Underground optical cable vibration multi-event identification method based on SE-DMRCB multi-scale feature extraction

Abstract

An underground optical cable vibration multi-event identification method based on SE-DMRCB multi-scale feature extraction solves the problems of nonlinearity, long and short term dependence, frequency domain feature overlapping and the like of underground optical cable vibration signals, and solves the problems that a vibration event identification model depends on single-scale feature extraction, so that the effect on complex and diverse event processing performance is poor and the like. According to the method, the features of different receptive fields are extracted at the same time by using a parallel multi-scale convolution architecture, deep feature learning is performed through 8 layers of stacked multi-scale residual convolution blocks, and each convolution block is realized by integrating key technologies such as multi-scale convolution, an SE attention mechanism and residual connection. According to the method, an event result is finally identified through a residual error convolution feature learning and classification decision module of data feature projection and multi-scale feature extraction in sequence. According to the method, the recognition accuracy of multiple events is optimal, and the application real-time performance requirement is met.

Description

Technical Field

[0001] This invention relates to the fields of multi-event recognition of vibration in underground optical cables and deep learning technology, specifically to a method for multi-event recognition of vibration in underground optical cables based on a deep multi-scale residual convolutional block incorporating Squeeze-and-Excitation Attention Mechanism (SE-DMRCB) that extracts multi-scale features. Background Technology

[0002] Underground fiber optic cables serve as the nervous system of urban communication networks, carrying critical functions such as power communication and public network communication. Statistics show that the total length of underground fiber optic cables worldwide exceeds one million kilometers, and the density of underground fiber optic cables continues to increase. However, underground fiber optic cables face diverse safety threats during operation, including construction excavation, vehicle traffic, and geological activity. These threats can lead to cable damage, communication outages, and even safety accidents, causing significant economic losses.

[0003] Distributed fiber optic sensing technology, especially phase-sensitive optical time-domain reflectometry (Φ-OTDR), leverages its advantages of high sensitivity, long-distance detection, high reliability, and low cost. Using optical fiber as the sensing medium, it detects various vibration events occurring along the fiber in real time and collects vibration signals. Analysis of these signals can then be used to provide early warnings of destructive actions. This technology has found applications in the field of security monitoring.

[0004] In recent years, researchers have begun to conduct experiments on buried optical cables under simulated conditions, using the Φ-OTDR system to collect signals of typical events such as walking, digging, and knocking, verifying the feasibility of the technology in complex geological and noise environments, and laying the foundation for the safety monitoring of buried optical cables. Optical fiber vibration event identification methods are mainly divided into two categories: machine learning methods and deep learning methods.

[0005] Machine learning methods primarily rely on manual feature extraction. Jia et al. used the K-Nearest Neighbors (KNN) algorithm to extract time and frequency domain features of signals and employed a Support Vector Machine (SVM) classifier to identify watering, knocking, climbing, pressing, and spurious events. Cao et al. extracted 32 features from Φ-OTDR signals and used an SVM model to identify background noise, digging, knocking, watering, shaking, and walking events. Wang et al. utilized a random forest classifier to learn the features of time-domain interference signals, effectively identifying watering, knocking, pressing, and interference-free events. However, these methods struggle to obtain distinguishable features manually when faced with events with similar vibration characteristics (such as digging and pickaxe blows), limiting their recognition performance. Deep learning methods, through adaptive feature extraction, can better distinguish similar events. Shi et al. (Shi Y, Chen J, Kang X, et al. An Φ-OTDR event recognition method based on Transformer[C] / / 2023 21st International Conference on Optical Communications and Networks (ICOCN). IEEE, 2023: 1-3.) proposed a method combining Transformer and Φ-OTDR to identify events such as background noise, cycling, washing, patting, and walking. Zhu et al. (Zhu C, Pu Y, Yang K, et al. Distributed optical fiber intrusion detection by image encoding and SwinT in multi-interference environment of long-distance pipeline[J]. IEEE Transactions on Instrumentation and Measurement, 2023, 72: 1-12.) introduced a method combining Swin Transformer with two-dimensional vibration signal images to identify environmental signals such as farmland, villages, and mountainous areas, as well as strong interference signals such as highways and railways. Wu et al. proposed the 1DCNN-BiLSTM method, using 1DCNN to extract the spatiotemporal features of the signal and BiLSTM to capture the spatial relationships between signal nodes.Tian et al. (Tian M, Dong H, Yu K. Attention based Temporal convolutional network for ϕ-OTDR event classification[C] / / 2021 19th International Conference on Optical Communications and Networks (ICOCN). IEEE, 2021: 1-3.) combined the attention mechanism with CNN and proposed the ATCN-BiLSTM model to identify events such as non-threatening backgrounds, climbing fences, and rain.

[0006] In summary, while existing optical cable vibration event recognition models have achieved some success in specific tasks, they still struggle to achieve both high accuracy and robustness in complex scenarios. The fundamental reasons are: single-scale convolutional structures cannot simultaneously capture multi-timescale features of vibration signals; insufficient model response to key channels weakens important features; gradient decay occurs during deep structure training; and the imbalance of classes in actual engineering data further affects the recognition accuracy of minority classes. Therefore, it is necessary to design a recognition model that can integrate multi-scale features, possess channel adaptability, and maintain stable gradient propagation in deep structures to achieve high-precision recognition of underground optical cable vibration signals in real-world scenarios. Summary of the Invention

[0007] To address the shortcomings of existing technologies in identifying vibration signals from underground optical cables, such as insufficient multi-scale feature extraction, inadequate utilization of channel features, and gradient vanishing in deep networks, which lead to poor performance in handling complex and diverse events, this invention provides a multi-event identification method for vibration of underground optical cables based on SE-DMRCB multi-scale feature extraction.

