Distributed optical vibration sensing early warning verification method for island multi-source environment

Abstract

The invention relates to the technical field of optical fiber sensing, in particular to a distributed optical vibration sensing early warning verification method for an island multi-source environment. Signals collected in different environments and different states are preliminarily divided into threat signals and non-threat signals, the effectiveness of data is further verified through frequency spectrum transformation, low-frequency noise and natural environment interference signals such as sea waves and wind sound are filtered out from the signals through a 5Hz high-pass filter, high-frequency noise is further suppressed by applying a wavelet denoising technology, and the non-threat signals are obtained. And useful signal features are reserved. A variational mode decomposition technology is adopted to carry out multi-scale decomposition on signals, and multiple classic features of a mode function obtained after decomposition are selected to extract feature vectors. Low-frequency effective signals are extracted, and high-frequency noise is removed, so that essential characteristics of the signals can be better captured. And the data subjected to signal processing is input into a CNN-LSTM-Attention model for deep learning analysis.

Description

Technical Field

[0001] The present invention relates to the field of optical fiber sensing technology, and more specifically, to a distributed optical vibration sensing early warning verification method for an island multi-source environment, in particular an event recognition method based on an optical fiber distributed sensing system and a combination of multiple advanced signal processing methods and deep learning. Background Art

[0002] As the global security situation becomes increasingly severe, the need for security in remote areas, especially isolated areas such as islands, is becoming increasingly urgent. Island areas are often key strategic locations. However, due to their remote location, weak security capabilities, and frequent exposure to harsh natural environments such as waves and sand, traditional monitoring methods often cannot provide rapid response and accurate identification of intrusion incidents.

[0003] Traditional island security monitoring methods mostly rely on manual patrols, video surveillance, or remote sensing. However, manual patrols struggle to provide all-weather, all-encompassing monitoring and are susceptible to environmental conditions, fatigue, and other factors. While video surveillance systems can provide real-time images, their effectiveness is limited in inclement weather or complex terrain. Furthermore, remote sensing monitoring methods, such as satellite imagery, while capable of providing wide coverage, are often costly and inefficient, making them unsuitable for rapid response missions.

[0004] With the development of fiber-optic sensing technology, fiber-optic distributed sensing systems (DAS) have become a research hotspot in the field of intrusion monitoring due to their high sensitivity, high spatial resolution, long-distance detection capabilities, and long-term stable operation. This technology uses optical fibers as sensors, exploiting changes in internal scattered signals to monitor external vibrations, temperature fluctuations, and other factors. However, while DAS technology has made significant progress in environmental monitoring, accurately distinguishing useful intrusion signals from background noise in complex environments, particularly those with multiple interference sources such as islands, remains a pressing challenge.

[0005] Currently, intrusion event identification in island environments relies heavily on traditional signal processing methods, such as time-domain analysis, frequency-domain analysis, and waveform matching. While these methods can extract certain signal features, they often struggle to effectively separate useful signals from background noise in complex environments. For example, the sound of waves, wind, and natural vibrations in the sandy beach area can overlap with intrusion signals, significantly reducing event classification accuracy. Furthermore, traditional methods have limited ability to extract spatiotemporal features, especially in complex environments like islands. The time-varying and spatially localized nature of signals hinders efficient processing and classification.

[0006] As data volume and complexity increase, traditional single models often struggle to meet the challenges of modern signal processing and pattern recognition. To address this issue, hybrid models combining multiple deep learning techniques have gained widespread application. In fiber-optic distributed sensing systems, signals contain both spatial and temporal information. To specifically process fiber-optic signals, a deep learning network architecture combining CNNs, LSTMs, and a self-attention mechanism was selected. This architecture addresses the problem of multi-source environmental event recognition, fully exploiting key signal features and achieving accurate identification of complex environmental events. CNNs extract local spatial features from sensor signals and identify signal variation patterns at different locations. LSTMs process time series data in the temporal channel, modeling temporal dependencies and capturing dynamic signal features. To further improve model performance, a self-attention mechanism dynamically adjusts the focus on different time steps or features, enabling the model to focus on the most critical information for the task, thereby enhancing its ability to analyze complex data. The adopted CNN-LSTM-Attention mechanism can simultaneously process the spatial and temporal dimensions of fiber optic sensing signals, and significantly improves the ability to recognize and predict environmental events through efficient feature extraction, time series modeling and dynamic information selection.

[0007] Fiber-optic distributed acoustic sensing technology has been widely used in ocean monitoring, but detecting intrusion events in island environments still faces numerous challenges. The unique beach geology and wave noise of island environments complicate event identification. Currently, research on intrusion event detection in complex, multi-source environments such as land, beach, and underwater is limited. This multi-source environment results in low signal recognition accuracy, necessitating the development of a new intelligent recognition method. Summary of the Invention

[0008] To address the challenges of intrusion event detection in island environments, this paper proposes an intelligent submarine cable sensing method based on a fiber-optic distributed sensing system combined with multiple advanced signal processing and deep learning methods. Specifically, the paper preprocesses and classifies various events. Signals collected in different environments and under different conditions are initially classified as threat signals and non-threat signals. Spectral transformation is then used to further verify the data validity. The signals are then filtered through a 5Hz high-pass filter to remove low-frequency noise and natural environmental interference such as waves and wind. Wavelet denoising is then applied to further suppress high-frequency noise, preserving useful signal features. Next, variational mode decomposition is used to perform multi-scale decomposition of the signals. Feature vectors are extracted from the resulting modal functions using various classic features. This extracts low-frequency valid signals and removes high-frequency noise to better capture the signal's essential characteristics. The processed data is then fed into a CNN-LSTM-Attention model for deep learning analysis, enabling intelligent signal recognition for fiber-optic distributed sensing systems in island environments. Finally, ablation experiments were conducted on different deep learning models, analyzing key parameters such as accuracy, precision, and F1 score, demonstrating the advanced nature of the method employed in this paper.

