the AsymConv-Densenet network is 90.87%, which is 0.72%, 3.97%, and 1.11% higher than that of the Densenet network, SE-Densenet network, and CBAM-Densenet network, respectively. In the walking event, the recall rates of the AsymConv-Densenet network and SE-Densenet network are both 98.33%, which are 2.30% and 1.51% higher than that of the Densenet network and CBAM-Densenet network, respectively. Therefore, in the four intrusion events of knocking, shaking, climbing, and walking, the asymmetric convolutional block can better improve the recognition performance of the Densenet network, reduce more missed reports, and stop the occurrence of intrusion events in time.

[0086] 6. Performance analysis of AsymConv-Densenet network;

[0087] like Figure 6 As shown, Figure 2 This is a comparison chart of the accuracy curves of the AsymConv-Densenet network of this implementation mode and four networks; among them, the Alexnet network, the Googlenet network and the Resnet-18 network are convolutional neural networks, and the Swintransformer network is an improved and optimized Transforme network.

[0088] The Alexnet network increases the depth of the network by adding convolution layers, but the network will produce gradient vanishing, resulting in insufficient feature transfer, which affects the accuracy of the Alexnet network in identifying Φ-OTDR events. The AsymConv-Densenet network alleviates the gradient vanishing problem of the Alexnet network through dense blocks in the network. From the curve changes, it can be seen that with the increase of the number of iterations, the accuracy of the AsymConv-Densenet network is completely higher than that of the Alexnet network. When the number of iterations reaches 30, the accuracy of the AsymConv-Densenet network is improved by 4.59% compared with the Alexnet network. In order to improve the feature extraction capability, the AsymConv-Densenet network and the Googlenet network both use convolution kernels of different sizes for convolution operations. When convolution kernels of different sizes are used for convolution operations, the amount of calculation in the training and verification stages of the Googlenet network is the same, while the AsymConv-Densenet network can simplify the amount of calculation in the verification stage. From the curve changes, it can be seen that the accuracy of the AsymConv-Densenet network is basically higher than that of the Googlenet network. When the iteration reaches 30 times, the accuracy of the AsymConv-Densenet network is improved by 1.51% compared with the Googlenet network. The Resnet network establishes a short-circuit connection between the front layer and the back layer. In order to solve this problem, the AsymConv-Densenet network establishes dense connections between all the front layers and the back layers. Both are used to solve the gradient disappearance problem. Compared with the short-circuit connection of the Resnet-18 network, the dense connection of the AsymConv-Densenet network allows the gradient to be transmitted more directly to the shallow network, which can more effectively alleviate the gradient disappearance. According to the changes in the curve, as the number of iterations increases, the accuracy of the AsymConv-Densenet network is better than that of the Resnet-18 network. When the number of iterations reaches 30, the accuracy of the AsymConv-Densenet network is improved by 2.20% compared with the Resnet-18 network. The Swintransformer network abandons the RNN and CNN structures, causing the network to lose the ability to capture local features. The AsymConv-Densenet network improves the ability to capture local features through asymmetric convolution blocks. When the number of iterations is close to 30, the accuracy of the AsymConv-Densenet network is improved by 2.32% compared to the Swintransformer network.

[0089] By comparing the maximum accuracy of the five networks, the maximum accuracy of AsymConv-Densenet network, Alexnet network, Resnet-18 network, Googlenet network, and Swintransformer network are 96.53%, 92.87%, 94.59%, 94.99%, and 94.31%, respectively. The maximum accuracy of AsymConv-Densenet network is 3.66%, 1.94%, 1.54%, and 2.22% higher than that of Alexnet network, Resnet-18 network, Googlenet network, and Swintransformer network, respectively. This experiment selects the maximum accuracy of the five networks and analyzes the recall rates of 8 Φ-OTDR events, as shown in Table 3.