[0008] A method for identifying multiple vibration events of underground optical cables based on SE-DMRCB multi-scale feature extraction includes a data preprocessing module, a data feature projection module, a deep residual convolution feature learning module for multi-scale feature extraction, and a classification decision module.

[0009] The data preprocessing module is used to perform sliding window segmentation and standardized fractionation on the original vibration signal to form an experimental dataset.

[0010] The data feature projection module extracts multi-timescale initial features of the vibration signal through parallel multi-scale convolution.

[0011] Multi-scale feature extraction deep residual convolution feature learning module: Uses 8 stacked residual convolution blocks to extract deep features and obtain deep feature representation of vibration data;

[0012] Each residual convolutional block includes multi-scale convolutional layers, standard convolutional layers, SE attention mechanism layers, and residual connection layers;

[0013] The classification decision module achieves the final identification of 8 types of vibration events through global average pooling and a fully connected network.

[0014] The beneficial effects of this invention are:

[0015] In the identification method described in this invention, the boundary features of events in the signal are not truncated by using a 50% overlap rate of the sliding window, thus preserving the edge features and temporal continuity of the events. This effectively increases the number of training vibration signal samples and improves the generalization ability of the model. Furthermore, Z-score standardization eliminates amplitude differences under different acquisition conditions, enhancing the model's adaptability to diverse events.

[0016] The identification method described in this invention achieves complementary fusion of features at different time scales and adaptive enhancement of key channel features through the collaborative design of a data feature projection module and a deep residual convolutional feature learning module for multi-scale feature extraction. Simultaneously, residual connections effectively alleviate the gradient vanishing problem. Experimental results show that the model achieves a 99.84% accuracy rate in identifying eight types of underground vibration events in a real substation scenario. Compared with other comparative models, it exhibits superior performance in accuracy, real-time performance, and robustness. Furthermore, ablation experiments further verify the rationality of the model architecture design and the synergistic effect between the modules. Attached Figure Description

[0017] Figure 1 This is a general framework diagram of the model of the present invention;

[0018] Figure 2 Construction diagram of residual convolutional feature learning module for multi-scale feature extraction;

[0019] Figure 3 This is a schematic diagram of the classification decision module;

[0020] Figure 4 Comparison graphs are provided for the ablation experiments; among them, (a), (b), (c), and (d) are graphs showing the loss rate and accuracy of the model (Baseline) and the ablation variant on the training set / validation set.

[0021] Figure 5 This is a confusion matrix and a graph showing the overall performance of the model (Baseline) of this invention.

[0022] Figure 6 The confusion matrix and overall performance of the ResNet-1D model used in the comparison are shown in the figure.

[0023] Figure 7 The confusion matrix and overall performance of the DenseNet-1D model used in the comparison are shown in the figure.

[0024] Figure 8 The confusion matrix and overall performance of the MobileNet-1D model used in the comparison are shown in the figure.

[0025] Figure 9 The confusion matrix and overall performance of the SignalTransformer model used in the comparison are shown in the figure.

[0026] Figure 10 The confusion matrix and overall performance of the BiLSTM model used in the comparison are shown in the figure.

[0027] Figure 11 The diagram shows the confusion matrix and overall performance of the VanillaCNN model used in the comparison. Detailed Implementation

[0028] Specific Implementation Method 1: Combination Figures 1 to 3 This embodiment describes a method for identifying multiple vibration events in underground optical cables based on SE-DMRCB multi-scale feature extraction. This method is implemented through the following steps:

[0029] Step S1. In the scenario of laying underground optical cables, for areas where vibrations frequently occur due to human threats, underground power optical cables with high safety requirements are selected. A Φ-OTDR system is deployed and the optical cable is used as the vibration sensing medium. Human vibration behavior is designed under real conditions, and the original vibration event signals of the underground optical cable are collected experimentally.

[0030] Step S2. Construct a multi-event identification model for underground optical cable vibration based on SE-DMRCB multi-scale feature extraction; the network architecture of this model is as follows: Figure 1 As shown, eight types of vibration events were performed in the experimental area of ​​the buried optical cable. After the vibration signal of the buried optical cable was collected using the Φ-OTDR system, it was input into the model. The data was processed by four main modules in sequence to identify the types of vibration events.

[0031] The model includes a data preprocessing module, a data feature projection module, a deep residual convolution learning module for multi-scale feature extraction, and a classification decision module;

[0032] The data preprocessing module is used to perform sliding window segmentation and standardized score (Z-score) on the original vibration signal to form an experimental dataset.

[0033] In this embodiment, the data preprocessing module preprocesses the original vibration signal in step S1 as follows:

[0034] Various types of vibration events were simulated above the buried experimental optical cable, and vibration signals sensed by the buried optical cable were acquired using a Φ-OTDR. The acquired buried vibration signals were then processed by the data preprocessing module in the model. Since the acquired vibration signals are long-term, continuous vibration signals, directly inputting the raw signals would make it difficult for the model to learn local features and would lead to gradient instability. Therefore, the raw vibration signal x∈R was preprocessed. 1×n The signal is processed, where n is the signal length (n=1100000). A sliding window strategy is used to segment the long time-series signal, and the window parameters are configured as follows: window size W. size =1024 sampling points, overlap rate r overlap =50%, Step size S stride =512 sampling points; the formula for calculating the generated window is:

[0035] +1

[0036] In this embodiment, the response intensity of the Φ-OTDR system to different underground optical cable vibration events is affected by factors such as the burial depth of the optical cable, the laying scenario, the location and distance of the event, the cable type, fiber loss, and soil medium, resulting in significant differences in signal amplitude distribution. Therefore, Z-score normalization is performed on the data of each window.

[0037]

[0038] Where μ and σ are the mean and standard deviation of the signal, respectively.