[0009] The core innovation of this invention lies in the proposal of an intelligent submarine cable sensing method that can be used in island environments. Fiber-optic distributed sensing technology is widely used in the field of ocean monitoring, but it is rarely used in the field of island signal recognition. This is because the island environment is different from the land environment. Different terrains have multiple influencing factors such as waves and sand buffering, which pose challenges to event identification. The method of the present invention is based on fiber-optic distributed sensing technology. It selects a variety of advanced data processing methods based on the different characteristics of different island environments. Combined with the constructed deep learning model, it achieves a high accuracy rate of 99.35% in signal recognition for the other three models. It provides another feasible way for fiber-optic signals to eliminate the influence of the classic island environment.

[0010] To achieve the above object, the present invention provides the following technical solutions:

[0011] The distributed optical vibration sensing early warning verification method for an island multi-source environment includes the following steps:

[0012] Step 1: Collect signal data. Fiber optic sensors deployed on the island acquire real-time vibration signals from different areas of the island. Trampling signals from three different island environments—the beach, underwater, and cobblestone paths—are collected and labeled as corresponding signals to create a training dataset.

[0013] Step 2: Preprocess the collected signal data, including: using a 5Hz high-pass filter to remove low-frequency noise and using wavelet transform to remove high-frequency noise;

[0014] Step 3: Variational mode decomposition is used for feature extraction, and VMD is used to decompose the signal and extract useful features;

[0015] Step 4: Input the preprocessed signal data into a deep learning model for event classification. The deep learning model combines CNN with LSTM and Attention mechanisms to build a CNN-LSTM-Attention network.

[0016] In step 5, the CNN-LSTM-Attention network first extracts the spatial features of the signal through CNN to capture local temporal-spatial patterns. Then, the LSTM is used to process the temporal dependencies of the signal and capture timing information. The Attention mechanism then weights the features output by the LSTM to highlight the most critical time steps or features. Finally, the fully connected layer maps the weighted features to event categories, completing the accurate classification of intrusion events.

[0017] This technical solution is further optimized, and step 1 specifically includes: a narrow-linewidth laser source generates a highly coherent continuous optical signal, which is modulated by an AOM and converted into a pulsed optical signal; then, the pulsed light is amplified by an EDFA to enhance the signal intensity, and the amplified pulse signal is further processed by a waveform amplifier to ensure signal quality; then, the signal is transmitted to the optical fiber sensing part through a circulator, and wavelength division multiplexing is performed through WDM to ensure the effective transmission and collection of signals of different wavelengths; the collected backscattered light signal is transmitted again through the circulator, received by a balanced photodetector, and converted into an electrical signal; the electrical signal is sent to a data acquisition card.

[0018] The present technical solution is further optimized. The high-frequency noise removal using wavelet transform in step 2 specifically includes three steps: 1) decomposition of the noisy signal, 2) threshold processing, and 3) signal reconstruction. The decomposition of the noise signal depends on the wavelet transform. Specifically, the relationship between the noisy signal and the wavelet transform signal is:

[0019]

[0020] Where g(t) is the noisy signal, φ(t) is the mother wavelet, p is the scaling factor, q is the translation factor, and W f (p,q) is the signal after wavelet transform, and its reconstruction formula is:

[0021]

[0022] in C φ is a constant.

[0023] This technical solution is further optimized. The wavelet threshold method used in the wavelet denoising is expressed as follows:

[0024]

[0025] where w p,q is the wavelet coefficient before denoising, is the wavelet coefficient after denoising, the threshold μ=θ 2 log(M), where M is the signal length and θ is an estimate of the noise level.

[0026] This technical solution is further optimized, and step 3 uses variational mode decomposition to extract features:

[0027] The signal is divided into different modal functions, and two modal functions with higher correlation coefficients are taken. After selecting the IMF to be used, it is necessary to select appropriate characteristic parameters to extract the eigenvector. The selected characteristic parameters mainly include center frequency CF, kurtosis K, peak PV, and short-term standard deviation STSD. First, CF is used to describe the main frequency position of a signal component, that is, the average frequency of energy in the signal spectrum; K is a statistic that describes the signal and is used to measure the steepness of the tail of the signal distribution. It reflects the "sharpness" of the data distribution compared to the normal distribution. Specifically, the kurtosis of the obtained IMF can be expressed by the following formula:

[0028]

[0029] where K i represents kurtosis, IMF i represents the i-th IMF signal, μ i is the mean of the ith IMF, E represents the expected value, and the peak value is used to represent the maximum absolute value s in the signal, which can reflect the extreme strength of the signal in order to enhance the signal discrimination;

[0030] STSD can be used to describe the statistics of the degree of signal fluctuation in a short time window and analyze the local dynamic characteristics of the signal. When calculating STSD, the signal is first divided into t different time windows. For each event window, the formula is as follows:

[0031]

[0032] Where STSD(t) is the short-term standard deviation of the t-th time window, x i Represents the signal value in the window, μ is the mean of the window, N represents the number of samples, and the four characteristic parameters selected above are obtained for the two retained IMFs. They are combined into the final eigenvector, and the eigenvector can be expressed as:

[0033] feature vector=[(CF1,K1,PV1,STSD1),(CF2,K2,PV2,STSD2)] (6)

[0034] By extracting their central frequency, kurtosis, peak value, short-term standard deviation and other classic features, the feature vector is extracted.

[0035] This technical solution is further optimized, and the CNN-LSTM-Attention network architecture is as follows:

[0036] S4.1 Input signal data shape is (1,8,473). The first convolution layer uses a 3x1 convolution kernel and outputs a feature map of 16 channels. The second convolution layer uses a 3x1 convolution kernel and outputs a feature map of 32 channels. The third convolution layer uses a 3x1 convolution kernel and outputs a feature map of 64 channels. Each convolution layer is followed by a ReLU activation function and a Dropout layer.

[0037] The S4.2 LSTM layer processes the features extracted by the CNN. The CNN output data is flattened and fed into the LSTM layer, where the time step represents each spatial channel. The LSTM layer uses a bidirectional structure with 64 hidden units to capture the temporal dependencies of the signal. Event classification is then performed through a fully connected layer.

[0038] S4.3 Attention: After the output of the LSTM layer, a self-attention mechanism is introduced to further improve model performance. The Attention mechanism calculates the attention weight of each time step, allowing the model to automatically focus on the most important time steps or spatial features. Specifically, the output of the LSTM layer passes through a fully connected layer to obtain an attention score for each time step, indicating the importance of the time step in event classification. Subsequently, these scores are normalized using the Softmax function to ensure that the sum of the attention weights of all time steps is 1.

