High-speed rail intrusion monitoring method based on distributed optical fiber sensing and fusion neural network

Abstract

The application discloses a high-speed rail intrusion monitoring method based on a distributed optical fiber sensing and a fusion neural network, and comprises the following steps: collecting disturbance signals of a high-speed rail perimeter by using a distributed optical fiber acoustic wave sensing system, performing segmentation processing on the collected disturbance signals, and constructing a sample data set containing multiple disturbance category labels; removing background noise contained in the sample data; constructing a high-speed rail intrusion detection model, wherein the high-speed rail intrusion detection model comprises a convolutional neural network, a bidirectional long short-term memory neural network, a splicing module, a RepMLP module and a softmax layer, and the occurrence probability of each disturbance category corresponding to the disturbance signal is predicted. The application realizes long-distance, high-precision, low-time-consumption and automatic abnormal disturbance monitoring of the high-speed rail perimeter.

Description

Technical Field

[0001] This invention belongs to the field of high-speed railway perimeter monitoring technology, specifically relating to a high-speed railway intrusion monitoring method based on distributed optical fiber sensing and fused neural networks. Background Technology

[0002] High-speed railways are generally characterized by long distances and complex geological conditions, which can lead to sudden disasters such as rockfalls and mudslides, posing a serious threat to the safety of high-speed train operations. Therefore, it is necessary to establish a real-time perimeter safety monitoring system along the railway line, with a focus on the safety of high-speed railway operations, to provide early warnings for emergencies.

[0003] As an important branch of Distributed Fiber Optic Sensing (DOFS) technology, DAS is highly suitable for perimeter security monitoring along high-speed railways due to its ability to achieve long-distance, distributed, real-time quantitative monitoring of dynamic vibrations along fiber optic lines. Fiber optic sensing monitoring data includes not only time-series data of disturbance occurrences but also spatial-series data of disturbance locations, providing rich information for subsequent analysis. To extract effective information from massive amounts of data, scholars both domestically and internationally have attempted to combine machine learning algorithms with DOFS technology. By manually defining certain features, appropriate machine learning algorithms, such as correlation vector machines, decision trees, and random forests, are selected to obtain the final classification results. Although these methods can achieve high recognition rates through manual feature extraction, they suffer from limitations in specificity and generalization ability.

[0004] With the rapid development of deep learning algorithms, their neural network structures have become more complex and diverse, enabling them to automatically extract deep features of signals, thereby reducing manual intervention and making them more adaptable to a wider range of changing environments. CNNs (Convolutional Neural Networks) are a type of feedforward neural network composed of input layers, convolutional layers, activation function layers, pooling layers, and fully connected layers, and are one of the representative algorithms of deep learning. CNNs do not require manually designed feature extractors; the shared parameters of the convolutional kernels within their hidden layers and the sparsity of inter-layer connections allow CNNs to perform gridded features with relatively low computational cost, and model performance can be improved by adjusting the number of layers and parameters. However, CNNs cannot extract time-series features, are prone to overfitting, and are relatively sensitive to signal noise, requiring additional techniques to improve the model's robustness. Compared to CNNs, LSTMs (Laser-Based Memory) are a special type of recurrent neural network designed to address the long-term dependency problem inherent in ordinary recurrent neural networks. Each recurrent unit has multiple gates, including input gates, output gates, forget gates, and hidden features. Due to their unique design structure, LSTMs are suitable for processing and analyzing correlations in important events with relatively long intervals and delays in time series. To fully consider the forward and backward dependencies at specific moments, we adopted a bidirectional LSTM (BiLSTM) model. However, BiLSTM ignores the spatial information in the data, and when processing long sequence data, it may encounter problems such as vanishing or exploding gradients, requiring the adoption of techniques to address these issues. Therefore, when using DAS to monitor the perimeter of high-speed railways, a single neural network model cannot effectively process the monitoring data, necessitating a new method to integrate the advantages of different neural networks.

[0005] Researchers in other fields have used methods such as concatenating and reclassifying the temporal and spatial features extracted by CNN and LSTM respectively. However, these established models cannot be directly transferred to the high-speed rail intrusion detection problem because high-speed rail intrusion data has characteristics such as few perturbation samples, short duration of perturbation events, small range of influence, and diverse types of perturbations. Without further optimization of the network structure, training high-speed rail intrusion data with only this established model will result in poor robustness, unsatisfactory accuracy, or even inability to distinguish between different types of perturbations due to the inability to address the local characteristics of different perturbation types.