[0039] During data preprocessing, a 50% overlap sliding window ensures that the boundary features of events in the signal are not truncated, preserving event edge features and temporal continuity. This effectively increases the number of training vibration signal samples and improves the model's generalization ability. Z-score standardization eliminates amplitude differences under different acquisition conditions, enhancing the model's adaptability to diverse events. The preprocessed vibration event data for the eight categories is then input into the next layer, the data feature projection module.

[0040] The data feature projection module extracts multi-timescale initial features of the vibration signal through parallel multi-scale convolution; the specific implementation process is as follows:

[0041] The data feature projection module described in this embodiment receives preprocessed data on eight types of vibration events and projects data features onto them. Because the vibration signals of different vibration event types in buried optical cables differ significantly across time scales—for example, knocking events exhibit short-duration pulse characteristics, walking events show periodic vibration characteristics, and knocking and digging events are impact actions on buried optical fibers, exhibiting similar signal characteristics and short-duration pulses in the time domain—a single convolutional kernel cannot simultaneously capture these multi-time-scale features. Therefore, a parallel multi-scale convolutional architecture is adopted to simultaneously extract features from different receptive fields. This module contains three one-dimensional convolutional branches of different sizes, using 3×1, 5×1, and 7×1 convolutional kernels respectively, with each branch outputting 21 to 22 feature channels. The multi-scale feature fusion process can be represented as:

[0042] ;

[0043] ;

[0044] Where ⊕ represents the channel dimension concatenation operation, k∈{3,5,7} is the convolution kernel size, L is the feature map length, and C out The number of output channels for each branch. Three different sized convolutional kernels capture high-frequency local detail features, mid-frequency transition features, and low-frequency global trend features, respectively. By concatenating features, complementary fusion of multi-scale information of vibration events is achieved, providing rich initial feature maps of vibration signals for the deep residual convolutional feature learning module for subsequent multi-scale feature extraction.

[0045] The deep residual convolution feature learning module for multi-scale feature extraction uses a deep stack of eight multi-scale residual convolution blocks (DMRCB) for deep feature extraction. Each residual convolution block includes a multi-scale convolutional layer, a standard convolutional layer, an SE attention mechanism layer, and a residual connection layer.

[0046] In this embodiment, the initial feature mapping of the buried optical cable vibration signal was completed through the data feature projection module. However, due to the multi-timescale, non-stationary, and high-noise characteristics of the Φ-OTDR signal, shallow convolution alone is insufficient to fully extract discriminative multi-event features. Therefore, a deep residual convolutional feature learning module based on SE-DMRCB multi-scale feature extraction was designed. This module consists of 8 stacked MRCB modules, and an SE attention mechanism is added to the multi-scale residual convolutional block. This mechanism can dynamically adjust the weights of different channels based on the global statistical information of the feature distribution, enhancing the network's attention to key channels and suppressing redundant or noisy channels, thereby improving the model's feature representation ability and generalization performance. Figure 2 As shown, each residual convolutional block Block i (i=1,…,8) adopts a design of "multi-scale convolution + standard convolution + SE attention mechanism + residual connection" to achieve hierarchical feature abstraction from local to global and from low-level to high-level. The multi-scale residual convolutional block includes the following structure:

[0047] The multi-scale convolutional layer receives the feature information stitched together by the data feature projection module and uses it as the input to the first residual convolutional block. Utilizing the rich initial feature representation of the buried optical cable vibration signal, it performs hierarchical scale refinement of the initial feature representation, further filtering and combining features. By employing parallel multi-scale convolution within the residual convolutional block, it continuously maintains a multi-scale receptive field in the feature space, avoiding the limitation of traditional single convolutional kernels that can only capture fixed-scale features. The output features of each multi-scale convolutional layer are:

[0048] =Concat(Conv1D 3×1 (F (i−1) ),Conv1D 5×1 (F (i−1) ),Conv1D 7×1 (F (i−1) ))

[0049] In the formula, F (i-1) For the output features of the (i-1)th layer, the number of output channels for each convolutional kernel is dynamically allocated based on the total number of channels in the current layer, ensuring that the number of channels after concatenation remains consistent with the input. This design avoids information loss caused by the homogenization of the receptive field in deep networks, enabling the network to capture multi-scale temporal features at different depths.

[0050] After the multi-scale convolutional layers, a light Dropout (p=0.1) is introduced to regularize the concatenated features, preventing overfitting during the feature extraction stage.

[0051] The standard convolutional layer regularizes the concatenated feature information, and then introduces a standard 3×1 convolutional layer for feature integration and nonlinear transformation. The features output by the standard convolutional layer are:

[0052] = Conv1D 3×1 (Dropout) , p=0.1))

[0053] The kernel size of 3×1 in the standard convolutional layer was determined through continuous trial and adjustment using a Bayesian optimization algorithm, i.e., by:

[0054] The optimal convolutional kernel size is found to fully capture the interaction relationships between multi-scale features. By deeply fusing the three different scale features extracted by multi-scale convolution, the interaction relationships between scales are learned. When it is necessary to change the number of feature channels, this convolutional layer plays the role of dimension mapping. Moreover, this convolutional layer increases the non-linear expression depth of the network. Through additional convolutional operations, high-frequency pulse features of impact events such as knocking, digging, and picking in vibration signals, as well as temporal rhythm features of periodic events such as walking and motorcycling, are further extracted. The shallow amplitude-time variation features are transformed into high-level discriminative features that can distinguish different vibration event types.

[0055] Batch normalization and ReLU activation are applied to the output features of the standard convolutional layer, resulting in the following output features:

[0056] =ReLU(BatchNorm1d( ));

[0057] Batch normalization can accelerate convergence and stabilize the training process by reducing internal covariate bias, while the ReLU activation function introduces nonlinearity to enhance the model's expressive power.