[0090] Table 3

[0091]

[0092]

[0093] Heavy object dropping events are not typical intrusion events, and they rarely occur, posing little threat to the perimeter security of power optical cables. The recall rate of this network is 99.36%, which is 0.24% lower than that of Resnet-18 and Googlenet networks. It still maintains a high recall rate to avoid underreporting when it occurs. Friction and piezoelectric ceramic vibration are often used to simulate severe weather such as strong winds and rain, and are non-intrusive events. The recall rates of friction and piezoelectric ceramic vibration of this network are 95.87% and 99.76% respectively, which are better than the other comparison networks. In environmental noise events, the recognition recall rate of the AsymConv-Densenet network is 98.01%, which is only higher than the Alexnet network and the Resnet-18 network. Environmental noise simulates the environmental conditions under normal weather and is a non-intrusive event. Therefore, low recall will not affect the identification of intrusion events.

[0094] The characteristics of shaking and knocking data images are that the amplitude is generated intermittently in the time domain. Compared with knocking, the amplitude of shaking is generated intermittently. In the frequency domain, the amplitude distribution of the two is similar. The Alexnet network, Resnet-18 network, and Googlenet network only contain symmetric convolution kernels, while the AsymConv-Densenet network contains not only symmetric convolution kernels, but also asymmetric convolution kernels in the horizontal and vertical directions. The AsymConv-Densenet network not only retains the advantages of symmetric convolution kernels in extracting local image features, but also can capture the lateral and longitudinal change characteristics of the data image. The Swin transformer network divides the input image into small blocks and uses a moving window method for information interaction, which can capture the changing characteristics of the image, but has high requirements for image resolution. Experiments show that for this experimental data set, the advantages of this network cannot be reflected. The recall rate of knocking events of the AsymConv-Densenet network is 99.84%, which is 1.51%, 0.56%, 0.40%, and 1.04% higher than those of the Alexnet network, Resnet-18 network, Googlenet network, and Swin transformer network, respectively. The recall rate of shaking events of the AsymConv-Densenet network is 90.15%, which is 2.38%, 7.38%, 5.63%, and 2.46% higher than those of the Alexnet network, Resnet-18 network, Googlenet network, and Swin transformer network, respectively. Therefore, for the recognition of two typical intrusion events, shaking and knocking, the recall rate of this network is better than that of other networks.

[0095] As a typical intrusion event in the perimeter security of power optical cables, the data image characteristics of climbing and walking are relatively similar and difficult to distinguish with the naked eye. Therefore, it is necessary to accurately identify such events and issue timely warnings to prevent such events from happening and reduce losses. The AsymConv-Densenet network uses convolution kernels of different sizes in asymmetric convolution blocks to differentially extract the features of the two types of events, and each layer in the dense block retains the output of all previous layers. Compared with other networks, it obtains more detailed features and can better distinguish the two. According to Table 3, the recall rate of AsymConv-Densenet network in climbing events is 90.87%, which is 0.55%, 2.14%, 5.32%, and 7.78% higher than that of Alexnet network, Resnet-18 network, Googlenet network, and Swintransformer network, respectively. In the detection and recognition of walking events, the recall rate of AsymConv-Densenet network is 98.33%, which is 2.46%, 3.26%, 0.72%, and 3.81% higher than that of Alexnet network, Resnet-18 network, Googlenet network, and Swintransformer network, respectively. In the recognition of climbing and walking events, the recall rate of AsymConv-Densenet network is higher than that of other networks, and AsymConv-Densenet network has obvious advantages.

[0096] Through the above analysis, compared with Alexnet, Resnet-18, Googlenet and Swintransformer networks, AsymConv-Densenet can more accurately identify four common intrusion events, namely knocking, shaking, climbing and walking, in the perimeter security of power optical cables, and issue timely alarms to reduce damage to power facilities.

[0097] In this embodiment, the asymmetric convolution block is combined with the Densenet network and applied to the Φ-OTDR vibration event recognition. The network uses dense connections and asymmetric convolution blocks to extract more distinguishing features, obtain a higher recall rate, and reduce missed reports. In order to explain and estimate the effectiveness and performance of the AsymConv-Densenet network, a series of experiments were conducted. The experimental results show that after the asymmetric convolution block optimizes the Densenet network, the overall performance is better than the Densenet network, SE-Densenet network, and CBAM-Densenet network. In several types of simulated events of power optical cable perimeter security, the network in this embodiment has the highest recall rate for identifying climbing, shaking, walking, and knocking. In the comparison of convolutional neural networks and SwinTransformer networks, the overall performance and recall rate of AsymConv-Densenet network in identifying climbing, walking, shaking and knocking are higher than those of Alexnet network, Resnet-18 network, Googlenet network and SwinTransformer network. Therefore, AsymConv-Densenet network is more suitable for application in power optical cable perimeter security intrusion incidents than Alexnet network, Resnet-18 network, Googlenet network and SwinTransformer network.