[0039] The output of the S4.4 Attention part is classified through a fully connected layer with 128 neurons. The fully connected layer outputs four categories, representing different types of intrusion events. The classification results are used through an activation function to obtain the probability of each category, and finally output the most likely event type.

[0040] This technical solution is further optimized, the step S4.1, for the CNN network part: the CNN part uses a total of three convolutional layers, the first convolutional layer conv1 receives input data and applies a 3x1 convolution kernel with a step size of 1, outputting a feature map of 16 channels, each convolution kernel will slide in the time dimension to extract local spatial and temporal features, because the size of the convolution kernel is 3x1, it can capture the dependencies between adjacent time steps, and each convolution operation produces a new feature map, which contains the spatial pattern of each time step; the second convolutional layer conv2 receives the output from the first convolutional layer, continues to use a 3x1 convolution kernel with a step size of 1, and outputs a feature map of 32 channels, here, the convolution kernel continues to slide in the time dimension to further learn the detailed features of the signal. In this way, the second convolution layer The convolution can extract the spatial features of the signal from a more abstract level, so that the model can capture more complex patterns; the third convolution layer conv3 uses a 3x1 convolution kernel with a step size of 1, and outputs a feature map of 64 channels, which further enhances the feature extraction ability of the signal and can learn complex patterns in the signal from a higher level; the ReLU activation function is used after each convolution layer to increase the nonlinear characteristics, so that the model can learn more complex features. After the convolution process, the feature map is further downsampled through the pooling layer MaxPool2d to reduce the amount of calculation while retaining important features. The data after each convolution layer passes through the Dropout layer with a dropout rate of 0.5 to prevent overfitting during training. The Dropout operation randomly discards 50% of the neurons to force the model to learn more robust features.

[0041] This technical solution is further optimized, and S4.2, for the LSTM part: the data output from the CNN is flattened, and the LSTM layer accepts the flattened features as input. The model uses 64 hidden units and adopts a bidirectional LSTM, that is, the past and future time step information is considered at the same time, and the time step is selected as the spatial channel of the data. In this case, LSTM can capture the changing dependencies of the signal in different spatial channels and process the long-term dependencies in the data of each channel. After the input features pass through the LSTM network, the output representation corresponds to the feature representation of each time step.

[0042] This technical solution is further optimized, and S4.3, for the self-attention part: the Attention mechanism is used to weight the features of the LSTM output: first, the output of the LSTM layer is a tensor with a shape of (batch_size, time_steps, hidden_size), where time_steps represents the time step and hidden_size is the number of hidden units of the LSTM; then, each time step feature of the LSTM is linearly transformed through a fully connected layer to generate a scalar representing the attention score of each time step. This score reflects the importance of the time step in the current task, and then, the attention score is normalized by the Softmax function to ensure that the sum of the attention weights of all time steps is 1; finally, the Attention mechanism multiplies the LSTM output of each time step by the corresponding attention weight to calculate the weighted time series feature.

[0043] This technical solution is further optimized. S4.4, for the classification part: after passing through the fully connected layer, the weighted time series features are mapped to the final classification results. The dimension of the output layer is 4, representing 4 types of intrusion events. Through the Softmax activation function, the output layer maps the results of the model to the probability value of each event type. Finally, the model predicts the most likely event type based on the probability value.

[0044] Different from the existing technology, the above technical solution has the following advantages:

[0045] First, a complete set of signal preprocessing and feature extraction methods is adopted to ensure that valuable signal features are extracted from complex environments. Signal preprocessing includes 5Hz high-pass filtering to remove low-frequency noise, such as natural environmental interference such as waves and wind, which helps to improve the signal-to-noise ratio of the signal. Next, the signal is further processed using wavelet denoising technology. The signal is decomposed into different frequency bands through multi-scale analysis, effectively removing high-frequency noise and retaining the local characteristics of the signal. It is particularly suitable for non-stationary signals and mutation signals. Finally, variational mode decomposition is applied to the multi-scale decomposition of the signal. The signal is decomposed through an adaptive mechanism and converted into an intrinsic mode function. The correlation coefficient is used as a criterion to retain the effective mode and extract features with high correlation from it, which helps to capture the essential characteristics of the signal, especially in complex multi-source interference environments. It can improve the accuracy and robustness of feature extraction. Secondly, the present invention combines advanced deep learning models, especially the CNN-LSTM-Attention network architecture, which can simultaneously process the spatial and temporal dimensions in the fiber optic sensing signal. The incorporation of a self-attention mechanism, which weights different time steps and features, enables the model to automatically focus on the most critical information, thereby improving the ability to identify and predict intrusion events in complex environments. By combining efficient signal preprocessing, precise feature extraction, and a deep learning model, the present invention achieves an event accuracy rate of 99.35%, significantly improving the accuracy of intrusion event recognition in complex environments such as islands, and is particularly capable of coping with the effects of multi-source interference and environmental noise. This method not only improves signal processing accuracy but also optimizes event classification performance, providing a new solution for real-time, accurate intrusion warnings. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A flow chart of the method adopted by the present invention;

[0047] Figure 2 Schematic diagram of event recognition and classification based on a multi-source island environment according to the present invention;

[0048] Figure 3 is a schematic diagram of a fiber optic DAS system;

[0049] Figure 4 A statistical chart for classifying and labeling events of samples;

[0050] Figure 5 Result diagram of spectrum transformation of signals in different regions;

[0051] Figure 6 Flowchart for performing variational mode decomposition method;

[0052] Figure 7 A schematic diagram of selecting characteristic parameters based on a unique environment;

[0053] Figure 8 Schematic diagram of eigenvectors after modal decomposition;

[0054] Figure 9 This is the architecture diagram of the constructed CNN-LSTM-Attention network;

[0055] Figure 10 The final result indicators after training and the comparison chart with the other three networks;

[0056] Figure 11 This is a diagram showing the impact of each feature on model performance;

[0057] Figure 12 Schematic diagram of the classification results of three deep learning models. DETAILED DESCRIPTION

[0058] In order to explain the technical content, structural features, achieved objectives and effects of the technical solution in detail, the following is a detailed description in conjunction with specific embodiments and accompanying drawings.