[0058] The SE attention mechanism learns the interdependencies between feature channels processed by batch normalization and ReLU activation function in the previous layer, adaptively recalibrating channel feature weights to enhance the response to high-frequency pulse feature channels of impact events such as knocking and picking, and temporal rhythm feature channels of periodic events such as walking and motorcycling, while suppressing redundant channels such as background noise and soil interference. The implementation process of the SE attention mechanism layer is as follows:

[0059] First, global descriptors of the vibration feature map in the time dimension are obtained through global average pooling (GAP):

[0060] z=GAP( )=(1 / L) ∈R C ;

[0061] in, Let z be the feature map of the i-th residual convolutional block after batch normalization and activation, where C is the number of channels, L is the feature map length, and z∈R. C This is the channel global descriptor vector.

[0062] Then, channel attention weights are learned through two fully connected layers. The first layer performs channel compression, and the second layer performs channel activation. The channel attention weights are as follows:

[0063] s=σ(W2 ReLU(W1 z));

[0064] In the formula, W1∈R （C / r）×C Let W2 be the weight matrix of the first fully connected layer, where W2 ∈ R. C×（C / r） Here is the weight matrix of the second fully connected layer, r is the compression ratio, and σ( ) is the Sigmoid activation function, which normalizes the attention weights to the [0,1] interval.

[0065] Finally, the learned channel attention weights s∈R C Recalibrate the original features through channel-by-channel multiplication:

[0066] ;

[0067] In the formula, ⊙ represents channel-by-channel multiplication. This is the fused feature after recalibration using the SE attention mechanism. The SE attention mechanism enhances key channels containing high-frequency pulse features of knocking and pickaxe strikes, as well as important channels capturing the temporal rhythmic feature patterns of events such as walking and vehicle running over, while suppressing redundant channels such as background noise and environmental vibrations. With only a small increase in parameters (2C² / r), compared to attention mechanisms like CBAM that require simultaneous modeling of both spatial and channel dimensions, the SE attention mechanism models vibration signal features only in the channel dimension, avoiding the overhead of spatial dimension calculations for the time series. Furthermore, the SE attention mechanism's computational complexity is low due to its global average pooling compression of the temporal dimension and its fully connected layer learning of channel weights. While effectively distinguishing similar vibration events (such as digging and pickaxe strikes, knocking and vehicle running over), it has a negligible impact on the model's inference speed in real-time monitoring scenarios using the Φ-OTDR system.

[0068] The feature information processed by the SE attention mechanism layer is used as input, and then further processed through downsampling and residual connections. The deep residual convolutional feature learning module for multi-scale feature extraction introduces max pooling layers in specific layers (Block 4 and Block 8) for downsampling, with a pooling kernel size of 2 and a stride of 2, by halving the feature map length. The downsampling operation gradually expands the temporal receptive field of the vibration signal, from capturing the instantaneous impact details of events such as knocking and picking, to learning the complete periodic vibration patterns of events such as walking and motorcycling, and then to establishing long-term dependencies of complex events such as digging and drilling. This achieves feature abstraction from local vibration details to global event semantics, while reducing computational complexity, enabling the network to capture the high-level discriminative features needed to distinguish different vibration event types.

[0069] In this embodiment, the recognition model introduces residual connections in each residual convolutional block to form a residual convolutional block. The original vibration signal feature representation from the previous layer's data feature projection module is transmitted to the output via a shortcut path and added to the main path features. The specific implementation of the residual connections varies depending on whether downsampling is performed, and can be categorized into two cases:

[0070] No downsampling layers (Blocks 1-3, 5-7), such as Figure 2 As shown, for residual convolutional blocks that do not undergo downsampling, the spatial dimension of the output features of the main path remains consistent with that of the input features, and the residual connections use the identity shortcut path:

[0071] ;

[0072] In the formula, For the output features of the (i-1)th layer, The main path outputs features, which are then processed through multi-scale convolutional layers, standard convolutional layers, batch normalization and activation functions, and an SE attention mechanism layer to extract a fused feature representation. This representation includes temporal magnitude features, multi-scale temporal features, frequency response features, and channel-weighted features. The shortcut path directly passes the input without any transformation.

[0073] ;

[0074] For multi-scale residual convolutional blocks containing downsampling layers, such as Figure 2 In Blocks 4 and 8, for residual convolutional blocks that undergo downsampling, the main path halves the spatial dimensions using max pooling after the SE attention mechanism layer. To ensure consistency in the dimensions of the added residuals, the shortcut path also needs to perform corresponding downsampling.

[0075] ;

[0076] ;

[0077] ;

[0078] In the formula, For the output features of the (i-1)th layer, The main path output features. Features sampled from the max pooling layer along the main path. This represents the features sampled by the max pooling layer under the shortcut path; kernel_size is the convolution kernel, stride is the stride; that is, the same max pooling operation as the main path is applied to the original vibration signal features output by the data feature projection module of the previous layer, ensuring that the two paths are aligned in the time dimension before element-by-element addition.

[0079] In this embodiment, the recognition model adopts a phased channel number design strategy: Blocks 1-2 maintain 64 channels, Blocks 3-4 have 128 channels, and Blocks 5-8 have 256 channels. Within each residual convolutional block, the multi-scale convolutional layer maintains the same number of channels as the input, and the transition in channel number is handled by the standard convolutional layer (3×1 convolutional kernel). Specifically, in Blocks 3 and 5, the standard convolutional layer maps the number of channels from 64 to 128 and from 128 to 256, respectively, to achieve the transition in channel number between different stages. However, the shortcut path of the residual connection jumps directly from the Block input to the output, bypassing all operations of the main path. Therefore, at the locations where the number of channels changes (Blocks 3 and 5), the shortcut path requires an independent channel matching mechanism, which uses a 1×1 convolution to map the number of channels of the input features to be consistent with the output of the main path, ensuring dimensional alignment when the residuals are added. This design allows residual connections within the same phase (Blocks 1-2, 3-4, 5-8) to achieve dimension matching using zero-parameter identity mapping or pooling operations. This avoids introducing an additional 1×1 convolutional layer in each block, reducing parameter count and computational overhead. Simultaneously, it provides the most direct gradient backpropagation path, effectively mitigating the gradient vanishing problem in deep networks. Through the stacking of eight residual convolutional blocks and residual connections, the network extracts vibration signal features layer by layer, from local waveform details to global event patterns, and from amplitude-time fundamental features to event type discrimination features in a hierarchical representation. Furthermore, gradients can be directly backpropagated to shallower layers via identity mapping, effectively alleviating the gradient vanishing problem in deep networks, accelerating network convergence, and maximizing the preservation of the original waveform signal features and event temporal features of the vibration signal input, thus enhancing feature reuse capabilities.