[0006] Therefore, we are currently working on designing an end-to-end fusion network structure suitable for processing high-speed rail signal data, which can effectively improve the model's feature extraction capability and anti-interference capability, and enable the system to maintain high accuracy and second-level response speed even in relatively complex environments. Summary of the Invention

[0007] Technical problem solved: This invention discloses a high-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural networks. It utilizes existing railway optical cables and a distributed optical fiber acoustic sensing system (DAS) to achieve real-time monitoring of signals related to sudden accidents and illegal intrusions along the high-speed rail line. By employing advanced neural network classification and recognition technologies such as convolutional neural networks (CNN), bidirectional long short-term memory neural networks (BiLSTM), and reparameterized multilayer perceptrons (RepMLP), it identifies specific intrusion types for common high-speed rail intrusion events and sends information such as the accident location to relevant monitoring personnel, thereby achieving long-distance, high-precision, and low-time-consuming automated monitoring of the high-speed rail perimeter.

[0008] Technical solution:

[0009] A high-speed rail intrusion detection method based on distributed optical fiber sensing and fused neural networks, the high-speed rail intrusion detection method comprising the following steps:

[0010] S1. Disturbance signals at the perimeter of the high-speed railway are collected using a distributed fiber optic acoustic sensing system. The collected disturbance signals are segmented and processed, and the data of several sampling points within an adjacent spatial range within 1 second are divided into a sample data to construct a sample dataset containing multiple disturbance category labels.

[0011] S2, targeting background noise characteristics, uses heuristic threshold wavelet denoising to remove background noise contained in the sample data;

[0012] S3. Construct a high-speed rail intrusion detection model. The high-speed rail intrusion detection model includes a convolutional neural network, a bidirectional long short-term memory neural network, a splicing module, a RepMLP module, and a softmax layer. A sample dataset is used to train and validate the high-speed rail intrusion detection model. The convolutional neural network and the bidirectional long short-term memory neural network are set in parallel to extract features from the sample data after background noise removal, obtaining spatial and temporal features of the sample data. The spatial features of the sample data characterize the changes and relationships of the disturbance signal between adjacent sampling points within a window, while the temporal features characterize the behavior of the disturbance signal over time within the same window. The splicing module splices the spatial and temporal features of the sample data, obtaining the spatiotemporal feature vector of the disturbance signal, which is then input into the RepMLP module. The RepMLP module utilizes the local capture characteristics of the convolutional layer to characterize the geometric details of the spatiotemporal feature vector, extracting instantaneous disturbance features of different categories of disturbance signals. The softmax layer classifies the extracted instantaneous disturbance features of different categories of disturbance signals and predicts the occurrence probability of each disturbance category corresponding to the disturbance signal.

[0013] S4. Acoustic wave signals from the perimeter of the high-speed railway are collected using a distributed fiber optic acoustic wave sensing system. After segmentation and denoising of the collected acoustic wave signals, they are imported into the high-speed railway intrusion detection model. The instantaneous disturbance features of the acoustic wave signals at different sampling points are extracted, and the occurrence probability of each disturbance category corresponding to the acoustic wave signals is output. The type with the highest occurrence probability is output as the classification result of the sampling point at the current moment.

[0014] Furthermore, in step S1, the collected disturbance signal is segmented, and the data of 5 sampling points in adjacent spatial ranges within 1 second are divided into one sample data.

[0015] Furthermore, the disturbance category labels include seven types: climbing, knocking, digging, drilling, foreign object intrusion, illegal entry, and abnormal weather effects.

[0016] Furthermore, in step S2, the process of removing background noise from the sample data using a heuristic threshold wavelet denoising method based on background noise characteristics includes the following steps:

[0017] S21, the coif5 wavelet function is used to perform three-level wavelet decomposition calculation on the disturbance signal;

[0018] S22, set the fixed threshold and the unbiased threshold;

[0019] S23: Quantize the threshold of the high-frequency coefficients of the wavelet decomposition and adjust it according to the noise level estimate of the first-level wavelet decomposition to remove wavelet coefficients that belong to noise; when the signal-to-noise ratio is greater than the preset signal-to-noise ratio threshold, a fixed threshold is used for noise reduction; otherwise, an unbiased threshold is used for noise reduction.

[0020] S24 is reconstructed using inverse wavelet transform to obtain the denoised disturbance signal.

[0021] Further, in step S2, the convolutional neural network includes a first two-dimensional convolutional layer, a second two-dimensional convolutional layer, a third two-dimensional convolutional layer, a pooling layer, a flattening layer, a first fully connected layer, a second fully connected layer, and a third fully connected layer connected in sequence.

[0022] The kernel size of the first, second, and third 2D convolutional layers is 3×3, with a stride of 1 and padding of 1. For input data of size 5×100, the three 2D convolutional layers increase the number of channels from 1 to 64 while keeping the size of the last two dimensions of the data unchanged at 5×100. The pooling layer is used to perform a 2×2 2D Max Pooling operation on the output of the third 2D convolutional layer, reducing the data size of each channel to 2×50. The flattening layer is used to flatten the output of the pooling layer to obtain a one-dimensional vector of length 6400. The first, second, and third fully connected layers process the one-dimensional vector in sequence, mapping it to a 64-dimensional spatial feature vector.