[0059] The present invention proposes a distributed optical vibration sensing early warning verification method for an island multi-source environment, see Figure 1 The figure shows a flow chart of the early warning method, which mainly includes the following steps:

[0060] Step 1: Build a DAS system and collect data: Figure 3 The figure is a schematic diagram of the fiber DAS system. A narrow-linewidth laser source generates a highly coherent continuous optical signal. This optical signal is modulated by an acousto-optic modulator (AOM) and converted into a pulsed optical signal. Subsequently, the pulsed light is amplified by an erbium-doped fiber amplifier (EDFA) to enhance the signal strength. The amplified pulse signal is further processed by a waveform amplifier to ensure signal quality. Next, the signal is transmitted to the fiber optic sensing part through a circulator. The sensing fiber used is 1200m long and wavelength division multiplexing is performed through WDM to ensure the effective transmission and acquisition of signals of different wavelengths. The collected backscattered light signal is transmitted again through the circulator, received by a balanced photodetector (PD), and converted into an electrical signal. The electrical signal is sent to a data acquisition card (DAQ) with a sampling rate of 2304Hz for further analysis and processing.

[0061] To ensure high-quality signal acquisition and analysis, the system incorporates multiple filters to remove excess noise. Furthermore, the final stage of the system includes an isolator and radio frequency amplifier (RFA) for signal amplification and final processing. Signal processing and data analysis are performed on a personal computer using MATLAB 2021a for data processing, vibration signal extraction, and analysis.

[0062] During the experiment, a 1200-meter-long optical fiber was pulled from the system and initially laid multiple turns around the sensing system (84-93 meters). The fiber then passed along a cobblestone path on the island (175-179 meters) and continued across a beach area (220-222 meters). On the beach, the fiber was looped to expand the detectable area and enhance the system's sensitivity to intrusions. The fiber was then lowered into shallow water (111-114.5 meters) and looped there. In the initial stages of the experiment, researchers simulated different trampling events on the cobblestone path, beach, and shallow water, generating vibration signals from human movement in these areas. These simulated events were generated by the same researcher, ensuring consistency and comparability across the various events and avoiding errors introduced by human variability. Signals collected from these three environments, including those from the background environment, were classified into four main categories and further categorized based on threat. Among them, non-threatening events include background signals without people stepping on them, and the three types of stepping signals on cobblestone roads, beaches, and shallow water areas are all threatening signals.

[0063] After data acquisition, the signals were input into MATLAB 2021a and Python 3.8 for subsequent data processing. The sample event types and statistics selected in the experiment are as follows Figure 4 As shown in the figure, the experimental environment consists of three areas: the offshore area, the beach area, and the stone road area in the center of the island. Based on these three environments, four types of event data were collected. The events are divided into two categories: threat signals and non-threat signals. Non-threat signals represent interference signals in the natural environment, while threat signals represent human interference events. Specifically, underwater signals, human trampling signals, beach area human trampling signals, and stone road area human trampling signals were collected in each area. Figure 2 This is a schematic diagram of the event recognition and classification based on a multi-source island environment. The duration of each event sample is 8 seconds, and the number of events collected is 127 underwater, 129 on the beach, and 127 on the stone road. Figure 4 The acquired signal has been saved as an HDF5 file.

[0064] Step 2: Data preprocessing: In order to verify the validity of the collected data and confirm whether the collected signals have sufficient discrimination to clearly identify the target signal from the background noise, the collected signals are subjected to spectrum analysis. Figure 5 The results of spectrum transformation of signals in different regions are shown. It can be seen that in the signal containing only background noise, the amplitude of the signal in the range of 0-5Hz is higher, while it is relatively flat in other frequency ranges (such as Figure 5(a)). In the pedaling signal in the cobblestone area, the target signal is mainly concentrated in the range of 5-35Hz, and the signal amplitude is significantly higher than the background noise, with the maximum amplitude being about 8×10^5 (as shown in Figure 2). Figure 5 (b)); For the stepping signal in the beach area, the target signal is concentrated in the range of 5-45Hz, with a maximum amplitude of 4×10^5 (as shown in Figure 5 (c)); in the underwater stepping signal, the target signal is concentrated in the range of 5-95Hz, with a maximum amplitude of 4×10^5 (as shown in Figure 5 (d)). This provides a reference for the subsequent signal processing process.

[0065] In the data processing process, the necessary data preprocessing is first performed to further extract eigenvectors. It is known that the noise signal exhibits high amplitude in the 0-5 Hz frequency band, indicating that environmental noise such as wind and waves is primarily concentrated in this frequency range. Therefore, to effectively suppress the influence of low-frequency phase noise, the signal is subjected to a 5 Hz high-pass filter. The filter is set to an 8th-order IIR high-pass filter to ensure filter performance. The filter applies zero-phase filtering to each column of data, eliminating phase delay and preserving the signal waveform. The sampling frequency is set to 2304 Hz.

[0066] Subsequently, the signal is further processed using wavelet denoising technology. Wavelet denoising is a widely used signal denoising method. It originated from Fourier transform. However, compared with traditional Fourier analysis denoising, it is more suitable for the denoising problem of non-stationary signals. Wavelet denoising is generally divided into three steps: 1) decomposition of noisy signals, 2) threshold processing and 3) signal reconstruction. The decomposition of noisy signals depends on wavelet transform. Wavelet transform is a time-frequency analysis method of signals with the characteristics of multi-resolution analysis. Specifically, the relationship between noisy signals and wavelet transformed signals is:

[0067]

[0068] Where g(t) is the noisy signal, φ(t) is the mother wavelet, p is the scaling factor, q is the translation factor, and W f (p,q) is the signal after wavelet transformation. Its reconstruction formula is:

[0069]

[0070] in C φ is a constant.