[0080] The classification decision module achieves the final identification of 8 types of vibration events through global average pooling and a fully connected network.

[0081] like Figure 3 As shown, the signal depth feature representation F(8)∈R of the last residual convolutional block (Block 8) in the residual convolutional feature learning module for obtaining multi-scale feature extraction is given by F(8)∈R. 256×256 Then, the classification decision module maps it to the probability distribution of 8 types of vibration events. First, it aggregates the time series dimensional information through global average pooling (GAP):

[0082] f global =GAP(F (8) )∈R 256 ;

[0083] In the formula, F (8)The output of the 8th residual convolutional block, compared to the flattening operation, demonstrates that the parameterless design of GAP significantly reduces the risk of overfitting (reducing the number of parameters by more than 98%), while also exhibiting natural robustness to the time shift of vibration signals. Subsequently, feature transformation and classification are performed through a two-layer fully connected network.

[0084] ;

[0085] h1=ReLU(BatchNorm1d(z1));

[0086] ;

[0087] Among them, f drop1 =Dropout(f global (p=0.3) represents heavy regularization of the input layer to prevent overfitting, f drop2 =Dropout(h1,p=0.15) performs light regularization on the hidden layer, b1 is the bias vector of the first layer, b2 is the bias vector of the second layer, and W1∈R 256×256 Let W2 be the weight matrix of the first fully connected layer, where W2 ∈ R. 8×256 Z is the weight matrix of the second fully connected layer. out The original scores are for each of the eight categories. Finally, the Softmax function is used to convert the original scores for each category into a probability distribution.

[0088] ;

[0089] Where z_i is the original score of the i-th category, which is the i-th element in the z_out vector.

[0090] The predicted output is the category with the highest probability. :

[0091] ;

[0092] Through the above design, an end-to-end mapping from the original vibration signal to the category probability distribution is completed: the vibration signal is subjected to multi-scale feature extraction to obtain a high-level discriminative feature representation, and the feature weights are adaptively recalibrated in the channel dimension through the SE attention mechanism. This strengthens the high-frequency pulse feature channel of transient impact events such as knocking and pickaxe hitting events, the temporal rhythm feature channel of periodic events such as walking and motorcycles, and the frequency response feature channel of different events, such as the high-frequency feature of electric drills, the low-frequency feature of car crushing, and the broadband feature of digging. At the same time, redundant channels such as background noise and medium propagation loss are suppressed. Finally, the classification decision module outputs the prediction results of 8 types of vibration events.

[0093] In this embodiment, a parallel multi-scale convolution structure is introduced into the data feature projection module, using three different sizes of convolution kernels: 3×1, 5×1, and 7×1, to capture high-frequency details, mid-frequency transitions, and low-frequency global features of vibration signals at different time scales, thereby achieving complementary fusion of multi-scale information.

[0094] In the deep feature learning stage, an 8-layer stacked residual convolutional block (SE-DMRCB) is constructed. Each convolutional block integrates the structure of "multi-scale convolution - standard convolution - SE attention mechanism - residual connection", realizing hierarchical feature abstraction from local to global. The SE attention mechanism strengthens the response of key features through adaptive recalibration of channel weights and works in conjunction with the residual connection, which not only improves the feature representation ability of the model, but also effectively alleviates the gradient vanishing problem.

[0095] By using the Bayesian optimization algorithm to automatically search for model parameters such as network depth, number of basic channels, convolutional kernel size, learning rate, and weight decay, the optimal configuration is obtained after 50 rounds of trials, ensuring that the model achieves a good balance between complexity and generalization ability.

[0096] In this implementation, a loss function and optimization strategy are designed.

[0097] The dataset after the data preprocessing module shows a significant class imbalance problem, as shown in Table 1, which presents the classification results of underground optical cable vibration events.

[0098] Table 1

[0099]

[0100] The background category accounts for 36.4%, and other categories account for approximately 8.8%. Category-weighted cross-entropy loss is used.

[0101] ;

[0102] Where N is the batch sample size, p i w is the predicted probability of the true class of sample i. yi =N total / (N classes ×N c ) represents the category weight, N total N represents the total number of samples. classes To represent the total number, N c Given the number of samples in class c, this strategy mitigates the model's bias towards the majority class by increasing the weight of the minority class and decreasing the weight of the majority class, thus promoting balanced learning across all classes. The optimizer chosen is AdamW, and the learning rate is l. r =3.94×10 −4 The weight decay λ = 4.14 × 10−3 The StepLR step learning rate decay strategy is adopted (decay by 50% every 20 epochs), and an early stopping mechanism (patience=15) is introduced during training. The batch size is 16, and training is carried out for 50 epochs until full convergence.

[0103] Specific Implementation Method Two: Combination Figure 4 and Figure 5 This embodiment is an example of the method described in Specific Embodiment One:

[0104] 1. Using a buried optical fiber cable of a substation as the experimental object, the experiment was conducted in an open grassland area next to the substation. The optical cable is 1.98 kilometers long, of which about 1.5 kilometers are buried underground. The cable is laid at a depth of about 2.95 meters and is protected by a conduit. The underground section of the optical cable was selected for the experiment. There are two roads around the experimental section, at a distance of 3 meters and 7 meters respectively. Pedestrians, bicycles and cars occasionally pass by on the two roads. During the experiment, the weather was sunny, the soil was dry, and there was little external interference, which is a typical urban underground optical cable environment.

