Optical cable intrusion detection method and device based on distributed optical fiber vibration sensing

Abstract

This invention provides a method and device for detecting optical cable intrusion based on distributed optical fiber vibration sensing. This approach addresses the technical challenges of optical fiber sensing and security monitoring, such as high false alarm rates, poor real-time performance, and the inability to meet the security requirements of long-distance optical cable monitoring systems. The method comprises: obtaining a first vibration signal within a preset amplitude range; performing frequency filtering on the first vibration signal to obtain a second vibration signal; performing feature extraction on the second vibration signal to obtain a high-dimensional feature vector; performing multi-scale time-series classification on the high-dimensional feature vector to obtain classified data; and determining the vibration state of the optical cable based on the classified data. This solution enables remote security monitoring of optical cables and improves the accuracy of detection results.

Description

Technical Field

[0001] The present invention relates to the technical field of optical fiber sensing and security monitoring, and in particular to an optical cable intrusion detection method and device based on distributed optical fiber vibration sensing. Background Art

[0002] With the accelerating pace of urbanization and the rapid development of the information society, the security of long-distance optical cables, as critical infrastructure for information transmission, has become increasingly important. Traditional security monitoring methods rely primarily on manual inspections and simple sensor alarm systems. Manual inspections are inefficient and subjective, making it difficult to provide real-time, comprehensive monitoring of long-distance optical cables. While simple sensor alarm systems achieve a degree of automated monitoring, they suffer from poor accuracy and reliability in complex environmental noise environments. They are unable to effectively distinguish between real vibration events and environmental disturbances such as noise from wind, rain, vehicles, and animal activity, and can easily be misidentified as actual intrusion or sabotage. False alarms and missed alarms are frequent, leading to unnecessary on-site inspections and significant waste of manpower and resources. Furthermore, the systems suffer from poor real-time performance. Existing systems often require a long time for data processing and analysis, making it difficult to accurately classify and respond to vibration events immediately, thus failing to meet the timeliness requirements of long-distance optical cable security.

[0003] Therefore, in order to ensure the safe operation of long-distance optical cables, it is urgent to develop an efficient and stable optical fiber vibration event classification system and device to improve the accuracy and real-time performance of monitoring, reduce the false alarm rate, and provide reliable technical support for the safety protection of long-distance optical cables. Summary of the Invention

[0004] The present invention provides an optical cable intrusion detection method and device based on distributed optical fiber vibration sensing, which solves the problems of high false alarm rate, poor real-time performance and difficulty in meeting long-distance optical cable security monitoring requirements in existing optical cable security monitoring.

[0005] In order to solve the above technical problems, the technical solutions of the present invention are as follows:

[0006] An embodiment of the present invention provides an optical cable intrusion detection method based on distributed optical fiber vibration sensing, comprising:

[0007] Acquire a first vibration signal within a preset amplitude range;

[0008] performing frequency filtering on the first vibration signal to obtain a second vibration signal;

[0009] performing feature extraction processing on the second vibration signal to obtain a high-dimensional feature vector;

[0010] Performing multi-scale time series classification processing on the high-dimensional feature vector to obtain classification data;

[0011] The vibration state of the optical cable is determined based on the classification data.

[0012] Optionally, obtaining a first vibration signal within a preset amplitude range includes:

[0013] The original vibration signal is obtained by installing distributed optical fiber vibration sensors at preset distances along the subway track;

[0014] According to a preset dynamic amplitude threshold, the original vibration signal is subjected to amplitude screening processing to obtain a first vibration signal.

[0015] Optionally, performing frequency filtering on the first vibration signal to obtain a second vibration signal includes:

[0016] Get the noise figure;

[0017] performing frequency band decomposition on the first vibration signal according to the noise coefficient to obtain a vibration sequence signal including vibration signals of multiple frequency bands;

[0018] Reconstructing the noise coefficient to obtain a target noise coefficient;

[0019] According to the target noise coefficient, frequency filtering is performed on the vibration sequence signal to obtain a second vibration signal.

[0020] Optionally, performing feature extraction processing on the second vibration signal to obtain a high-dimensional feature vector includes:

[0021] performing frame division and windowing processing on the second vibration signal to obtain a windowed short time frame;

[0022] Performing time-frequency domain feature extraction processing on the windowed short time frame to obtain a time-frequency feature matrix;

[0023] Embedding coding is performed on the time-frequency feature matrix to obtain a high-dimensional feature vector.

[0024] Optionally, performing multi-scale time series classification processing on the high-dimensional feature vector to obtain classification data includes:

[0025] The high-dimensional feature vector is input into a trained multi-scale time series classification model for processing to obtain classification data.

[0026] Optionally, the multi-scale time series classification model is trained through the following process:

[0027] Obtaining original vibration sample signals through distributed optical fiber vibration sensors installed along the subway track at preset distances;

[0028] Performing amplitude screening processing on the original vibration sample signal to obtain a first vibration sample signal;

[0029] performing frequency filtering on the first vibration sample signal to obtain a second vibration sample signal;

[0030] Classify the second vibration signal according to the subway vibration sample type label, the mining vibration sample type label, and the climbing vibration type label to obtain a vibration event sample library;

[0031] Performing feature extraction on sample data in the vibration event sample library to obtain a time-frequency domain feature matrix;

[0032] A multi-scale time series classification model is obtained according to the time-frequency domain feature matrix.

[0033] Optionally, the time-frequency domain feature matrix is processed to obtain fused discriminant features, including:

[0034] Performing multi-scale time series feature extraction on the time-frequency domain feature matrix to obtain multi-scale eigenvalues;

[0035] Performing spatiotemporal attention weighting on the multi-scale feature values to obtain attention-weighted features;

[0036] Performing bidirectional temporal modeling on the attention weighted features to obtain a prediction probability;

[0037] According to the predicted probability, focal loss optimization and parameter update are performed on the classification model with preset parameters to obtain a multi-scale time series classification model.

[0038] An embodiment of the present invention further provides an optical cable intrusion detection device based on distributed optical fiber vibration sensing, comprising:

[0039] An acquisition module, configured to acquire a first vibration signal within a preset amplitude range;

[0040] a processing module configured to perform frequency filtering on the first vibration signal to obtain a second vibration signal; perform feature extraction on the second vibration signal to obtain a high-dimensional feature vector; and perform multi-scale time series classification on the high-dimensional feature vector to obtain classified data;

[0041] A determination module is used to determine the vibration state of the optical cable according to the classification data.

[0042] An embodiment of the present invention further provides a computing device, comprising: a processor and a memory storing a computer program, wherein the computer program executes the above method when executed by the processor.