[0071] The modulus maximum method, spatial correlation method and wavelet threshold method are the three main methods of wavelet filtering denoising. Compared with the first two methods, the wavelet threshold method has the advantages of simple algorithm, small amount of calculation and good filtering effect. It is especially suitable for signals with low signal-to-noise ratio. Therefore, the research and application of wavelet threshold method in signal filtering are very extensive. In this experiment, the soft threshold denoising method is adopted. It helps to smooth the signal by scaling the wavelet coefficients instead of completely discarding them, avoiding the mutations and distortions that may be caused by hard threshold denoising, better preserving the details of the signal, and providing a more natural denoising effect. The expression of soft threshold denoising is:

[0072]

[0073] where w p,q is the wavelet coefficient before denoising, is the wavelet coefficient after denoising, the threshold μ=θ 2 log(M), where M is the signal length and θ is the estimated noise level. In the experiments of this embodiment, the wavelet basis for wavelet denoising was coif-4, and the number of decomposition layers was set to 8. The coif-4 wavelet has good symmetry and compact support, which can reduce boundary effects and preserve the main features of the signal during denoising. The 8-layer decomposition further enhances the extraction of signal details while avoiding oversmoothing. These parameters were obtained based on previous experience.

[0074] Step 3: Use variational mode decomposition for feature extraction: After denoising, use variational mode decomposition to further process the signal. The feature extraction method used in this stage is shown in the figure below. Figure 6 As shown in the figure, the signal preprocessed with wavelet denoising is processed using variational mode decomposition to further reduce the data dimension, extract the sample signal features, and retain the effective frequencies. Four characteristic parameters are then extracted: center frequency, peak value, kurtosis, and short-term standard deviation. These features are then combined into a feature vector. Finally, these feature vectors are fed into a deep learning model for further classification or recognition tasks.

[0075] Variational mode decomposition (VMD) is an advanced method for analyzing nonstationary signals. It employs an adaptive mechanism to decompose the signal based on its local characteristics. During modal decomposition, the signal is gradually decomposed into a series of intrinsic mode functions (IMFs), each representing a distinct frequency component of the signal.

[0076] The signal is divided into different modal functions, and the two modal functions with the highest correlation coefficients are selected. After selecting the appropriate IMF, appropriate characteristic parameters need to be selected to extract the eigenvector. The characteristic parameters selected mainly include center frequency (CF), kurtosis (K), peak value (PV), and short-term standard deviation (STSD). The selection of these parameters is determined according to the specific environment. Figure 7 It reflects the four selected characteristic parameters and the environmental regions they focus on. Specifically, the center frequency determines the main frequency position of the signal and is used to distinguish signals from the three environments. The peak value reflects the maximum signal strength. The unique sandy environment of the beach and the shallow water may buffer the signal, causing the signal peak to drop. The kurtosis reflects the steepness of the signal. It focuses on the changes in the beginning and end characteristics of the signal in the beach environment. The short-term standard deviation reflects the periodic fluctuations of the signal. Factors such as gusts of wind in the cobblestone area on land and waves in the shallow area can cause periodic fluctuations in the signal. It is mainly used to distinguish signals from these two environments.

[0077] The CF describes the dominant frequency position of a signal component, that is, the average frequency of the energy in the signal spectrum. K is a statistic describing the signal, used to measure the steepness of the tail of the signal distribution. It reflects the degree of "sharpness" of the data distribution compared to the normal distribution. Specifically, the kurtosis of the obtained IMF can be expressed as follows:

[0078]

[0079] where K i represents kurtosis, IMF i represents the i-th IMF signal, μ i is the mean of the ith IMF, and E represents the expected value. The peak value is used to represent the maximum absolute value s in the signal, which can reflect the extreme strength of the signal to enhance the signal's discrimination.

[0080] STSD can be used to describe the statistical value of the signal's fluctuation degree within a short time window and analyze the local dynamic characteristics of the signal. When calculating STSD, first divide the signal into t different time windows. For each event window, there is a formula:

[0081]

[0082] Where STSD(t) is the short-term standard deviation of the t-th time window, x i Represents the signal value in the window, μ is the mean of the window, and N represents the number of samples. The four characteristic parameters selected above can be obtained for the two retained IMFs and combined into the final eigenvector. The eigenvector can be expressed as:

[0083] feature vector=[(CF1,K1,PV1,STSD1),(CF2,K2,PV2,STSD2)] (6)

[0084] By extracting classic features such as their center frequency, kurtosis, peak value, and short-term standard deviation, we can extract feature vectors. This method can accurately reflect the main characteristics of the signal, thereby improving the accuracy of the subsequent recognition system.

[0085] like Figure 8 As shown in Figure 2, VMD is used to decompose the original signal into five mode functions. Figure 8 (a) shows the five decomposed mode functions, labeled IMF1 to IMF5 from low frequency to high frequency, and their respective center frequencies are labeled. It can be observed that the center frequencies gradually increase from 3 Hz to 341.4 Hz. Figure 8 (b) shows the correlation coefficient between each IMF and the original signal. The correlation coefficient decreases from low frequency to high frequency, indicating that the effective signal is mainly concentrated in the low frequency component. Based on the experience of predecessors, the first two components with correlation coefficients exceeding 0.4 are selected and the remaining three components are discarded to further improve the efficiency of the model. The two components retained are IMF1 and IMF2, as shown in Figure 2. Figure 8 (c) shown. Figure 8 (d) shows the power spectral density of the signal before and after processing. It can be seen that the majority of the low-frequency signal is retained, while the power spectral density of the high-frequency signal is reduced by an average of 15.6 dB. The signal reduction is particularly significant at the center frequencies of the two discarded IMF components (from -17.84 dB to -11.71 dB at 100 Hz, and from -19.01 dB to -16.3 dB at 200 Hz). This result demonstrates excellent denoising capabilities, paving the way for subsequent feature vector extraction.