[0105] 2. To collect signals from various underground vibration events, a 1-kilometer G652D optical fiber was used to connect the Φ-OTDR system to the optical cable. The pulse width was set to 32ns, the frequency to 1000Hz, and the spatial resolution to 2m. Eight types of vibration events were designed for the experiment: background noise, walking, knocking, digging, motorcycle, pickaxe, drill, and cart. The collected raw vibration signals were processed by sliding window segmentation and Z-score normalization, generating 24,589 samples. These samples were divided into a training set (17,213 samples), a validation set (3,683 samples), and a test set (3,699 samples) in a 7:1.5:1.5 ratio. One-hot encoding and random shuffling were used for training to ensure the model's generalization and stability.

[0106] 3. Model experimental environment and evaluation indicators;

[0107] All experiments were conducted on a device equipped with an NVIDIA GeForce RTX 2080 GPU (64GB VRAM) and CUDA version 11.3. All models used the same training parameters: batch size of 16, maximum training epochs of 50, learning rate of 0.000394, optimizer Adam, and loss function weighted cross-entropy.

[0108] To comprehensively evaluate model performance, four metrics—precision (P), recall (R), F1 score (F1-Score), and accuracy (A)—were selected for quantitative analysis. Where: P represents the proportion of actually positive samples among those predicted as positive; R represents the proportion of correctly identified samples among those actually positive; F1 is the harmonic mean of the two, used to comprehensively evaluate the model's classification ability; and A represents the overall proportion of correctly identified samples. The calculation formulas are as follows:

[0109]

[0110]

[0111]

[0112]

[0113] Where: N TP N represents the number of samples that are actually positive and whose predicted results are also positive. TN N represents the number of samples that were actually negative and whose predicted results were negative. FP N represents the number of samples that were actually negative but were predicted to be positive. FN This represents the number of samples that are actually positive but are predicted to be negative.

[0114] 4. Ablation experiment;

[0115] First, the model hyperparameter optimization experiment;

[0116] To obtain the optimal model architecture, this invention employs a Bayesian optimization algorithm to automatically search for key model parameters. The optimization space of the model includes: network depth (4-12 layers), basic number of channels (32-128), convolutional kernel size (3, 5, 7), and learning rate (1e). -5 up to 1e -2 ) and weight decay (1e -5 up to 1e -2 After 50 rounds of trials, the optimal configuration was obtained as shown in Table 2. Table 2 shows the search range and optimal values ​​of the model parameters.

[0117] Table 2

[0118]

[0119] The optimization process shows that a deeper network (8 layers) with a moderate number of channels (64) can achieve a good balance between model complexity and generalization ability. The use of a smaller kernel (kernel size=3) in the standard convolutional layer captures local features while maintaining a smaller number of parameters.

[0120] Then, the ablation experiment design was carried out;

[0121] To comprehensively verify the effectiveness of each key component in the model, this invention designed a series of ablation experiments. All ablation experiments used a pre-constructed dataset of vibration events from underground optical cables. This dataset contains 24,589 samples across 8 vibration event categories, strictly divided into training, validation, and test sets in a 7:1.5:1.5 ratio. The ablation experiments focused on four aspects: Multi-Scale Convolution (MSC), Network Depth, Squeeze and Excitation (SE) attention mechanism, and Channel Configuration. All experiments were trained for 50 epochs using AdamW (lr=3.94×10⁻⁶). -4 The loss function is a weighted cross-entropy function, and an early stopping mechanism (patience=15) is used to prevent overfitting.

[0122] Finally, the results of the ablation experiment;

[0123] Table 3 shows the complete results of 15 ablation variant experiments. Accuracy (A) and F1 score (F1-Score) were selected to evaluate the ablation variant experiments. The results show that the model (Baseline) configuration of this invention (MSC[3,5,7] +SE + Depth8 + Channels64) achieved the best performance, with a test accuracy of 0.9984 and an F1-Score of 0.9984.

[0124] Table 3

[0125]

[0126] To further understand the synergistic effects among the components, this method analyzes the performance of different ablation variants. Figure 4 The performance comparison of the model and different ablation variants on the training and validation sets is shown. Figure 4As shown, the accuracy of a single convolutional kernel (kernel size=3,5,7) decreased by 1.52%-13.60% compared to the baseline, indicating that single-scale features are difficult to simultaneously capture the local and global patterns of optical cable vibration signals. Analysis shows that a single convolutional kernel is insufficient to capture the multi-scale features of vibration signals. While smaller convolutional kernels (kernel size=3) can capture high-frequency details, their ability to perceive global patterns is limited. Larger convolutional kernels (kernel size=5, kernel size=7) can perceive longer time dependencies, but their resolution for local features is insufficient. Multi-scale convolution, by using convolutional kernels of different sizes in parallel, achieves the coordinated capture of local details and global patterns, significantly outperforming any single configuration. Experiments show that the three-kernel combination [3,5,7] achieves the best balance between performance and efficiency, significantly outperforming single-kernel, dual-kernel, and quad-kernel combinations. It captures local detail features (K3), mid-frequency transition features (K5), and low-frequency global trend features (K7), respectively, covering the main feature scales of optical cable vibration signals. In SE attention mechanism ablation, the recognition accuracy was 99.84% with the SE attention mechanism introduced, decreasing to 91.05% after removal, indicating that the SE attention mechanism effectively enhances key features and suppresses noise interference. In network depth ablation, the recognition accuracy was 80.37%, 98.08%, and 99.84% for depths of 4, 6, and 8 layers, respectively, showing that the 8-layer structure achieves the best balance between feature representation capability and training stability. Regarding channels, the accuracy was 81.51%, 88.05%, 99.84%, 83.54%, and 81.75% for 32, 48, 64, 96, and 128 channels, respectively, with 64 channels showing the best performance-to-computational cost balance.