[0043] An embodiment of the present invention further provides a computer-readable storage medium storing instructions, which, when executed on a computer, enable the computer to execute the above method.

[0044] The technical solution of the present invention includes at least the following effects:

[0045] The above-mentioned scheme of the present invention obtains a first vibration signal within a preset amplitude range; performs frequency filtering processing on the first vibration signal to obtain a second vibration signal; performs feature extraction processing on the second vibration signal to obtain a high-dimensional feature vector; performs multi-scale time series classification processing on the high-dimensional feature vector to obtain classification data; determines the vibration state of the optical cable based on the classification data; and adopts a joint time-space modeling method to achieve stable monitoring performance under complex environmental noise, thereby improving the intelligence level of long-distance optical cable security monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 is a flow chart of an optical cable intrusion detection method based on distributed optical fiber vibration sensing provided by an embodiment of the present invention;

[0047] Figure 2 This is a flow chart of the intrusion monitoring process of the optical cable intrusion detection method based on distributed optical fiber vibration sensing provided by an embodiment of the present invention;

[0048] Figure 3 is a structural diagram of an optical cable intrusion detection device based on distributed optical fiber vibration sensing provided by an embodiment of the present invention;

[0049] Figure 4 It is a structural diagram of a computing device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0050] Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0051] like Figure 1 and Figure 2 As shown, an embodiment of the present invention provides an optical cable intrusion detection method based on distributed optical fiber vibration sensing, comprising:

[0052] Step 11, obtaining a first vibration signal within a preset amplitude range;

[0053] Step 12: performing frequency filtering on the first vibration signal to obtain a second vibration signal;

[0054] Step 13: performing feature extraction processing on the second vibration signal to obtain a high-dimensional feature vector;

[0055] Step 14, performing multi-scale time series classification processing on the high-dimensional feature vector to obtain classification data;

[0056] Step 15: Determine the vibration state of the optical cable according to the classification data.

[0057] In this embodiment, the original vibration signal is first acquired through optical fiber vibration sensors distributed along the optical cable, and the amplitude of the original vibration signal is processed to obtain a first vibration signal. These sensors can be fiber Bragg grating sensors, accelerometers, etc. When a vibration event occurs around the optical cable (such as digging, climbing, knocking, etc.), the sensor can sense and convert these physical vibrations into electrical signals or optical signals. To ensure that various possible vibration conditions can be accurately captured, the signal acquisition frequency needs to be set according to actual needs. Generally speaking, for high-frequency vibration events, such as rapid digging movements, a higher acquisition frequency is required, which can reach thousands of times per second or even higher; for low-frequency vibration events, such as slow human movement, the acquisition frequency can be appropriately reduced, but it is also necessary to ensure that the vibration characteristics can be accurately recorded.

[0058] Since the collected raw vibration signals often contain a large amount of noise, such as environmental noise and sensor noise, filtering techniques can be used for denoising. For example, low-pass filtering can remove high-frequency noise while retaining the useful low-frequency components of the vibration signal. Band-pass filtering can filter vibration signals within a specific frequency range and remove noise from other irrelevant frequencies. To improve the quality of the vibration signal and facilitate subsequent feature extraction, signal enhancement processing is also required. For example, signal amplification technology can be used to amplify weak vibration signals to an appropriate amplitude, or signal smoothing technology can be used to smooth the signal waveform, reduce glitches and mutations, and improve signal stability. Through the above processing, a second vibration signal is obtained.

[0059] Analyze the characteristics of the second vibration signal from a time domain perspective. Common time domain characteristics include mean, variance, peak value, peak-to-peak value, rise time, fall time, etc.

[0060] The second vibration signal is converted from the time domain to the frequency domain by Fourier transform and other methods, and the frequency domain features are extracted.

[0061] Frequency domain features primarily include frequency components and power spectral density. For example, different vibration events exhibit distinct frequency distributions in the frequency domain. By analyzing these frequency components, different vibration types can be distinguished. Combining time and frequency domain information, methods such as wavelet transforms are used to extract time-frequency domain features. Wavelet transforms provide local characteristics of the signal at different times and frequencies, more comprehensively reflecting the characteristics of the vibration signal. The extracted time, frequency, and time-frequency domain features are combined to form a high-dimensional feature vector. This feature vector provides a comprehensive and detailed description of the vibration signal's characteristics.

[0062] After obtaining the high-dimensional feature vector, you need to select an appropriate classification model to classify it. Different classification models have different characteristics and applicable scenarios, so you need to choose one based on your specific situation. You can train the classification model using sample data with known categories. The training data should include various types of vibration events, their corresponding high-dimensional feature vectors, and class labels.

[0063] Through training, the classification model learns the characteristic patterns of different vibration event categories, enabling it to accurately classify new high-dimensional feature vectors. The high-dimensional feature vector to be classified is input into the trained classification model, which then makes a judgment based on the learned characteristic patterns and outputs a classification result. The classification result can be a different vibration event type, such as digging, climbing, or knocking, or an assessment of the vibration event's severity, such as normal, suspicious, or dangerous.

[0064] Based on the classification results, the optical cable vibration status is determined and corresponding alarm rules are set. For example, if the classification result is "dangerous," a high-level alarm is triggered; if the classification result is "suspicious," a low-level alarm is triggered; if the classification result is "normal," no alarm is triggered. The alarm rules control the operating status of the security monitoring device. The security monitoring device can be an audible and visual alarm, a text message alarm module, an email alarm system, etc. When an alarm is triggered, the alarm device will issue a corresponding alarm signal, such as a loud siren, flashing warning lights, or sending an alarm text message or email, to promptly notify relevant personnel for action.

[0065] This technical solution accurately detects vibration conditions around optical cables by acquiring vibration signals and performing a series of processing. From initial signal acquisition to preprocessing to remove noise and other interference, and then to feature extraction to generate high-dimensional feature vectors, this approach enables more detailed and accurate analysis of vibration signals. This allows precise identification of different types and intensities of vibration, helping to accurately determine whether the optical cable is in a safe state. Classification data is then determined based on the high-dimensional feature vectors, used to determine the cable's vibration status and control the operation of the alarm device, enabling intelligent judgment and decision-making. The system automatically determines whether to issue an alarm based on the analysis results, eliminating the need for human intervention. This improves the timeliness and accuracy of monitoring and reduces the impact of human factors. The entire process rapidly processes vibration signals. Once an anomaly is detected, the alarm device is activated, allowing personnel to take swift action to prevent damage to the optical cable and ensure the safe and stable operation of the communication network. The vibration signal preprocessing and feature extraction process enhances the system's adaptability to diverse environments and vibration types. Regardless of environmental fluctuations, the system accurately analyzes vibration signals, ensuring reliable monitoring and broad application in complex environments and across various optical cable lines.