[0086] Step 4: CNN-LSTM-Attention network model: Figure 9 The CNN-LSTM-Attention network model used consists of three main components: First, the CNN extracts local spatial features from the fiber optic sensor signal and identifies the signal's changing patterns at different locations. Next, the LSTM processes the time series data, modeling the dependencies between signals over time and capturing dynamic features. To further enhance the model's performance, the self-attention mechanism dynamically adjusts the focus on different time steps or features, allowing the model to focus on the most critical information, thereby improving its ability to parse complex data. The network specifically includes the following:

[0087] S4.1. For the CNN network: Feature extraction yields a feature vector with 473 spatial channels, each with 8 features. The input data shape for this model is (473, 8). The CNN uses three convolutional layers. The first convolutional layer (conv1) receives the input data and applies a 3x1 convolution kernel with a stride of 1, outputting a 16-channel feature map. Each convolution kernel slides along the temporal dimension (473), extracting local spatial and temporal features. Because the convolution kernel is 3x1 in size, it can capture the dependencies between adjacent time steps. Each convolution operation produces a new feature map containing the spatial pattern of each time step. The second convolutional layer (conv2) receives the output from the first convolutional layer and continues to apply a 3x1 convolution kernel with a stride of 1, outputting a 32-channel feature map. Here, the convolution kernel continues to slide along the temporal dimension, further learning the detailed features of the signal. In this way, the second convolution layer can extract the spatial features of the signal at a more abstract level, allowing the model to capture more complex patterns. The third convolutional layer (conv3) uses a 3x1 convolution kernel with a stride of 1 and outputs a feature map of 64 channels. It further enhances the feature extraction capability of the signal and can learn complex patterns in the signal from a higher level. The ReLU activation function is used after each convolutional layer to increase nonlinear characteristics, allowing the model to learn more complex features. After the convolution process, the feature map is further downsampled through the pooling layer (MaxPool2d) to reduce the amount of computation while retaining important features. The data after each convolution layer passes through the Dropout layer with a dropout rate of 0.5 to prevent overfitting during training. The Dropout operation randomly discards 50% of the neurons, forcing the model to learn more robust features.

[0088] S4.2. For the LSTM layer: The CNN output data is flattened. The LSTM layer accepts the flattened features as input. The model uses 64 hidden units and employs a bidirectional LSTM (bidirectional = True), which considers both past and future time steps. Time steps are selected as spatial channels of the data. In this case, the LSTM can capture the dependencies between signal changes across different spatial channels and handle long-term dependencies within the data of each channel. After the input features pass through the LSTM network, the resulting output represents the feature representation corresponding to each time step.

[0089] S4.3. For the self-attention part: The Attention mechanism is used to weight the features output by the LSTM. First, the output of the LSTM layer is a tensor of shape (batch_size, time_steps, hidden_size), where time_steps represents the time step length and hidden_size is the number of hidden units in the LSTM. Each time step feature of the LSTM is then linearly transformed through a fully connected layer, generating a scalar representing the attention score for each time step. This score reflects the importance of that time step in the current task. The attention scores are then normalized using the Softmax function to ensure that the sum of the attention weights for all time steps is 1. Finally, the Attention mechanism multiplies the LSTM output of each time step by the corresponding attention weight to calculate the weighted time series features. In this way, the model can automatically focus on the most critical time steps, enhancing its attention to important information, thereby improving its ability to identify and predict intrusion events in complex environments.

[0090] S4.4. For the classification phase: After passing through the fully connected layer, the weighted time series features are mapped into the final classification results. The output layer has a dimension of 4, representing the four intrusion event types. Using the Softmax activation function, the output layer maps the model results into probability values for each event type. Ultimately, the model predicts the most likely event type based on the probability values.

[0091] Figure 10 This is the result obtained after training. As can be seen from this figure, the CNN-LSTM-Attention model performs better than the other three comparison models in multiple evaluation indicators. Specifically, the accuracy of CNN-LSTM-Attention reached 99.35%, far exceeding other models, showing its strong performance in event recognition tasks. At the same time, the F1 score is 0.9934, indicating that the model performs well in balancing accuracy and recall, and can effectively reduce false positive and false negative errors. The precision and recall rates are 0.9936 and 0.9934 respectively, further proving that the model has high accuracy and completeness in identifying intrusion events. In comparison, although the other three models (LSTM, CNN+LSTM and LSTM-Attention) also achieved high evaluation results, they were not as good as CNN-LSTM-Attention in terms of accuracy, F1 score and precision. Among them, LSTM had the lowest indicators, indicating that it was not as effective as the deep learning model that combined convolution and attention mechanisms when processing signals in complex environments. Based on the network model adopted above, the classification results of the event were obtained. The specific classification results are as follows Figure 12As shown in Figure 2, the experimental results show that CNN-LSTM-Attention has better separation for the dataset. For the CNN network, it is impossible to distinguish the signal with label 0 from the signal with label 3 ( Figure 12 (a)), for the CNN-LSTM network, the two signals still have some overlap, and some extreme signals of label 3 also overlap with the signal labeled label 1 ( Figure 12 (b)), for the CNN-LSTM-Attention network, we can see that the four labels are completely separated ( Figure 12 (c)). The performance of CNN+LSTM and LSTM-Attention is relatively close, but neither can surpass CNN-LSTM-Attention. This shows that the introduction of CNN and Attention mechanisms has significant advantages in handling intrusion event recognition tasks in multi-source environments.

[0092] At the same time, the contribution of the four features to the model is also analyzed, such as Figure 11 This figure shows the impact of different feature removals on the performance of the CNN-LSTM-Attention model, analyzing the impact of feature removal on model accuracy. Removing the peak feature has the least impact, maintaining accuracy at 98.04%, with an F1 score of 0.9802, and precision and recall of 0.9817 and 0.9803, respectively. This indicates that removing the peak feature has minimal impact on model performance. Removing the short-term standard deviation feature slightly reduces accuracy to 97.39%, with an F1 score of 0.9736. This decrease in precision and recall indicates that the short-term standard deviation has a certain impact on model performance. After removing skewness, accuracy drops to 94.12%, with an F1 score of 0.9413, and precision and recall also decrease, demonstrating that skewness significantly impacts model performance. Finally, removing the center frequency has the greatest impact, with the accuracy dropping to 92.16% and the F1 score being 0.9220. The precision and recall rates have dropped significantly, indicating that the center frequency feature contributes the most to the model, and removing this feature has a significant impact on model performance.

[0093] In order to solve the problem that the time-varying and spatial locality of the signal make these methods unable to process and classify efficiently, the present invention combines a variety of advanced signal processing technologies to pre-process the signal and extract the eigenvector. After the signal is denoised by a 5Hz high-pass filter, the signal is analyzed at multiple scales using wavelet denoising technology to effectively separate the noise and useful signal in different frequency bands, further removing the high-frequency noise while retaining the local characteristics of the signal. Variational mode decomposition is a method suitable for non-stationary signal analysis. It introduces an adaptive mechanism to decompose the signal based on the local signal characteristics. The signal is gradually decomposed into a series of intrinsic mode functions. The eigenvectors are extracted based on the characteristics of different intrinsic mode functions. This method can better capture the essential characteristics of the signal, especially when processing complex signals, and can suppress interference while ensuring information retention.