[0127] In summary, the ablation experiments verified the rationality of the model architecture design. Each component improved the model performance in different ways, and there was a good synergistic effect between the components.

[0128] 5. Comparative experiment;

[0129] To comprehensively evaluate the performance and stability of the model, this implementation method conducted the following two sets of comparative experiments of different types:

[0130] First, model repeatability experiments;

[0131] To verify the stability of the model training, six independent experiments were conducted with the same dataset partitioning and hyperparameter configuration. The evaluation metrics included precision (P), recall (R), and F1 score. Table 4 shows the results of the repeatability comparison experiment of the model of this invention.

[0132] Table 4

[0133]

[0134] The model achieved a mean validation accuracy of 99.51% and a standard deviation of only 0.30%, indicating that the model can converge stably under different random initialization conditions and has good training consistency. The sixth experiment showed the best performance, with the final loss reduced to 0.56%, and validation and test accuracies of 99.81% and 99.84%, respectively. As shown in Table 1, the model performed excellently across all eight categories of vibration events, achieving 100% accuracy in the "background" and "dig" categories, while the F1 score for the other six categories was above 99.54%. Out of 3699 test samples, only 6 samples were misclassified, resulting in a misclassification rate as low as 0.16%.

[0135] Then, it is compared with existing models;

[0136] To verify the comprehensive performance advantages of this model, six representative deep learning models were selected for comparative experiments, including VanillaCNN, ResNet-1D, DenseNet-1D, MobileNet-1D, BiLSTM, and SignalTransformer. All models used the same dataset partitioning and hyperparameter settings, with a maximum training epoch of 50. An early stopping mechanism (patience=15) was employed to prevent overfitting. The evaluation metrics were F1 score and accuracy (A). Table 5 shows the experimental results of the comparative models. The model of this invention significantly outperforms other comparative models in both recognition accuracy and F1 score, achieving a test accuracy of 99.84%, which is significantly better than other comparative models. From the perspective of model efficiency, the inference time of 10.55ms meets the vibration signal processing efficiency requirements of the Φ-OTDR system at 1000Hz, providing a technical solution for real-time processing and rapid identification of vibration signals in underground optical cable monitoring scenarios.

[0137] Table 5

[0138]

[0139] Figures 5-11 The confusion matrices of each comparative model for eight types of vibration events in buried optical cables are presented. For example... Figure 11 As shown, VanillaCNN exhibits significant confusion in the Pickaxe and Cart classes of complex events; Figure 8 MobileNet-1D has high computational efficiency, but it still causes a lot of confusion and misjudgment in the Walk and Dig classes. Figure 6 The ResNet-1D model introduces residual and dense connection structures and Figure 7DenseNet-1D alleviates the gradient vanishing problem, but confusion still exists in semantically similar events, such as Cart class and Drill class, Walk class and Dig class; Figure 10 The BiLSTM model, as a representative of recurrent neural networks, performs well in processing periodic vibration events, achieving high recognition rates in the Walk and Motorcycle classes, but suffers from severe confusion in the Background and Walk classes, indicating that LSTM has difficulty distinguishing between low-frequency periodic signals and background noise. Figure 9 The SignalTransformer model captures global dependencies through a self-attention mechanism, outperforming BiLSTM overall, but it lacks the ability to extract local transient features. In contrast, the proposed model (a) achieves significant advantages across all categories through the collaborative design of multi-scale residual convolutional blocks, SE attention mechanism, and residual connections. The diagonal proportion of all eight event categories exceeds 0.9950. Compared to traditional CNNs, lightweight networks, residual networks, and temporal models, the proposed model improves the diagonal proportion of the confusion matrix by 2-25 percentage points, especially in high-confusion classes such as Cart and Pickaxe, while maintaining consistency in low-confusion vibration events such as Walk and Dig. This provides a solution for the identification and classification of vibration signals from buried optical fibers in practical engineering.

[0140] This implementation addresses the problems of insufficient multi-scale feature extraction, inadequate utilization of channel features, and gradient vanishing in deep networks in underground optical cable vibration event identification. It proposes a multi-event identification model for underground optical cable vibration based on SE-DMRCB multi-scale feature extraction. The model achieves complementary fusion of features at different time scales and adaptive enhancement of key channel features through the collaborative design of a data feature projection module and a deep residual convolutional feature learning module for multi-scale feature extraction. Simultaneously, residual connections effectively alleviate the gradient vanishing problem. Experimental results show that the model achieves an accuracy of 0.9984 for identifying eight types of underground vibration events in a real substation scenario. Compared with other comparative models, it exhibits superior performance in accuracy, real-time performance, and robustness. Furthermore, ablation experiments further verify the rationality of the model architecture design and the synergistic effect between the modules.

[0141] The current model utilizes vibration signals collected by the Φ-OTDR system. In actual engineering, a single optical cable is often laid for several kilometers or even tens of kilometers, and multiple vibration events may occur simultaneously at different locations. For example, while a vehicle is running over the cable at one end, construction excavation activities may be taking place at the other end. The current model is designed based on the single-event assumption. Future research needs to focus on spatiotemporal decoupling and localization techniques in multi-event concurrent scenarios: on the one hand, by separating features in the spatial dimension, independent identification of events at different locations can be achieved; on the other hand, a temporal attention mechanism can be introduced to capture complex patterns of alternating or overlapping events within a short time window at the same location, thereby improving the model's adaptability in long-distance monitoring scenarios.