[0066] In an optional embodiment of the present invention, in step 11, obtaining a first vibration signal within a preset amplitude range may include:

[0067] Step 111, obtaining original vibration signals by using distributed optical fiber vibration sensors installed at preset distances along the subway track;

[0068] Step 112: Perform amplitude screening processing on the original vibration signal according to a preset dynamic amplitude threshold to obtain a first vibration signal.

[0069] In this example, hardware deployment is first performed. Distributed fiber optic sensors are deployed every 200 meters along the subway track, covering a 10-kilometer cable segment. The sampling rate is set to 2 kHz. The sensors collect vibration signals in real time. The raw signals contain the following components:

[0070] (1) Subway train vibration: The frequency is concentrated in 20~200Hz, showing a periodic continuous waveform;

[0071] (2) Environmental noise: such as wind noise (low frequency), vehicle traffic (medium frequency), etc.

[0072] (3) Potential intrusion signals: such as impact vibration (digging behavior, frequency band 50~300Hz) or intermittent vibration (climbing behavior, frequency band 10~100Hz);

[0073] The original vibration signal sequence collected can be expressed as: raw (t), where t is the time index.

[0074] The original vibration signal x is collected raw (t), the dynamic amplitude threshold method is used to filter and process the original vibration signal to remove transient impulse noise, such as lightning strikes, equipment interference, etc., to obtain the target vibration signal; specifically,

[0075] Based on the amplitude of the signal over the past N frames (N is 1000 in this example), calculate the historical signal amplitude mean μ and the historical signal amplitude standard deviation σ:

[0076] ;

[0077] ;

[0078] Among them, x raw (t i ) is the i-th signal in the original vibration signal sequence.

[0079] The threshold U is set according to the historical signal amplitude mean μ and the historical signal amplitude standard deviation σ:

[0080] U = μ ± 3σ;

[0081] By setting the threshold U, only signal segments with amplitudes within this range are retained. The screening process can be expressed as:

[0082] ;

[0083] Among them, x filtered (t) is the first vibration signal.

[0084] This embodiment deploys distributed fiber optic sensors along the subway tracks and combines them with dynamic amplitude threshold screening technology to accurately extract effective vibration signals and suppress instantaneous interference, ensuring full coverage of the monitoring area and high-precision data collection. It also dynamically sets thresholds based on the historical signal mean and standard deviation to filter out abnormal amplitude noise such as lightning strikes and equipment interference, while retaining subway vibration and potential intrusion signals, providing low-noise, high signal-to-noise ratio input data for subsequent processing, significantly improving the anti-interference capability and reliability of the detection system.

[0085] In an optional embodiment of the present invention, in step 12, performing frequency filtering on the first vibration signal to obtain the second vibration signal may include:

[0086] Step 121, obtaining the noise coefficient;

[0087] Step 122: performing frequency band decomposition on the first vibration signal according to the noise coefficient to obtain a vibration sequence signal including vibration signals of multiple frequency bands;

[0088] Step 123, reconstructing the noise coefficient to obtain a target noise coefficient;

[0089] Step 124: Perform frequency filtering on the vibration sequence signal according to the target noise coefficient to obtain a second vibration signal.

[0090] In this embodiment, the first vibration signal is first marked, and segments with a duration of not less than 0.5s are retained to avoid transient noise residue, and a valid vibration signal segment set is obtained. , where each segment corresponds to a time window;

[0091] A five-layer decomposition is performed using a wavelet basis to remove high-frequency detail coefficients (>500 Hz) and reconstruct low-frequency signals to retain the effective components of subway vibration and intrusion events. The decomposition process can be expressed as:

[0092] ;

[0093] Among them, c A5 (t) is the fifth-layer approximation coefficient, representing the low-frequency effective signal (such as subway vibration); c Dj (t) is the detail coefficient of the jth layer (j=1,2,...,5), which represents high-frequency noise (such as wind noise and equipment interference).

[0094] For detail coefficient c Dj (t) (j=1,2,...,5) performs hard threshold processing to remove components with frequencies > 500Hz; only the low-frequency approximate coefficient c is retained A5 (t), reconstructed denoised signal;

[0095] Signal frequency filtering can be expressed as:

[0096] ;

[0097] Among them, ϕ 5,m (t) is the translation version of the 5th layer wavelet basis function; c A5 (t m ) is the translation version of the 5th layer approximation coefficient; m is the translation index; is the reconstructed noise-reduced signal, that is, the second vibration signal after preprocessing.

[0098] In this embodiment, the wavelet multi-scale decomposition and hard threshold filtering technology are used to achieve the technical effect of effectively separating low-frequency effective signals and suppressing high-frequency noise. The first vibration signal is divided into multiple frequency bands by 5-layer wavelet decomposition, and the high-frequency detail coefficients with a frequency greater than 500Hz are eliminated, while the low-frequency approximate coefficients are retained. At the same time, hard threshold processing and signal reconstruction are used to eliminate transient noise residues, ensuring that the signal-to-noise ratio of the signal after noise reduction is significantly improved, providing a pure time-frequency data basis for subsequent feature extraction.

[0099] In an optional embodiment of the present invention, in step 13, performing feature extraction processing on the second vibration signal to obtain a high-dimensional feature vector may include:

[0100] Step 131, performing frame division and windowing processing on the second vibration signal to obtain a windowed short time frame;

[0101] Step 132, performing time-frequency domain feature extraction processing on the windowed short time frame to obtain a time-frequency feature matrix;

[0102] Step 133: Perform embedded coding processing on the time-frequency feature matrix to obtain a high-dimensional feature vector.

[0103] In this embodiment, the second vibration signal (long time series signal) is first divided into short time segments to reduce spectrum leakage and facilitate time-frequency analysis. The frame parameters are: frame length 100ms, frame shift 50ms, and Hamming window smoothing;

[0104] In step 131, the frame division and windowing process can be expressed as follows:

[0105] ;

[0106] in, is the short time frame after windowing; t m is the starting time of the mth frame; N is the number of sampling points corresponding to the frame length; w(n) is the Hamming window function; where,

[0107] ;

[0108] After obtaining the windowed short-time frame, the Mel spectrum (64 dimensions) and Mel frequency cepstral coefficient (12 dimensions) are calculated for each frame signal and fused into a 76-dimensional feature vector.