[0023] Furthermore, a two-dimensional regularization layer and a ReLU activation function layer are connected after the two-dimensional convolutional layer; the two-dimensional regularization layer is used to perform batch normalization on the output to reduce the adverse effects of data distribution migration, and the ReLU activation function layer is used to introduce non-linear fitting ability.

[0024] A dropout layer is connected between the first fully connected layer and the second fully connected layer. The dropout layer is used to improve generalization ability and avoid overfitting to the input data.

[0025] Furthermore, the bidirectional long short-term memory neural network includes two stacked LSTM networks, each with an input size of 5 and a hidden layer size of 16; the second LSTM network takes the output of the first LSTM network as its input, and each LSTM network inputs the sequence data in both forward and reverse directions and obtains outputs in both directions.

[0026] Input data of size 5×100 is fed into the first layer of the LSTM network, and finally the outputs of two bidirectional layers, a total of four hidden layers of size 16, are obtained. These are then concatenated to obtain a time feature vector of length 64.

[0027] Furthermore, the RepMLP module includes a fourth fully connected layer, a convolutional layer, a regularization layer, and a fifth fully connected layer connected in sequence;

[0028] The RepMLP module processes the spatiotemporal feature vector of the perturbation signal and inputs the resulting logit vector of length 7 into the softmax layer.

[0029] Furthermore, in step S4, the occurrence probability of each disturbance category corresponding to the output acoustic signal is determined. If the occurrence probability of all types is lower than the preset probability threshold, it is determined to be a non-destructive disturbance. If the occurrence probability of at least one type is higher than the preset probability threshold, the type with the highest occurrence probability is output as the classification result of the sampling point at the current time, and the location of the sampling point is determined as the disturbance point location.

[0030] Beneficial effects:

[0031] First, the high-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network of the present invention can directly utilize the existing railway optical cable to meet the real-time monitoring needs along the high-speed rail line without the need for additional equipment, thereby reducing monitoring costs and subsequent maintenance expenses.

[0032] Secondly, the high-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network of the present invention selects the heuristic threshold wavelet method to denoise the data according to the characteristics of the high-speed rail environment. It can perform adaptive denoising processing according to the signal characteristics, eliminate the influence of background noise on the classification results in complex environments, and improve the overall signal-to-noise ratio of the data.

[0033] Third, the high-speed rail intrusion detection method based on distributed optical fiber sensing and fused neural network of the present invention automatically extracts and analyzes the segmented DAS signal using a fused neural network model, makes full use of the global features of the signal, and enhances the local capture characteristics of the model by adding a RepMLP module to replace the single fully connected layer, which helps to accurately distinguish different types of disturbances and improve classification accuracy. It is more reasonable than a single CNN or LSTM model and has the advantages of high efficiency, comprehensiveness and strong robustness, making it more suitable for actual high-speed rail perimeter security applications. Attached Figure Description

[0034] Figure 1 This is a flowchart of the high-speed rail intrusion detection method based on distributed optical fiber sensing and fused neural network of the present invention;

[0035] Figure 2 The images show a comparison of the noise reduction effects; (a) is a schematic diagram of the original signal, and (b) is a schematic diagram of the signal after noise reduction. Detailed Implementation

[0036] The following embodiments are provided to enable those skilled in the art to more fully understand the present invention, but do not limit the invention in any way.

[0037] This invention discloses a high-speed rail intrusion detection method based on distributed optical fiber sensing and fused neural networks, the high-speed rail intrusion detection method comprising the following steps:

[0038] S1. Disturbance signals at the perimeter of the high-speed railway are collected using a distributed fiber optic acoustic sensing system. The collected disturbance signals are segmented and processed, and the data of several sampling points within an adjacent spatial range within 1 second are divided into a sample data to construct a sample dataset containing multiple disturbance category labels.

[0039] S2, targeting background noise characteristics, uses heuristic threshold wavelet denoising to remove background noise contained in the sample data;

[0040] S3. Construct a high-speed rail intrusion detection model. The high-speed rail intrusion detection model includes a convolutional neural network, a bidirectional long short-term memory neural network, a splicing module, a RepMLP module, and a softmax layer. A sample dataset is used to train and validate the high-speed rail intrusion detection model. The convolutional neural network and the bidirectional long short-term memory neural network are set in parallel to extract features from the sample data after background noise removal, obtaining the spatial and temporal features of the sample data. The spatial features of the sample data are used to characterize the changes and relationships of the disturbance signal between adjacent sampling points within a window, while the temporal features are used to characterize the change behavior of the disturbance signal over time within the same window. The splicing module splices the spatial and temporal features of the sample data, obtaining the spatiotemporal feature vector of the disturbance signal, which is then input into the RepMLP module. The RepMLP module utilizes the local capture characteristics of the convolutional layer to characterize the geometric details of the spatiotemporal feature vector, extracting the instantaneous disturbance features of different categories of disturbance signals. The softmax layer classifies the extracted instantaneous disturbance features of different categories of disturbance signals and predicts the occurrence probability of each disturbance category corresponding to the disturbance signal.