[0094] The present invention is applied to intrusion event monitoring in complex environments such as islands. It combines optical fiber distributed sensors to obtain vibration signals, pre-processes the signals through a complete set of signal processing methods, then performs scale decomposition to extract feature vectors, and uses deep learning models for data processing and event classification, thereby providing an efficient and real-time intrusion warning system.

[0095] With the increase in data volume and complexity, traditional single models are often unable to cope with the challenges of modern signal processing and pattern recognition. To solve this problem, hybrid models that combine multiple deep learning technologies have been widely used. In fiber optic distributed sensing systems, signals contain both spatial channel information and time channel information. In order to process fiber optic signals in a targeted manner, the present invention selects a deep learning network architecture that combines CNN, LSTM and self-attention mechanism to fully explore the key features in the signal for the problem of multi-source environmental event recognition and achieve accurate recognition of complex environmental events. The present invention uses CNN to extract local spatial features from sensor signals and identify the change patterns of signals at different locations. LSTM is responsible for processing time series data in the time channel. It can model the dependency of signals over time and capture the dynamic features in the signals. In order to further improve model performance, the self-attention mechanism dynamically adjusts the degree of attention to different time steps or features, so that the model can focus on the most critical information in the task, thereby enhancing the model's ability to parse complex data. The CNN-LSTM-Attention mechanism adopted in the present invention can simultaneously process the spatial and temporal dimensions of optical fiber sensing signals, and significantly improves the recognition and prediction capabilities of environmental events through efficient feature extraction, time series modeling, and dynamic information selection.

[0096] It should be noted that, in this document, relational terms such as first and second, etc., are used solely to distinguish one entity or operation from another, and do not necessarily require or imply any actual relationship or order between these entities or operations. Furthermore, the terms "include," "comprise," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or terminal device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or terminal device. Without further limitation, elements defined by the phrase "include..." or "comprising..." do not exclude the presence of additional elements in the process, method, article, or terminal device comprising the elements. Furthermore, in this document, "greater than," "less than," "exceeding," etc., are understood to exclude the number itself; "above," "below," "within," etc., are understood to include the number itself.

[0097] Although the above embodiments have been described, those skilled in the art may make additional changes and modifications to these embodiments once they know the basic creative concepts. Therefore, the above descriptions are merely embodiments of the present invention and do not limit the scope of patent protection of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention's description and drawings, or directly or indirectly applied in other related technical fields, are also included in the scope of patent protection of the present invention.

Claims

1. A distributed optical vibration sensing early warning verification method for an island multi-source environment is characterized by: The following steps are involved: Step 1: Collect signal data. Fiber optic sensors deployed on the island acquire real-time vibration signals from different areas of the island. Trampling signals from three different island environments—the beach, underwater, and cobblestone paths—are collected and labeled as corresponding signals to create a training dataset. Step 2: Preprocess the collected signal data, including: using a 5Hz high-pass filter to remove low-frequency noise and using wavelet transform to remove high-frequency noise; Step 3: Variational mode decomposition is used for feature extraction, and VMD is used to decompose the signal and extract useful features; Step 4: Input the preprocessed signal data into a deep learning model for event classification. The deep learning model combines CNN with LSTM and Attention mechanisms to build a CNN-LSTM-Attention network. In step 5, the CNN-LSTM-Attention network first extracts the spatial features of the signal through CNN to capture local temporal-spatial patterns. Then, the LSTM is used to process the temporal dependencies of the signal and capture timing information. The Attention mechanism then weights the features output by the LSTM to highlight the most critical time steps or features. Finally, the fully connected layer maps the weighted features to event categories, completing the accurate classification of intrusion events.

2. The distributed optical vibration sensing early warning verification method for an island multi-source environment according to claim 1, characterized in that: The step 1 specifically includes: a narrow-linewidth laser source generates a highly coherent continuous optical signal, which is modulated by an AOM and converted into a pulsed optical signal; the pulsed light is then amplified by an EDFA to enhance the signal strength, and the amplified pulse signal is further processed by a waveform amplifier to ensure signal quality; the signal is then transmitted to the optical fiber sensing part through a circulator, and wavelength division multiplexing is performed through WDM to ensure the effective transmission and collection of signals of different wavelengths; the collected backscattered light signal is transmitted again through the circulator, received by a balanced photodetector, and converted into an electrical signal; the electrical signal is sent to a data acquisition card.

3. The distributed optical vibration sensing early warning verification method for an island multi-source environment according to claim 1, characterized in that: The high-frequency noise removal using wavelet transform in step 2 specifically includes three steps: 1) decomposition of the noisy signal, 2) threshold processing and 3) signal reconstruction; the decomposition of the noise signal depends on the wavelet transform, specifically, the noisy signal g(t) and the wavelet transform signal W f The relationship between (p,q) is: Where φ(t) is the mother wavelet, p is the scaling factor, and q is the translation factor. represents the set of real numbers, and its reconstruction formula is: in C φ is a constant, W h are the continuous wavelet transform coefficients.

4. The distributed optical vibration sensing early warning verification method for an island multi-source environment according to claim 3, characterized in that: The wavelet threshold method used in the wavelet denoising is as follows: where w p,q is the wavelet coefficient before denoising, is the wavelet coefficient after denoising, the threshold μ=θ 2 log(M), where M is the signal length and θ is an estimate of the noise level.