[0109] The specific implementation process of step 132 includes:

[0110] (1) Performing short-time Fourier transform on the windowed short-time frame:

[0111] ;

[0112] Among them, f s is the sampling frequency; f is the frequency component; X (m) (f) is the signal after short-time Fourier transform.

[0113] (2) Perform Mel filter bank mapping:

[0114] ;

[0115] Among them, H b(f) is the frequency response of the b-th Mel filter; is the Mel spectrum.

[0116] (3) To calculate the Mel-frequency cepstral coefficients, first take the logarithm of the Mel-frequency spectrum: ; Then perform discrete cosine transform to extract Mel frequency cepstral coefficients:

[0117] ;

[0118] Among them, MFCC (m) (c) is the Mel frequency cepstral coefficient;

[0119] (4) 64-dimensional Mel spectrum with 12 Mel-frequency cepstral coefficients MFCC (m) (c) Horizontally splicing on the feature dimension to obtain the time-frequency feature matrix F (m) ;

[0120] In step 133, the time-frequency feature matrix F is obtained (m) After that, it is embedded and encoded, that is, through the cascade of convolution and recurrent neural networks, the local frequency domain pattern and the global temporal dependency are integrated to generate a compact high-dimensional representation. Specifically, the first layer is a 1D convolution layer, and the 1D convolution kernel (width 5, step size 2) extracts the local frequency domain pattern; the second layer is a bidirectional LSTM layer, and the bidirectional LSTM (hidden unit 64) captures the temporal dependency; the output is a 256-dimensional high-dimensional vector to represent the spatiotemporal characteristics of the vibration event.

[0121] This embodiment achieves effective extraction and optimized representation of high-dimensional spatiotemporal features through frame windowing, time-frequency feature fusion, and deep learning embedded coding. Specifically, this includes: 1) frame windowing to reduce spectral leakage and segment short-term signals; 2) extracting a 76-dimensional Mel-spectrogram (64-dimensional) and MFCC (12-dimensional) to fuse time-frequency features; and 3) cascading 1D convolution (local frequency domain patterns) and bidirectional LSTM (global temporal dependency) encoding to generate a 256-dimensional high-dimensional vector that comprehensively characterizes the spatiotemporal characteristics of vibration events, providing robust and discriminative input features for multi-scale classification.

[0122] In an optional embodiment of the present invention, in step 14, performing multi-scale time series classification processing on the high-dimensional feature vector to obtain classification data may include:

[0123] Step 141: Input the high-dimensional feature vector into a trained multi-scale time series classification model for processing to obtain classification data.

[0124] In this embodiment, the high-dimensional feature vector is input into the multi-scale time series classification model for processing, and finally the classification data is obtained. The process includes two stages: feature extraction and weighting processing, and adaptive fusion processing;

[0125] The multi-scale time series classification model is trained through the following process:

[0126] Step 1: obtaining original vibration sample signals by installing distributed optical fiber vibration sensors at preset distances along the subway track;

[0127] Step 2: performing amplitude screening processing on the original vibration sample signal to obtain a first vibration sample signal;

[0128] Step 3: performing frequency filtering on the first vibration sample signal to obtain a second vibration sample signal;

[0129] Step 4: Classify the second vibration signal according to the subway vibration sample type label, the mining vibration sample type label, and the climbing vibration type label to obtain a vibration event sample library;

[0130] Step 5: extracting features from the sample data in the vibration event sample library to obtain a time-frequency domain feature matrix;

[0131] Step 6: Obtain a multi-scale time series classification model based on the time-frequency domain feature matrix.

[0132] Specifically, the amplitude screening process of the original vibration sample signal in step 2 includes the following process:

[0133] Dynamic threshold alarm: Adaptively adjusts the classification confidence threshold (default 0.9, reduced to 0.7 when noise is high) based on the ambient noise level (e.g., increased noise in rainy or snowy weather).

[0134] Alarm rules: A real-time alarm is triggered when three consecutive frames are detected as "digging" or a single frame is detected as "climbing" with a confidence level greater than 0.95.

[0135] The method of performing frequency filtering on the first vibration sample signal in step 3 is consistent with the filtering method in step 12;

[0136] In step 4, the pre-processed vibration signals are classified according to the subway vibration sample type label, mining vibration sample type label, and climbing vibration type label to obtain a vibration event sample library, including: classification and labeling by vibration type to construct a standardized sample library. The classification rules are as follows:

[0137] (1) Subway vibration. It is characterized by a continuous vibration waveform with a duration of 30-60 seconds and a main frequency of 80 Hz. The signal segment when the train passes is extracted and marked as "subway".

[0138] (2) Excavation vibration. It is characterized by impact vibration, lasting 0.5-2 seconds, and a frequency range of 50-300 Hz. The simulated excavation experimental data is marked as "excavation".

[0139] (3) Climbing vibration. It is characterized by intermittent vibration lasting 5-10 seconds with a frequency range of 10-100 Hz; it simulates climbing behavior data and is labeled “climbing”.

[0140] The labeled vibration events are output as a dataset, represented as:

[0141] y k ∈{subway vibration, excavation vibration, climbing vibration};

[0142] A standardized sample library was constructed, containing three types of labels (subway vibration, excavation vibration, and climbing vibration), 1,000 groups of samples in each type, and divided into training set, validation set, and test set in the ratio of 8:1:1.

[0143] In step 5, feature extraction is performed on the sample data in the vibration event sample library to obtain a time-frequency domain feature matrix; specifically, the vibration event sample library is processed according to the process of step 13 to obtain a time-frequency domain feature matrix;

[0144] In step 6, a multi-scale time series classification model is obtained based on the time-frequency domain feature matrix. Specifically, a classification model with preset parameters is first established. The model structure is a stack of 4 TimeMixer blocks, each of which contains dilated convolution (expansion factor 2) and a local attention mechanism to extract multi-scale time series features. An 8-head self-attention mechanism models long-range dependencies and weightedly fuses features from different time steps. The model is pre-trained on the UCI-HAR public vibration dataset and then fine-tuned in the subway scenario. The training strategy uses focal loss (α=0.25, γ=2) to alleviate category imbalance. AdamW (initial learning rate 3e-4, weight decay 1e-5) is used. During the fine-tuning phase, the first 3 layers of TimeMixer blocks are frozen, and only the top layer and the classifier are trained.

[0145] In this stage, a multi-scale time series analysis model is constructed to extract and classify the vibration signal in a refined manner, thereby obtaining fused discriminative features. The specific implementation process includes:

[0146] (1) Standardize the input data:

[0147] ;

[0148] Where μ is the mean, σ is the standard deviation, is the standard value, x i is the element in the time-frequency domain feature matrix.