[0041] S4. Acoustic wave signals from the perimeter of the high-speed railway are collected using a distributed fiber optic acoustic wave sensing system. After segmentation and denoising of the collected acoustic wave signals, they are imported into the high-speed railway intrusion detection model. The instantaneous disturbance features of the acoustic wave signals at different sampling points are extracted, and the occurrence probability of each disturbance category corresponding to the acoustic wave signals is output. The type with the highest occurrence probability is output as the classification result of the sampling point at the current moment.

[0042] This embodiment first uses a DAS system to collect disturbance signals around the high-speed rail perimeter, constructing a dataset containing seven classification labels, which is then divided into a test set and a validation set. Second, for background noise features, a heuristic threshold wavelet denoising method is used to denoise the data. Third, a deep neural network is built by fusing CNN and BiLSTM, and CNN and BiLSTM are used in parallel to extract data features and their correlation with preceding and following time periods. The results are concatenated to obtain a fused feature vector. Finally, RepMLP is used to output the occurrence probability of each type, and the location of the disturbance is determined based on the classification results of each sampling point. Based on this model, actual collected data is used for testing, and network parameters are adjusted using stochastic mini-batch gradient descent. A residual network model is added, and the RepMLP parameter settings are adjusted to optimize the model parameters until convergence, ultimately resulting in a high-accuracy high-speed rail intrusion detection model with a low false positive rate.

[0043] See Figure 1 The specific steps of this high-speed rail intrusion detection method are as follows:

[0044] I. Model Training Phase:

[0045] Step 1: Sample data can be collected from real-world scenarios or experimental scenarios. This invention utilizes existing railway optical cables as the sensing medium and employs DAS to achieve long-distance, distributed, real-time quantitative monitoring of perimeter disturbances without requiring additional equipment, thus reducing monitoring costs and subsequent maintenance expenses.

[0046] Considering the limited amount of disturbance data during actual high-speed rail operation, this embodiment incorporates disturbance data from experimental scenarios. In these scenarios, seven common disturbance types along the high-speed rail perimeter are fully simulated, and data is collected in real-time using a DAS system. In the simulated environment, optical fibers are laid in an S-shape along the perimeter fence and buried underground near the perimeter to achieve full perimeter coverage. The DAS system is connected to one end of the optical fiber to ensure that both the signal collection and transmission components of the DAS system are functioning correctly.

[0047] Around the deployed optical fiber, multiple disturbance experiments were conducted, including: (1) climbing, (2) knocking, (3) digging, (4) drilling, (5) foreign object intrusion, (6) illegal entry, and (7) abnormal weather noise. The signals collected from the optical fiber were then labeled with the above seven categories based on their specific types. Data from five adjacent sampling points within a 1-second spatial range were divided into a basic event window, i.e., one sample. A signal dataset containing the seven disturbance categories was constructed and divided into training and validation sets, including as many samples of each category as possible. The corresponding two-dimensional data size was 5×100. 100 samples were collected for each of the above types, and the training and validation sets were divided in a 7:3 ratio.

[0048] In actual high-speed rail monitoring scenarios, disturbance events are not instantaneous but have short durations, typically within seconds. Therefore, in addition to considering the location of the disturbance event, it is also necessary to consider the temporal dimension. Based on this principle, this embodiment segments the collected samples, dividing the data from five adjacent sampling points within 1 second into one sample data. Then, a fusion neural network is used to process the sample data to accurately represent the spatial and temporal characteristics of high-speed rail disturbances.

[0049] Step 2: Perform wavelet denoising on the constructed dataset. In the specific implementation of denoising, wavelet denoising uses a heuristic threshold for a three-level wavelet transform, and adjusts the threshold based on the noise level estimate from the first-level wavelet decomposition. In the high-speed rail intrusion detection scenario, due to coupling effects and many other factors, the data contains a large amount of background noise. This noise consists of small perturbations unrelated to the overall signal characteristics in the time dimension, with random and variable amplitudes. Its frequency domain characteristics are mainly 1-10Hz, which can cause these perturbations to be submerged in noise, affecting recognition accuracy. To address this noise problem, a wavelet transform denoising method based on heuristic thresholds is used to preprocess the dataset. The threshold is adjusted according to the signal characteristics to adaptively reduce the negative impact of background noise. Wavelet functions are used to decompose the data at multiple lower scales, and the threshold is adaptively adjusted based on the noise level estimate from the first-level wavelet decomposition. Then, the estimated wavelet coefficients at each scale are used to reconstruct the denoised signal, thereby improving the overall signal-to-noise ratio and eliminating the impact of background noise on the classification results in complex environments.