[0098] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. A Φ-OTDR power optical cable perimeter security intrusion event identification method based on AsymConv-Densenet network, characterized by: The method is implemented by the following steps: Step 1: Convert the collected Φ-OTDR signals of different events into time-frequency diagrams to construct a Φ-OTDR signal vibration event image dataset; Step 2: construct an AsymConv-Densenet network model, and train the network model to obtain a trained network model; The network model consists of a convolutional layer, a dense connection module, a conversion module, a feature extraction module, a global pooling layer and a fully connected layer; The dense connection module includes dense block 1, dense block 2 and dense block 3; The conversion module includes conversion block 1, conversion block 2 and conversion block 3; The feature extraction module includes an asymmetric convolution block 1, an asymmetric convolution block 2 and an asymmetric convolution block 3; Step 3: Use the trained network model to identify the Φ-OTDR signal and realize the identification of the vibration event of the Φ-OTDR signal; the specific identification process is as follows: The time-frequency graph of the Φ-OTDR signal is input into the network model, and the frequency components, time-varying features and texture features in the time-frequency graph are extracted through the convolution layer to obtain a feature graph; the feature graph sequentially captures the global feature graph of different vibration events through dense block 1, conversion block 1, asymmetric convolution block 1, dense block 2, conversion block 2, asymmetric convolution block 2, dense block 3, conversion block 3 and asymmetric convolution block 3, and the global feature graph is averaged by the pixel values ​​of each global feature graph through the global pooling layer to obtain a feature vector of fixed length; After the feature vector passes through the fully connected layer, SoftMax is used to output the probability corresponding to each label in the Φ-OTDR data, and the vibration event label with the maximum probability is used as the output result, finally realizing the recognition of the vibration event of the Φ-OTDR signal.

2. The Φ-OTDR power optical cable perimeter security intrusion event identification method based on AsymConv-Densenet network according to claim 1 is characterized in that: In step 1, the collected Φ-OTDR signals of different events are converted into time-frequency diagrams through pre-emphasis, windowing and framing, spectral subtraction and short-time Fourier transform, and then constructed into a Φ-OTDR signal vibration event image dataset.

3. The Φ-OTDR power optical cable perimeter security intrusion event identification method based on AsymConv-Densenet network according to claim 1 is characterized in that: In step 2, the three asymmetric convolution blocks of the feature extraction module each include three parallel layers, and the convolution kernel sizes are 3×3 convolution kernel, 1×3 convolution kernel and 3×1 convolution kernel respectively; in the process of training the network model, the feature map passes through the asymmetric convolution block 1, and the feature map obtained by the conversion block 1 is subjected to multi-scale analysis through the 1×3 convolution kernel, the 3×1 convolution kernel and the 3×3 convolution kernel to extract the detailed information to distinguish the vibration events, and the features are fused through splicing and batch normalization; The weight parameters of the trained asymmetric convolution block 1 remain unchanged. The 1×3 convolution kernel, the 3×1 convolution kernel, and the 3×3 convolution kernel are added together, and the obtained mixed 3×3 convolution is used for feature extraction and batch normalization. The asymmetric convolution block 1 does not change the size and dimension of the feature map output by the conversion block 1.

4. The Φ-OTDR power optical cable perimeter security intrusion event identification method based on AsymConv-Densenet network according to claim 1 is characterized in that: Each layer in each dense block in the dense connection module is a nonlinear combination function, and the nonlinear combination function consists of a normalization layer, a rectified linear unit and a 3×3 convolution layer; Each layer in a dense block can process all the features obtained by the previous layer. For the kth layer, its forward layer has x0,x1,x2,...,x k-1 Combination function, the formula for dense connection is: X k =H k ([x0,x1,x2,…,x k-1 ]) where [x0,x1,x2,...,x k-1 ] is the concatenation of the feature layers in the channel dimension, H k () is a combined operation.

Images