[0149] (2) Multi-scale feature extraction, dilated convolution operation, to obtain multi-scale feature values:

[0150] ;

[0151] Where d is the dilation factor, J is the size of the convolution kernel, y[t] is the output feature at time point t after the convolution operation, w[k] is the weight parameter of the kth position of the convolution kernel, and x[t] is the eigenvalue of the element in the time-frequency domain feature matrix at time point t.

[0152] After the convolution operation on the vibration signal, when the output features at time point t pass through multiple layers of TimeMixer blocks, the original vibration pattern (such as the low-frequency components of subway vibration) is retained through residual connections, while adding local time series features (such as the impact waveform of excavation vibration); in adversarial domain adaptation, residual connections ensure that the differences in vibration characteristics between the source domain (Line A concrete track) and the target domain (Line B steel rail) are effectively modeled, avoiding the loss of key information during domain alignment; by fusing shallow details (such as signals after high-frequency noise suppression) with deep abstract features (such as spatiotemporal dependencies), the model's ability to discriminate intrusion events such as "climbing" and "excavation" is enhanced. The residual connection formula is:

[0153] ;

[0154] Among them, y l is the output of the current convolutional layer, y l-1 is the output of the previous convolutional layer (the input of the current convolutional layer), f(y l-1 ) is the result of the current convolutional layer processing the input.

[0155] (3) Calculate the attention weights to obtain the attention-weighted features. The attention weights are the core component of the attention mechanism and are used to dynamically assign the importance weights of different input features. Their functions include: automatically focusing on key areas in the input data (such as specific time points or frequency bands of vibration signals) according to task requirements, suppressing irrelevant noise; capturing the correlation between long-range features (such as the continuous waveform of subway vibration and the intermittent pattern of climbing vibration); extracting complementary features from different subspaces (such as time domain, frequency domain, and energy domain) through the multi-head attention mechanism to enhance the model's expressive power. The calculation formula is:

[0156] ;

[0157] in, , T is the parameter for adjusting weight distribution, α ij is the attention weight of the i-th query on the j-th key, d k is the dimension of the key vector, M j is the key vector of the j-th time step, is the query vector transpose of the i-th time step, e ik is the similarity set between the query and all keys.

[0158] (4) Perform bidirectional temporal modeling on the attention weighted features to obtain the predicted probability; wherein the gating mechanism is:

[0159] ;

[0160] ;

[0161] Among them, p t is the output of the forget gate, which means the predicted probability; i t is the input gate output; σ is the activation function, preferably the sigmoid function; W f 、W i is the weight matrix, corresponding to the weight parameters of the forget gate and input gate respectively; h t-1 is the hidden state at the previous moment; x t is the eigenvalue of the element in the time-frequency domain feature matrix at time point t; b f 、b i are bias terms, corresponding to the bias parameters of the forget gate and input gate, respectively, and are used to adjust the activation threshold of the gate.

[0162] (5) During the model training process, there is a class imbalance problem due to the significant difference in sample size between subway vibration (strong persistence, many samples) and climbing / excavation vibration (intermittent, few samples); at the same time, the vibration characteristics of different subway lines (such as concrete track line A and steel rail line B) are distributed differently, that is, there is also a cross-domain adaptation problem. Therefore, it is necessary to calculate the difference between the predicted result and the true label through a loss function (such as focal loss), clarify the optimization target, and convert the loss value into a gradient signal to drive the model parameter update and gradually improve the classification accuracy, thereby obtaining a multi-scale time series classification model. In the multi-scale time series classification model, the use of the loss function runs through the entire training process, and the loss value calculation formula is:

[0163] ;

[0164] Among them, p t is the model's predicted probability of the correct category; α t To balance the loss weights of different categories; γ is the weight for adjusting the difficult and easy samples; For comprehensive p t , α t The final loss value after γ is used for back propagation optimization model. After that, the optimizer can be used to update the parameters according to the gradient.

[0165] To address the differences in vibration characteristics between different subway lines (e.g., concrete rails on Line A and steel rails on Line B), we employed adversarial domain adaptation: a domain discriminator: a fully connected network (two layers, 128 nodes per layer) distinguishes between the source domain (Line A) and the target domain (Line B); a gradient reversal layer: a gradient reversal layer inserted after the feature extractor maximizes the domain discriminator loss to obfuscate domain features; and joint training: alternately optimizing the classification loss (cross entropy) and the domain discrimination loss (adversarial loss). Ultimately, the target domain test accuracy increased by 12.3%. The key technologies implemented in this phase are as follows:

[0166] (1) Maximum mean difference (MMD): MMD is used to quantify the difference between two data distributions (such as the source domain and the target domain). By mapping the data to a high-dimensional reproducing kernel Hilbert space (RKHS) and calculating its mean distance, MMD can determine whether the distributions are similar. In optical cable intrusion detection, MMD guides adversarial domain adaptation training to minimize the distribution difference of vibration characteristics between the source domain (concrete track of Line A) and the target domain (steel rail of Line B), forcing the model to learn domain-independent common features, thereby improving the generalization ability across line scenarios and solving the problem of classification performance degradation caused by environmental differences. The calculation formula is:

[0167] ;

[0168] Where MMD is the maximum mean difference, n s is the number of source domain samples, is the kernel space mapping function of the source domain sample, is the kernel space mapping function of the target domain sample.

[0169] (2) Adversarial loss: In the multi-scale time series classification model, the purpose of calculating the adversarial loss is to measure the effect of domain alignment. By maximizing the loss of the domain discriminator to confuse the domain features, the model is forced to learn common features that are irrelevant to the domain, thereby reducing the distribution difference of vibration features between the source domain (such as the concrete track of Line A) and the target domain (such as the steel rail of Line B), improving the generalization ability of the model in different subway line scenarios, and solving the problem of classification performance degradation caused by environmental differences. The calculation formula is:

[0170] ;

[0171] Among them, L adv To combat losses; is the expectation of the target domain samples, ensuring that the loss calculation covers all target data; D(G(x)) is the discriminant function, which is used to estimate the source domain probability of the target domain features and measure the domain alignment effect; G(x) is the feature extraction function.