[0050] The specific steps of wavelet denoising are as follows: 1. Wavelet decomposition of the signal: the coif5 wavelet function is used to perform three-level wavelet decomposition calculation on the signal; 2. Select an appropriate threshold, the formula of which is shown in (2). When the signal-to-noise ratio is large, a fixed threshold is selected, and otherwise an unbiased threshold is selected to make it more suitable for denoising processing of actual data; 3. Perform threshold quantization of the high-frequency coefficients of wavelet decomposition and adjust them according to the noise level estimate of the first-level wavelet decomposition to remove wavelet coefficients that belong to noise; 4. Reconstruct the signal by inverse wavelet transform to obtain the denoised signal. Figure 2 The comparison between the effects before and after denoising is shown, where (a) is a schematic diagram of the original signal and (b) is a schematic diagram of the denoised signal. Let W be a vector, and let the set of squares of its wavelet decomposition coefficients be arranged in ascending order, as shown in formula (1):

[0051] W = [ω1, ω2, ..., ω] n (1)

[0052] In the formula, S represents the sum of squares of the wavelet coefficients. make The heuristic threshold T is then expressed as:

[0053]

[0054] Where σ is the mean square error of the noise signal and N is the length of the noise signal.

[0055] Step 3: Construct a CNN network containing three convolutional layers, one pooling layer, and three fully connected layers. For a 5×100 input sample, it outputs a feature vector of length 64. Each convolutional layer is followed by a Batch Normalization layer and a ReLU activation function layer to regularize the data and improve the system's generalization ability. Considering the complex variations and large amplitude changes in high-speed rail signals, we also added a Dropout module to the fully connected layers to prevent overfitting and overtraining.

[0056] Step 4: Construct a BiLSTM network and use two bidirectional LSTM layers to analyze the changes in the signal over time and extract temporal features. Specifically, a stacked two-layer LSTM network is used, with each layer having an input size of 5 and an output length of 16. In the model, the four data results (two layers and two layers) are concatenated to obtain a feature vector of length 64.

[0057] Step 5: Replace the fully connected layer with a RepMLP module. After concatenating the two feature vectors to obtain a mixed feature vector, input the result into the RepMLP module. By combining the advantages of fully connected layers and convolutional layers, it takes into account both global features and local information, thereby obtaining the occurrence probability of each of the seven categories. The type with the highest occurrence probability is output as the classification result.

[0058] Step 6: Test the above fused neural network using the denoised dataset. Adjust the network parameters and optimize the model parameters and network structure using the stochastic mini-batch gradient descent method until the model converges, resulting in a high-precision high-speed rail intrusion detection model with a low false positive rate.

[0059] In actual high-speed rail intrusion monitoring data collection, a large amount of data is undisturbed or contains only background noise, with relatively little actual disturbance data. This makes it impossible to distinguish multiple disturbance types when only a single feature is extracted from the dataset. To fully extract sample features and reduce reliance on manual intervention, a parallel CNN and BiLSTM network structure is used to obtain the spatiotemporal features of the samples, and RepMLP is used to replace the traditional fully connected layers. The specific approach is as follows:

[0060] Due to the diverse types of disturbances present in actual high-speed rail monitoring scenarios, manual feature extraction methods rely excessively on subjective experience, resulting in low efficiency and poor robustness. Therefore, this paper considers using deep learning network models to automatically extract data features. Because high-speed rail intrusion data is characterized by a small number of disturbance samples, short duration of disturbance events, small impact range, and diverse disturbance types, spatial structure information and temporal characteristics should be fully considered when extracting features. Furthermore, since high-speed rail disturbance signals generally exhibit short-duration strong vibration characteristics, manifested as an instantaneous increase in amplitude in the time domain, this feature alone is insufficient to distinguish between various disturbance types. Therefore, this embodiment employs CNN and BiLSTM networks for feature extraction, primarily aiming to extract temporal, frequency, and shallow morphological features from the acquired signals, and obtaining classification results through feature concatenation.

[0061] Since CNNs can automatically learn and extract the spatial features of input data, and considering the similarity of the frequency domain distribution of signals at adjacent sampling points within a window, this embodiment uses CNNs to process and characterize the changes and connections between adjacent sampling points of high-speed rail signals. Considering the real-time requirements of high-speed rail, the network structure needs to be optimized as much as possible while realizing the feature extraction function.