[0142] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0143] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A method for multi-event recognition of underground cable vibration based on SE-DMRCB multi-scale feature extraction, characterized in that: The implementation process of the method is: Collecting the original vibration event signal of the buried optical cable, taking the original vibration event signal as the input of the buried optical cable vibration multi-event recognition model based on SE-DMRCB multi-scale feature extraction; Constructing a buried optical cable vibration multi-event recognition model based on SE-DMRCB multi-scale feature extraction; the multi-event recognition model includes a data preprocessing module, a data feature projection module, a multi-scale feature extraction deep residual convolution feature learning module, and a classification decision module; The original vibration event signal in step one is input into the data preprocessing module; the data preprocessing module cuts the original vibration event signal into a sliding window and normalizes the score to form an experimental data set; The data feature projection module extracts the multi-time scale initial features of the vibration signal through parallel multi-scale convolution; The multi-scale feature extraction deep residual convolution feature learning module uses 8 layers of stacked residual convolution blocks for deep feature extraction to obtain the deep feature representation of the vibration data; The classification decision module realizes the final 8-class vibration event recognition through global average pooling and a fully connected network.

2. The underground optical cable vibration multi-event recognition method based on SE-DMRCB multi-scale feature extraction according to claim 1, characterized in that: The data preprocessing module uses a sliding window strategy to segment the long time series signal, and sets the window overlap rate to 50%.

3. The underground optical cable vibration multi-event recognition method based on SE-DMRCB multi-scale feature extraction according to claim 1, characterized in that: The data feature projection module includes three one-dimensional convolution branches of different sizes, which use 3×1, 5×1 and 7×1 convolution kernels respectively, and each convolution branch outputs 21 to 22 feature channels; The multi-scale feature fusion process is represented as: ； ； In the formula, represents a channel dimension splicing operation, k e {3, 5, 7} is a convolution kernel size, L is a feature map length, C out is the output channel number of each convolution branch. The three different size convolution kernels capture high frequency local detail features, medium frequency transition features and low frequency global trend features respectively, and realize complementary fusion of multi-scale information of vibration events through feature splicing.

4. The underground optical cable vibration multi-event recognition method based on SE-DMRCB multi-scale feature extraction according to claim 1, characterized in that: In the multi-scale feature extraction deep residual convolution feature learning module, each residual convolution block includes a multi-scale convolution layer, a standard convolution layer, an SE attention mechanism layer and a residual connection layer; The multi-scale convolution layer receives the feature information output by the data feature projection module, and performs hierarchical scale refinement to obtain output features of the multi-scale convolution layer ; The standard convolution layer regularizes the output features of the multi-scale convolution layer, and introduces a standard 3x1 convolution layer to integrate and nonlinearly transform the features; and the features output by the standard convolution layer are obtained ; The output features of the standard convolutional layer are batch normalized and ReLU activated using batch normalization and an activation function to obtain batch-normalized output features ; The SE attention mechanism learns the interdependence between channels based on the normalized feature information, adaptively recalibrates the channel feature weight, and obtains the recalibrated feature.

5. The underground optical cable vibration multi-event recognition method based on SE-DMRCB multi-scale feature extraction according to claim 4, characterized in that: The process of the SE attention mechanism to obtain the recalibrated feature is: First, the global average pooling GAP operation is used to obtain the channel global descriptor of the vibration feature in the time dimension: z = GAP = (1 / L) ∈ R C ; where C is the number of channels, L is the characteristic length, and z ∈ R C is the vector of channel global descriptors; Then, learn the channel attention weight through two fully connected layers: the first layer performs channel compression, and the second layer performs channel excitation: the channel attention weight is as follows: s = σ(W2 ReLU(W1 z)) In the formula, W1 is a first layer full connection weight matrix, W2 is a second layer full connection weight matrix, σ( ) is a Sigmoid activation function, and attention weights are normalized to the interval [0, 1]. Finally, the learned channel attention weight s recalibrates the original feature through channel-by-channel multiplication to obtain the calibrated feature; As follows: ； In the formula, ⊙ represents a channel-by-channel multiplication, is the feature after recalibration by the SE attention mechanism.

6. The underground optical cable vibration multi-event recognition method based on SE-DMRCB multi-scale feature extraction according to claim 5, characterized in that: The feature processed by the SE attention mechanism layer is taken as input, and the next step is processed through downsampling and residual connection; residual connection is introduced in each residual convolution block to form a residual convolution block, and the multi-time scale initial features of the data feature projection module of the previous layer are transmitted to the output end through a shortcut path and added to the main path features; the specific implementation of the residual connection is divided into two cases according to whether the downsampling is performed: No downsampling layer, namely: Block 1-3 and Block 5-7, for the residual convolution block without downsampling, the spatial dimensions of the output features of the main path and the input features remain unchanged, and the residual connection adopts an identity shortcut path; ； wherein is the output feature of the i−1th layer, is the main path output feature, the shortcut path directly passes the input without any transformation; Containing a downsampling layer, namely: Block 4 or 8, for the residual convolution block performing downsampling, the main path is halved in spatial dimension through maximum pooling after the SE attention mechanism layer, and the shortcut path is correspondingly downsampled: ； ； ； In the formula, is the output feature of the i-1th layer, is the main path output feature, is the feature sampled by the max-pooling layer under the main path, is the feature sampled by the max-pooling layer under the shortcut path; kernel_size is a convolution kernel, and stride is a step length; that is, the same max-pooling operation as the main path is applied to the initial feature output by the data feature projection module of the previous layer, so as to ensure that the two paths are aligned in the time dimension and then added element by element.

7. The underground optical cable vibration multi-event recognition method based on SE-DMRCB multi-scale feature extraction according to claim 1, characterized in that: The classification decision module maps the signal depth features of the last residual convolution block in the deep residual convolution feature learning module of the multi-scale feature extraction to the probability distribution of 8 types of vibration events, and predicts the class with the maximum output probability.

Images