[0172] (3) Contrast loss: The core of contrast loss is to shorten the feature distance of similar samples and push the feature distance of unrelated samples farther, so as to encourage the model to learn more discriminative feature expressions. In optical cable intrusion detection, it can be used to enhance the feature consistency of similar vibration signals (such as the same intrusion type) and distinguish the feature differences of different types of signals, thereby improving the classification accuracy and robustness of the model for intrusion events. Especially in small sample or noise interference scenarios, it can effectively alleviate the overfitting problem and enhance the interpretability of the feature space. The calculation formula is:

[0173] ;

[0174] Among them, L cont is the contrast loss value; y ij is the sample pair label; d ij is the Euclidean distance of sample pairs in the feature space; m is the margin threshold.

[0175] The multi-scale time series analysis model undergoes channel pruning (removing 20% of redundant convolution kernels) and dynamic quantization, reducing the model size to 35% of its original size. Edge deployment: Equipped with an NVIDIA Jetson Xavier NX edge device and an integrated inference engine, single-shot inference latency is ≤8ms. Dynamic threshold alarms: Adaptively adjust the classification confidence threshold (default 0.9, lowered to 0.7 in high-noise conditions) based on ambient noise levels (e.g., increased noise in rainy or snowy weather). Alarm rules: trigger a real-time alarm if three consecutive frames are detected as "mining" or a single frame is detected as "climbing" with a confidence level greater than 0.95. Key technologies implemented in this phase are as follows:

[0176] (1) Model quantization and linear quantization: Model quantization converts high-precision floating-point parameters (such as 32 bits) into low-precision values (such as 8-bit integers), compressing the model size and improving computational efficiency, making it suitable for mobile or edge devices. Linear quantization is a common method that maps the floating-point range to discrete integer values based on uniform intervals. While reducing storage and computing power consumption, it maintains the numerical distribution characteristics as much as possible, balancing model inference speed and accuracy loss, thereby optimizing deployment performance in resource-constrained scenarios.

[0177] The calculation formula is:

[0178] ;

[0179] in, , Q(x) is the integer value after quantization; round(x / Δ) is the discrete mapping function from floating point to integer; Δ is the quantization step size; x max 、x min is the dynamic range of the floating-point weight, ensuring that quantization covers all data; b is the number of quantization bits;

[0180] (2) Dynamic threshold adjustment: Dynamic threshold adjustment adaptively updates the classification / detection threshold by real-time analysis of data distribution or model output, solving the problem of false alarms or missed detections caused by environmental changes (such as noise fluctuations and signal offsets) in fixed thresholds. In optical cable intrusion detection, it can dynamically optimize the decision boundary based on the vibration signal characteristics, balance sensitivity and specificity, and enable the model to maintain stable performance in complex scenarios (such as different lines and weather conditions), thereby improving the recognition accuracy of intrusion events and the robustness of the system. The calculation formula is:

[0181] ;

[0182] Among them, τ is the dynamically adjusted threshold; τ0 is the default initial threshold; α is the noise sensitivity coefficient, and the larger α is, the more sensitive the threshold is to noise changes; N is the current environmental noise intensity; N0 is the noise baseline threshold.

[0183] (3) Real-time processing delay: Real-time processing delay refers to the time interval from the system receiving data to outputting results, which directly affects the response speed and decision-making timeliness. In scenarios such as optical cable intrusion detection, low latency can quickly identify anomalies and trigger alarms, reducing the risk of missed or false alarms of intrusion events and ensuring the real-time performance of the security system. By optimizing algorithm complexity, hardware acceleration or data transmission efficiency, latency can be reduced, avoiding data backlogs and decision failures caused by processing delays, thereby improving system reliability and user experience. The calculation formula is:

[0184] ;

[0185] Among them, t total For real-time processing delay, t preprocess is the preprocessing time, t feature is the time consumption for feature extraction, t inference It takes time to reason about the model.

[0186] The technical solution of this embodiment acquires vibration signals, preprocesses them to remove noise and other interference, and then extracts features to obtain high-dimensional feature vectors, making the analysis of vibration signals more detailed and accurate, thereby accurately identifying vibrations of different types and intensities. Classification data is determined based on the high-dimensional feature vectors, which are then used to determine the vibration state of the optical cable and control the operating state of the alarm device, thus achieving intelligent judgment and decision-making. The system can automatically determine whether to issue an alarm based on the analysis results without the need for human intervention, improving the timeliness and accuracy of monitoring and reducing the impact of human factors. The entire process can quickly process vibration signals, and once an abnormality is detected, the alarm device can be promptly controlled to issue an alarm, allowing relevant personnel to quickly take measures to prevent damage to the optical cable and ensure the safe and stable operation of the communication network.

[0187] In an optional embodiment of the present invention, in step 15, determining the vibration state of the optical cable according to the classification data may include:

[0188] Step 151, parsing the classification data to obtain a structured classification result;

[0189] Step 152, performing logical judgment on the classification results according to preset rules to obtain a preliminary vibration state;

[0190] Step 153 : performing false alarm suppression processing on the preliminary vibration state according to historical data, obtaining an optimized vibration state and outputting it.

[0191] In this embodiment, the format of the classification data is first parsed, the category distribution of each time window is counted, and key information is extracted to obtain a structured classification result, such as: {category: "intrusion event", confidence: 0.85, time window: t1~t2}.

[0192] If the classification confidence exceeds a preset threshold, such as confidence > 0.8, it is marked as a valid event. The classification results of multiple time windows are checked for continuity to eliminate instantaneous noise interference. For example, three consecutive windows are marked as "intrusion events." Combined with the spatial positioning data of the fiber optic sensor, such as the physical location of the event, it is verified whether it matches the known sensitive area. Finally, the preliminary vibration status is output, such as: {Status: "Suspected intrusion", Location: "Pile number K10+200", Confidence: 0.9}.

[0193] After obtaining the preliminary vibration status, the system checks whether the current event resembles historical false alarm patterns (such as noise during a specific construction period). If the system deploys multiple fiber optic sensors, the classification results of the other sensors must be cross-validated. The judgment threshold is then dynamically adjusted based on the ambient noise baseline, and the optimized vibration status is output, such as: {Status: "Confirmed Intrusion", Location: "Pile K10+200", Confidence: 0.95}. If the status is "Intrusion," an alarm is triggered and the event details (time, location, and confidence level) are recorded. If the status is "Non-Intrusion," only a log is recorded for subsequent analysis.

[0194] like Figure 3 As shown, an embodiment of the present invention further provides an optical cable intrusion detection device 30 based on distributed optical fiber vibration sensing, comprising:

[0195] An acquisition module 31 is configured to acquire a first vibration signal within a preset amplitude range;

[0196] The processing module 32 is configured to perform frequency filtering on the first vibration signal to obtain a second vibration signal; perform feature extraction on the second vibration signal to obtain a high-dimensional feature vector; and perform multi-scale time series classification on the high-dimensional feature vector to obtain classification data.