[0062] For input data of size 5×100, to fully extract shallow morphological features, three convolutional layers are used to convolve activation functions to enhance the nonlinear effect of the network. At the same time, three fully connected layers are set to link the outputs of the convolutional layers and integrate temporal, frequency, and morphological features to finally establish a feature vector of length 64. The specific parameters are shown in Table 1.

[0063] Table 1

[0064]

[0065] First, the data passes through three 2D convolutional layers, each with a kernel size of 3×3, a stride of 1, and padding of 1. This setup increases the number of channels from 1 to 64 while maintaining the size of the last two dimensions at 5×100. Then, a 2D Max Pooling operation of size 2×2 is performed on the output of the three convolutional layers, reducing the data size of each channel to 2×50. Finally, the data is flattened to obtain a 1D vector of length 64×2×50=6400, which then passes through three fully connected layers with output sizes of 1024, 256, and 64, respectively. This vector is then mapped to a 64-dimensional feature vector.

[0066] To improve model accuracy, each convolutional layer is followed by a two-dimensional Batch Normalization layer and a ReLU activation function layer. The former performs batch normalization on the output to reduce the adverse effects of data distribution shifts, while the latter introduces non-linear fitting capabilities. Additionally, a 0.5 Dropout operation is used in the fully connected layers to avoid overfitting the input data and improve the generalization ability of the CNN module.

[0067] In real-world high-speed rail scenarios, disturbances are not instantaneous but continuous. Therefore, when classifying signals, their temporal changes should be a crucial factor, and long-term dependencies at specific moments should also be fully considered. Since BiLSTM can consider the contextual information of the input sequence and analyze long-term dependencies to process and predict time-series data, the applicant proposes using a two-layer BiLSTM to analyze the signal's temporal changes within the input window, extracting its temporal features and establishing a feature vector of length 64. The BiLSTM is set to two layers, with each layer having an input size of 5 and a hidden layer size of 16. This involves stacking two LSTM layers, with the second layer using the output of the first layer as input. Each layer receives the sequence data in both forward and reverse directions, producing outputs in both directions. The input data passes through this network, ultimately resulting in four bidirectional hidden layer outputs of size 16, which, when concatenated, yield a feature vector of length 4 × 16 = 64.

[0068] Research shows that, considering both features extracted by CNN and BiLSTM, a parallel CNN and BiLSTM network structure yields more accurate and reasonable classification results compared to a single network model, while also improving computational efficiency. Therefore, the applicant proposes to concatenate the two feature vectors to obtain a fused feature vector of length 128.

[0069] Finally, the feature classification results are generated. Considering that various common disturbance types in high-speed rail exhibit similar behavior in both the time and frequency domains, and are characterized by short duration and small range of influence, the accuracy of identification cannot meet practical needs when only their spatiotemporal features are extracted. Furthermore, in the same sample, the actual disturbance data accounts for only 10% or even less of the total data, while the traditional single fully connected layer, which efficiently models long-distance dependencies and location patterns to achieve classification, performs poorly in small-area feature recognition. Therefore, the above network structure needs further optimization to consider both global feature extraction and local information perception. The RepMLP module is an effective combination of fully connected layers and convolutional layers, which can fully utilize the global representation and location perception capabilities of fully connected layers and the local structure extraction capabilities of convolutions, enhancing local capture characteristics and improving classification performance. Therefore, the applicant intends to use the RepMLP module to replace the fully connected layer used for classification, with specific parameters shown in Table 2.

[0070] Table 2

[0071]

[0072] Finally, the logit vector of length 7 is passed through a softmax layer to obtain the classification probability of the sample belonging to each category. The category with the highest probability is the classification result. This allows us to determine whether the input signal is a destructive disturbance and its disturbance category.

[0073] For each sampling point, the data from its five adjacent sampling points within 1 second are input into the aforementioned fusion network to obtain the classification result of that sampling point at that moment. The sampling point classified as a destructive disturbance is the potential location of the disturbance at the current moment, and an alarm is issued accordingly.

[0074] II. Practical Application Stage:

[0075] Step 1: Deploy the DAS system at the high-speed railway site that needs monitoring, connect one end of the existing optical cable to the DAS system, and conduct actual tests to ensure that the signal collection and transmission parts of the DAS system are operating normally, maintaining a low packet loss rate and a transmission speed at the second level.

[0076] Step 2: Input the real-time high-speed rail perimeter data collected by the DAS system into the pre-modulated wavelet noise reduction module for noise reduction processing.