5. The distributed optical vibration sensing early warning verification method for an island multi-source environment according to claim 1, characterized in that: Step 3 uses variational mode decomposition to perform feature extraction: The signal is divided into different modal functions, and two modal functions with higher correlation coefficients are taken. After selecting the IMF to be used, it is necessary to select appropriate characteristic parameters to extract the eigenvector. The selected characteristic parameters mainly include center frequency CF, kurtosis K, peak value PV, and short-term standard deviation STSD. First, CF is used to describe the main frequency position of a signal component, that is, the average frequency of energy in the signal spectrum; K is a statistic that describes the signal and is used to measure the steepness of the tail of the signal distribution. It reflects the "sharpness" of the data distribution compared to the normal distribution. Specifically, the kurtosis of the obtained IMF can be expressed by the following formula: where K i represents kurtosis, IMF i represents the i-th IMF signal, μ i is the mean of the ith IMF, E represents the expected value, and the peak value is used to represent the maximum absolute value s in the signal, which can reflect the extreme strength of the signal in order to enhance the signal discrimination; STSD can be used to describe the statistics of the degree of signal fluctuation in a short time window and analyze the local dynamic characteristics of the signal. When calculating STSD, the signal is first divided into t different time windows. For each event window, the formula is as follows: Where STSD(t) is the short-term standard deviation of the t-th time window, x i Represents the signal value in the window, μ is the mean of the window, N represents the number of samples, and the four characteristic parameters selected above are obtained for the two retained IMFs. They are combined into the final eigenvector, and the eigenvector can be expressed as: feature vector=[(CF1,K1,PV1,STSD1),(CF2,K2,PV2,STSD2)] (6) Among them CF 1,2 ,K 1,2 ,PV 1,2 ,STSD 1,2 Represent the central frequency, kurtosis, peak value, and short-term standard deviation of the two retained IMFs; By extracting their central frequency, kurtosis, peak value, short-term standard deviation and other classic features, the feature vector is extracted.

6. The distributed optical vibration sensing early warning verification method for an island multi-source environment according to claim 1, characterized in that: The CNN-LSTM-Attention network architecture is as follows: S4.1 Input signal data shape is (1,8,473). The first convolution layer uses a 3x1 convolution kernel and outputs a feature map of 16 channels. The second convolution layer uses a 3x1 convolution kernel and outputs a feature map of 32 channels. The third convolution layer uses a 3x1 convolution kernel and outputs a feature map of 64 channels. Each convolution layer is followed by a ReLU activation function and a Dropout layer. The S4.2 LSTM layer processes the features extracted by the CNN. The CNN output data is flattened and fed into the LSTM layer. The time step represents each spatial channel, and the LSTM layer is used to capture the features of each spatial channel. The LSTM layer adopts a bidirectional structure with 64 hidden units to capture the temporal dependencies of the signal. Event classification is finally performed through the fully connected layer. S4.3 Attention: After the output of the LSTM layer, a self-attention mechanism is introduced to further improve model performance. The Attention mechanism calculates the attention weight of each time step, allowing the model to automatically focus on the most important time steps or spatial features. Specifically, the output of the LSTM layer passes through a fully connected layer to obtain an attention score for each time step, indicating the importance of the time step in event classification. Subsequently, these scores are normalized using the Softmax function to ensure that the sum of the attention weights of all time steps is 1. The output of the S4.4 Attention part is classified through a fully connected layer with 128 neurons. The fully connected layer outputs four categories, representing different types of intrusion events. The classification results are used through an activation function to obtain the probability of each category, and finally output the most likely event type.

7. The distributed optical vibration sensing early warning verification method for an island multi-source environment according to claim 6, characterized in that: The step S4.1, for the CNN network part: the CNN part uses a total of three convolutional layers. The first convolutional layer conv1 receives input data and applies a 3x1 convolution kernel with a step size of 1, outputting a feature map of 16 channels. Each convolution kernel slides in the time dimension to extract local spatial and temporal features. Because the size of the convolution kernel is 3x1, it can capture the dependencies between adjacent time steps. Each convolution operation produces a new feature map, which contains the spatial pattern of each time step; the second convolutional layer conv2 receives the output from the first convolutional layer, continues to use a 3x1 convolution kernel with a step size of 1, and outputs a feature map of 32 channels. Here, the convolution kernel continues to slide in the time dimension to further learn the detailed features of the signal. In this way, the second layer of convolution can be used from a more abstract The spatial features of the signal are extracted at different levels, so that the model can capture more complex patterns; the third convolutional layer conv3 uses a 3x1 convolution kernel with a step size of 1 and outputs a feature map of 64 channels, which further enhances the feature extraction ability of the signal and can learn complex patterns in the signal from a higher level; the ReLU activation function is used after each convolutional layer to increase the nonlinear characteristics, so that the model can learn more complex features. After the convolution process, the feature map is further downsampled through the pooling layer MaxPool2d to reduce the amount of calculation while retaining important features. The data after each convolution layer passes through the Dropout layer with a dropout rate of 0.5 to prevent overfitting during training. The Dropout operation randomly discards 50% of the neurons to force the model to learn more robust features.

8. The distributed optical vibration sensing early warning verification method for an island multi-source environment according to claim 6, characterized in that: S4.2, for the LSTM part: the data output from the CNN is flattened, and the LSTM layer accepts the flattened features as input. The model uses 64 hidden units and adopts a bidirectional LSTM, that is, the past and future time step information is considered at the same time, and the time step is selected as the spatial channel of the data. In this case, LSTM can capture the changing dependencies of the signal in different spatial channels and process the long-term dependencies in the data of each channel. After the input features pass through the LSTM network, the output represents the feature representation corresponding to each time step.

9. The distributed optical vibration sensing early warning verification method for an island multi-source environment according to claim 6, characterized in that: S4.3, for the self-attention part: The Attention mechanism is used to weight the features of the LSTM output: First, the output of the LSTM layer is a tensor with a shape of (batch_size, time_steps, hidden_size), where time_steps represents the time step and hidden_size is the number of hidden units of the LSTM; then, each time step feature of the LSTM is linearly transformed through a fully connected layer to generate a scalar representing the attention score of each time step. This score reflects the importance of the time step in the current task. Then, the attention score is normalized by the Softmax function to ensure that the sum of the attention weights of all time steps is 1; finally, the Attention mechanism multiplies the LSTM output of each time step by the corresponding attention weight to calculate the weighted time series feature.

10. The distributed optical vibration sensing early warning verification method for an island multi-source environment according to claim 6, characterized in that: S4.4, for the classification part: After passing through the fully connected layer, the weighted time series features are mapped to the final classification results. The dimension of the output layer is 4, representing 4 types of intrusion events. Through the Softmax activation function, the output layer maps the model results to the probability value of each event type. Finally, the model predicts the most likely event type based on the probability value.

Images