[0197] The determination module 33 is configured to determine the vibration state of the optical cable according to the classification data.

[0198] Optionally, the acquisition module 31 is specifically configured to:

[0199] The original vibration signal is obtained by installing distributed optical fiber vibration sensors at preset distances along the subway track;

[0200] According to a preset dynamic amplitude threshold, the original vibration signal is subjected to amplitude screening processing to obtain a first vibration signal.

[0201] Optionally, the processing module 32 is specifically configured to:

[0202] Get the noise figure;

[0203] performing frequency band decomposition on the first vibration signal according to the noise coefficient to obtain a vibration sequence signal including vibration signals of multiple frequency bands;

[0204] Reconstructing the noise coefficient to obtain a target noise coefficient;

[0205] According to the target noise coefficient, frequency filtering is performed on the vibration sequence signal to obtain a second vibration signal.

[0206] Optionally, the processing module 32 is further specifically configured to:

[0207] performing frame division and windowing processing on the second vibration signal to obtain a windowed short time frame;

[0208] Performing time-frequency domain feature extraction processing on the windowed short time frame to obtain a time-frequency feature matrix;

[0209] Embedding coding is performed on the time-frequency feature matrix to obtain a high-dimensional feature vector.

[0210] Optionally, the processing module 32 is further specifically configured to:

[0211] The high-dimensional feature vector is input into a trained multi-scale time series classification model for processing to obtain classification data.

[0212] Optionally, the multi-scale time series classification model is trained through the following process:

[0213] Obtaining original vibration sample signals through distributed optical fiber vibration sensors installed along the subway track at preset distances;

[0214] Performing amplitude screening processing on the original vibration sample signal to obtain a first vibration sample signal;

[0215] performing frequency filtering on the first vibration sample signal to obtain a second vibration sample signal;

[0216] Classify the second vibration signal according to the subway vibration sample type label, the mining vibration sample type label, and the climbing vibration type label to obtain a vibration event sample library;

[0217] Performing feature extraction on sample data in the vibration event sample library to obtain a time-frequency domain feature matrix;

[0218] A multi-scale time series classification model is obtained according to the time-frequency domain feature matrix.

[0219] Optionally, the time-frequency domain feature matrix is processed to obtain fused discriminant features, including:

[0220] Performing multi-scale time series feature extraction on the time-frequency domain feature matrix to obtain multi-scale eigenvalues;

[0221] Performing spatiotemporal attention weighting on the multi-scale feature values to obtain attention-weighted features;

[0222] Performing bidirectional temporal modeling on the attention weighted features to obtain a prediction probability;

[0223] According to the predicted probability, focal loss optimization and parameter update are performed on the classification model with preset parameters to obtain a multi-scale time series classification model.

[0224] It should be noted that this device is a device corresponding to the above method, and all implementation methods in the above method embodiment are applicable to this embodiment and can achieve the same technical effect.

[0225] like Figure 4 As shown, an embodiment of the present invention further provides a computing device 40, including a processor 41, a memory 42, and a program or instruction stored in the memory 42 and executable on the processor 41. When the program or instruction is executed by the processor 41, the various processes of the embodiment of the optical cable intrusion detection method based on distributed optical fiber vibration sensing are implemented, and the same technical effects are achieved. To avoid repetition, they are not described here. It should be noted that the computing device in the embodiment of the present invention includes the mobile electronic device and the non-mobile electronic device described above.

[0226] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0227] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0228] In the embodiments provided by the present invention, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed can be through some interface, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0229] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0230] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0231] If the functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product, stored in a storage medium, includes instructions for causing a computer device (such as a personal computer, server, or network device) to execute all or part of the steps of the various embodiments of the method of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, mobile hard drives, ROM, RAM, magnetic disks, or optical disks.

[0232] In addition, it should be pointed out that in the apparatus and method of the present invention, it is obvious that each component or each step can be decomposed and / or recombined. These decompositions and / or recombinations should be regarded as equivalent solutions of the present invention. Moreover, the steps of performing the above-mentioned series of processing can naturally be performed in chronological order according to the order of description, but they do not necessarily need to be performed in chronological order, and some steps can be performed in parallel or independently of each other. For those of ordinary skill in the art, it can be understood that all or any steps or components of the method and apparatus of the present invention can be implemented in hardware, firmware, software or a combination thereof in any computing device (including a processor, storage medium, etc.) or a network of computing devices. This can be achieved by those of ordinary skill in the art using their basic programming skills after reading the description of the present invention.

[0233] Therefore, the purpose of the present invention can also be achieved by running a program or a group of programs on any computing device. The computing device can be a well-known general-purpose device. Therefore, the purpose of the present invention can also be achieved simply by providing a program product containing program code for implementing the method or device. That is to say, such a program product also constitutes the present invention, and the storage medium storing such a program product also constitutes the present invention. Obviously, the storage medium can be any well-known storage medium or any storage medium developed in the future. It should also be pointed out that in the device and method of the present invention, it is obvious that each component or each step can be decomposed and / or recombined. These decompositions and / or recombinations should be regarded as equivalent schemes of the present invention. In addition, the steps of performing the above-mentioned series of processing can naturally be performed in chronological order according to the order of description, but do not necessarily need to be performed in chronological order. Certain steps can be performed in parallel or independently of each other.