[0077] Step 3: Input the real-time data within a spatiotemporal window after noise reduction into the trained neural network recognition module. CNN and BiLSTM are used in parallel to extract feature vectors, which are then concatenated to obtain a hybrid feature vector. This hybrid vector is then input into RepMLP to predict the probability of occurrence for each category. The category with the highest probability is output as the classification result for that sampling point at that moment. When an abnormal disturbance is detected, the midpoint of the event window is designated as the disturbance location, and a warning signal is automatically sent to alert relevant personnel to the location and type of intrusion. Based on the actual data collected on-site, the noise reduction threshold is further adjusted, and the model parameters are optimized to achieve high robustness, ultimately resulting in a high-accuracy high-speed rail intrusion detection model with a low false positive rate.

[0078] In the high-speed rail intrusion scenario, the duration of a single disturbance is short, and the proportion of disturbance data in each data sample is small. Therefore, data segmentation is necessary, and features are extracted from a subset of samples sequentially. Furthermore, since different network structures have different feature points for data features, a parallel neural network model structure is chosen to simultaneously acquire temporal and spatial features. For each data sample, it is divided into multiple 5×100 data points and input into a CNN containing three convolutional layers, one pooling layer, and three fully connected layers. Features of this window are extracted, and a feature vector of length 64 is constructed to represent the changes and relationships of the high-speed rail signal between adjacent sampling points within the window. Simultaneously, the 5×100 data points within this window are input in parallel into a two-layer BiLSTM to obtain a feature vector of length 64, representing the time-varying behavior of the high-speed rail signal within the same window.

[0079] Traditional network structures use fully connected layers as the output, relying on distance and location patterns and failing to capture local information. However, in high-speed rail intrusion detection, the range of disturbance types is short, and different disturbance types have similar frequency domain representations. Simply concatenating feature vectors cannot further improve recognition accuracy, requiring further structural optimization. Therefore, after concatenating feature vectors from CNN and BiLSTM to obtain a 128-bit spatiotemporal feature vector, the result is input into the RepMLP module, replacing the single fully connected layer. This fully utilizes the local capture characteristics of convolutional layers to characterize the geometric details of the data, focusing on extracting instantaneous disturbance features. Finally, the classification probability of the sample belonging to each category is obtained, and the category with the highest probability is the classification result. This allows determination of whether the input signal is a destructive disturbance and its category. When an abnormal disturbance is detected, an alarm function is activated and the disturbance location is located.

[0080] In practical implementation, the aforementioned fusion neural network model is first trained using the denoised training set data until convergence. The validation set data is used to determine hyperparameters and evaluate the training results to determine suitable model parameters. In real-world applications, the 5×100 real-time data of the high-speed rail perimeter collected at each sampling point is input into the denoising module, then into the trained network model to obtain the predicted classification results, i.e., whether it is a destructive disturbance and the disturbance category. Based on the classification results of each sampling point, the location of the disturbance is determined, i.e., the deployment location of sampling points whose classification results are destructive disturbances. For several consecutive anomalous sampling points, the system selects the center point of their geometric distance to predict the location of the intrusion.

[0081] After testing, the aforementioned high-speed rail perimeter intrusion detection system based on CNN and BiLSTM achieved a real-time data classification speed on the order of seconds. In practical applications, it can effectively respond to emergencies and issue alarms, preventing accidents. Due to the relatively shallow network layers and few parameters, and the distributed parallel computing capabilities of the CNN and BiLSTM modules, the system boasts a high processing speed. The addition of a noise reduction module further reduces the impact of background noise, minimizing the influence of environmental factors such as high pedestrian traffic or inclement weather, and maintaining ideal classification accuracy, reaching over 95%.

[0082] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.

Claims

1. A high-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural networks, characterized in that, The high-speed rail intrusion detection method includes the following steps: S1. Disturbance signals at the perimeter of the high-speed railway are collected using a distributed fiber optic acoustic sensing system. The collected disturbance signals are segmented and processed, and the data of several sampling points within an adjacent spatial range within 1 second are divided into a sample data to construct a sample dataset containing multiple disturbance category labels. S2, targeting background noise characteristics, uses heuristic threshold wavelet denoising to remove background noise contained in the sample data; S3. Construct a high-speed rail intrusion detection model. The high-speed rail intrusion detection model includes a convolutional neural network, a bidirectional long short-term memory neural network, a splicing module, a RepMLP module, and a softmax layer. A sample dataset is used to train and validate the high-speed rail intrusion detection model. The convolutional neural network and the bidirectional long short-term memory neural network are set in parallel to extract features from the sample data after background noise removal, obtaining the spatial and temporal features of the sample data. The spatial features of the sample data are used to characterize the changes and relationships of the disturbance signal between adjacent sampling points within a window, while the temporal features are used to characterize the change behavior of the disturbance signal over time within the same window. The splicing module splices the spatial and temporal features of the sample data, obtaining the spatiotemporal feature vector of the disturbance signal, which is then input into the RepMLP module. The RepMLP module utilizes the local capture characteristics of the convolutional layer to characterize the geometric details of the spatiotemporal feature vector, extracting the instantaneous disturbance features of different categories of disturbance signals. The softmax layer classifies the extracted instantaneous disturbance features of different categories of disturbance signals and predicts the occurrence probability of each disturbance category corresponding to the disturbance signal. S4. Acoustic wave signals from the perimeter of the high-speed railway are collected using a distributed fiber optic acoustic wave sensing system. After segmentation and denoising of the collected acoustic wave signals, they are imported into the high-speed railway intrusion detection model. The instantaneous disturbance features of the acoustic wave signals at different sampling points are extracted, and the occurrence probability of each disturbance category corresponding to the acoustic wave signals is output. The type with the highest occurrence probability is output as the classification result of the sampling point at the current moment.

2. The high-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network according to claim 1, characterized in that, In step S1, the collected disturbance signal is segmented, and the data of 5 sampling points in adjacent spatial ranges within 1 second are divided into one sample data.

3. The high-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network according to claim 1, characterized in that, The disturbance category labels include seven types: climbing, knocking, digging, drilling, foreign object intrusion, illegal entry, and abnormal weather effects.

4. The high-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network according to claim 1, characterized in that, In step S2, the process of removing background noise from the sample data using the heuristic threshold wavelet denoising method based on background noise characteristics includes the following steps: S21, the coif5 wavelet function is used to perform three-level wavelet decomposition calculation on the disturbance signal; S22, set the fixed threshold and the unbiased threshold; S23: Quantize the threshold of the high-frequency coefficients of the wavelet decomposition and adjust it according to the noise level estimate of the first-level wavelet decomposition to remove wavelet coefficients that belong to noise; when the signal-to-noise ratio is greater than the preset signal-to-noise ratio threshold, a fixed threshold is used for noise reduction; otherwise, an unbiased threshold is used for noise reduction. S24 is reconstructed using inverse wavelet transform to obtain the denoised disturbance signal.

5. The high-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network according to claim 1, characterized in that, In step S2, the convolutional neural network includes a first two-dimensional convolutional layer, a second two-dimensional convolutional layer, a third two-dimensional convolutional layer, a pooling layer, a flattening layer, a first fully connected layer, a second fully connected layer, and a third fully connected layer connected in sequence. The kernel size of the first, second, and third 2D convolutional layers is 3×3, with a stride of 1 and padding of 1. For input data of size 5×100, the three 2D convolutional layers increase the number of channels from 1 to 64 while keeping the size of the last two dimensions of the data unchanged at 5×100. The pooling layer is used to perform a 2×2 2D Max Pooling operation on the output of the third 2D convolutional layer, reducing the data size of each channel to 2×50. The flattening layer is used to flatten the output of the pooling layer to obtain a one-dimensional vector of length 6400. The first, second, and third fully connected layers process the one-dimensional vector in sequence, mapping it to a 64-dimensional spatial feature vector.

6. The high-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network according to claim 5, characterized in that, Each two-dimensional convolutional layer is followed by a two-dimensional regularization layer and a ReLU activation function layer. The two-dimensional regularization layer is used to perform batch normalization on the output to reduce the adverse effects of data distribution migration, and the ReLU activation function layer is used to introduce non-linear fitting ability. A dropout layer is connected between the first fully connected layer and the second fully connected layer. The dropout layer is used to improve generalization ability and avoid overfitting to the input data.

7. The high-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network according to claim 1, characterized in that, The bidirectional long short-term memory neural network includes two stacked LSTM networks, each with an input size of 5 and a hidden layer size of 16. The second LSTM network takes the output of the first LSTM network as its input, and each LSTM network inputs the sequence data in both forward and reverse directions and obtains outputs in both directions. Input data of size 5×100 is fed into the first layer of the LSTM network, and finally the outputs of two bidirectional layers, a total of four hidden layers of size 16, are obtained. These are then concatenated to obtain a time feature vector of length 64.

8. The high-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network according to claim 1, characterized in that, The RepMLP module includes a fourth fully connected layer, a convolutional layer, a regularization layer, and a fifth fully connected layer connected in sequence. The RepMLP module processes the spatiotemporal feature vector of the perturbation signal and inputs the resulting logit vector of length 7 into the softmax layer.

9. The high-speed rail intrusion monitoring method based on distributed optical fiber sensing and fused neural network according to claim 1, characterized in that, In step S4, the occurrence probability of each disturbance category corresponding to the output acoustic signal is judged. If the occurrence probability of all types is lower than the preset probability threshold, it is judged as a non-destructive disturbance. If the occurrence probability of at least one type is higher than the preset probability threshold, the type with the highest occurrence probability is output as the classification result of the sampling point at the current time, and the location of the sampling point is determined as the disturbance point location.

Images