[0234] The above is a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for detecting optical cable intrusion based on distributed optical fiber vibration sensing, characterized in that: include: Acquire a first vibration signal within a preset amplitude range; performing frequency filtering on the first vibration signal to obtain a second vibration signal; performing feature extraction processing on the second vibration signal to obtain a high-dimensional feature vector; Performing multi-scale time series classification processing on the high-dimensional feature vector to obtain classification data; determining a vibration state of the optical cable based on the classification data; The high-dimensional feature vector is subjected to multi-scale time series classification processing to obtain classification data, including: The high-dimensional feature vector is input into a trained multi-scale time series classification model for processing to obtain classification data; the multi-scale time series classification model is trained through the following process: Obtaining original vibration sample signals through distributed optical fiber vibration sensors installed along the subway track at preset distances; Performing amplitude screening processing on the original vibration sample signal to obtain a first vibration sample signal; performing frequency filtering on the first vibration sample signal to obtain a second vibration sample signal; Classify the second vibration signal according to the subway vibration sample type label, the mining vibration sample type label, and the climbing vibration type label to obtain a vibration event sample library; Performing feature extraction on sample data in the vibration event sample library to obtain a time-frequency domain feature matrix; Multi-scale time series feature extraction is performed on the time-frequency domain feature matrix to obtain multi-scale eigenvalues; the calculation formula is: ； in, y [ t ] is the time point after the convolution operation t Output features of J is the convolution kernel size; w [ k ] is the convolution kernel k The weight parameter of each position; k is a natural number; x [ td⋅k ] is the element in the time-frequency domain feature matrix at the time point td⋅k The characteristic value of d is the expansion factor; The multi-scale feature values are subjected to spatiotemporal attention weighting to obtain attention weighted features; the calculation formula is: ； in, α ij For the i Query pair j The attention weight of each key; e ik is the similarity set between the query and all keys; T To adjust the weight distribution parameters; ; For the i The query vector transpose of time steps; M j For the j The key vector of time steps; d k is the dimension of the key vector; The attention weighted features are subjected to bidirectional temporal modeling to obtain the predicted probability; the calculation formula is: ； ； in, p t is the output of the forget gate, which means the predicted probability; i t is the output of the input gate; σ is the activation function; W f 、 W i is the weight matrix, corresponding to the weight parameters of the forget gate and input gate respectively; h t-1 is the hidden state at the previous moment; x t is the element in the time-frequency domain feature matrix at the time point t The characteristic value of b f 、 b i is the bias term, corresponding to the bias parameters of the forget gate and input gate, respectively, used to adjust the activation threshold of the gate; According to the predicted probability, the classification model with preset parameters is subjected to focal loss optimization and parameter update to obtain a multi-scale time series classification model; wherein the loss function is calculated as follows: ； in, p t is the model’s predicted probability for the correct category; α t To balance the loss weights of different categories; γ To adjust the weight of difficult and easy samples; For comprehensive p t 、 α t 、 γ The final loss value after θ is used for back-propagation optimization of the model.

2. The optical cable intrusion detection method based on distributed optical fiber vibration sensing according to claim 1 is characterized in that: The obtaining of a first vibration signal within a preset amplitude range includes: The original vibration signal is obtained by installing distributed optical fiber vibration sensors at preset distances along the subway track; According to a preset dynamic amplitude threshold, the original vibration signal is subjected to amplitude screening processing to obtain a first vibration signal.

3. The optical cable intrusion detection method based on distributed optical fiber vibration sensing according to claim 1 is characterized in that: Performing frequency filtering on the first vibration signal to obtain a second vibration signal includes: Get the noise figure; performing frequency band decomposition on the first vibration signal according to the noise coefficient to obtain a vibration sequence signal including vibration signals of multiple frequency bands; Reconstructing the noise coefficient to obtain a target noise coefficient; According to the target noise coefficient, frequency filtering is performed on the vibration sequence signal to obtain a second vibration signal.

4. The optical cable intrusion detection method based on distributed optical fiber vibration sensing according to claim 1, characterized in that: Performing feature extraction processing on the second vibration signal to obtain a high-dimensional feature vector includes: performing frame division and windowing processing on the second vibration signal to obtain a windowed short time frame; Performing time-frequency domain feature extraction processing on the windowed short time frame to obtain a time-frequency feature matrix; Embedding coding is performed on the time-frequency feature matrix to obtain a high-dimensional feature vector.

5. An optical cable intrusion detection device based on distributed optical fiber vibration sensing, characterized in that: include: An acquisition module, configured to acquire a first vibration signal within a preset amplitude range; a processing module configured to perform frequency filtering on the first vibration signal to obtain a second vibration signal; perform feature extraction on the second vibration signal to obtain a high-dimensional feature vector; and perform multi-scale time series classification on the high-dimensional feature vector to obtain classified data; a determination module, configured to determine a vibration state of the optical cable based on the classification data; The high-dimensional feature vector is subjected to multi-scale time series classification processing to obtain classification data, including: The high-dimensional feature vector is input into a trained multi-scale time series classification model for processing to obtain classification data; the multi-scale time series classification model is trained through the following process: Obtaining original vibration sample signals through distributed optical fiber vibration sensors installed along the subway track at preset distances; Performing amplitude screening processing on the original vibration sample signal to obtain a first vibration sample signal; performing frequency filtering on the first vibration sample signal to obtain a second vibration sample signal; Classify the second vibration signal according to the subway vibration sample type label, the mining vibration sample type label, and the climbing vibration type label to obtain a vibration event sample library; Performing feature extraction on sample data in the vibration event sample library to obtain a time-frequency domain feature matrix; Multi-scale time series feature extraction is performed on the time-frequency domain feature matrix to obtain multi-scale eigenvalues; the calculation formula is: ； in, y [ t ] is the time point after the convolution operation t Output features of J is the convolution kernel size; w [ k ] is the convolution kernel k The weight parameter of each position; k is a natural number; x [ td⋅k ] is the element in the time-frequency domain feature matrix at the time point td⋅k The characteristic value of d is the expansion factor; The multi-scale feature values are subjected to spatiotemporal attention weighting to obtain attention weighted features; the calculation formula is: ； in, α ij For the i Query pair j The attention weight of each key; e ik is the similarity set between the query and all keys; T To adjust the weight distribution parameters; ; For the i The query vector transpose of time steps; M j For the j The key vector of time steps; d k is the dimension of the key vector; The attention weighted features are subjected to bidirectional temporal modeling to obtain the predicted probability; the calculation formula is: ； ； in, p t is the output of the forget gate, which means the predicted probability; i t is the output of the input gate; σ is the activation function; W f 、 W i is the weight matrix, corresponding to the weight parameters of the forget gate and input gate respectively; h t-1 is the hidden state at the previous moment; x t is the element in the time-frequency domain feature matrix at the time point t The characteristic value of b f 、 b i is the bias term, corresponding to the bias parameters of the forget gate and input gate, respectively, used to adjust the activation threshold of the gate; According to the predicted probability, the classification model with preset parameters is subjected to focal loss optimization and parameter update to obtain a multi-scale time series classification model; wherein the loss function is calculated as follows: ； in, p t is the model’s predicted probability for the correct category; α t To balance the loss weights of different categories; γ To adjust the weight of difficult and easy samples; For comprehensive p t 、 α t 、 γ The final loss value after θ is used for back-propagation optimization of the model.

6. A computing device, characterized in that include: A processor and a memory storing a computer program, wherein when the computer program is executed by the processor, the method according to any one of claims 1 to 4 is performed.

7. A computer-readable storage medium, characterized in that The device stores instructions, which, when executed on a computer, enable the computer to execute the method according to any one of claims 1 to 4.

Images