Multi-dimensional depth feature extraction and recognition method of DAS signal

Abstract

The application discloses a multi-dimensional depth feature extraction and recognition method of DAS signals and belongs to the technical field of distributed optical fiber sensing and application. First, distributed optical fiber sensing signals are collected, and the distributed optical fiber sensing signals are segmented according to time periods to obtain a two-dimensional time-space sample signal matrix; a time signal of each space point is converted into a time-frequency spectrum diagram by using a short-time Fourier transform to obtain a three-dimensional time-frequency-space sample; a sample data set is constructed by using the obtained time-frequency-space signal; the sample data set is classified to obtain a training set, a verification set and a test set; a basic three-dimensional CNN network model matched with the DAS time-frequency-space sample input and an optimal attention module are constructed, and the structure is optimized to form a 3D-ACNN network model; the model is trained, verified and tested to complete high-precision and rapid recognition of the DAS signals. The end-to-end network is used to automatically extract depth features from three dimensions of time, frequency and space simultaneously by using the attention mechanism, so that the recognition performance is improved, and the calculation efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of distributed optical fiber sensing and its applications, and more specifically to the field of multidimensional depth feature extraction and recognition methods for DAS signals. Background Technology

[0002] Distributed optical fiber acoustic / vibration sensing (DAS) systems, based on the phase-sensitive optical time-domain reflectometry (Φ-OTDR) principle, utilize widely deployed optical fiber cables to provide a high-density, low-cost sensing network over a large area. In recent years, DAS systems have been widely applied in fields such as seismic wave detection in the oil and gas industry, perimeter security intrusion detection, urban communication cables, and traffic monitoring, forming a new type of fiber optic Internet of Things (IoT) for ground-based monitoring applications.

[0003] The current state of intelligent processing of DAS signals and its research can be broadly divided into two stages: the first stage (before 2017) saw the initial development of numerous traditional but useful temporal feature extraction and classifier design works; the second stage, with the rapid development of artificial intelligence (AI) technology, has seen a surge in the application of deep learning techniques to DAS to achieve more efficient detection and recognition. These include one-dimensional (1D), two-dimensional (2D) convolutional neural networks (CNNs), probabilistic neural networks (PNNs), You-Only-Look-Once (YOLO), dual-path networks (DPNs), attention-based long short-term memory (ALSTM), multi-scale CNNs combined with hidden Markov models (mCNN-HMM), 1D-CNNs combined with bidirectional LSTMs (1D-CNN-Bi-LSTM), deep belief networks, and transfer learning methods. These methods have improved the convenience of automatically extracting hidden features of signal targets and yielded better recognition results. However, most current techniques still focus on updating feature learning tools without considering the richness and comprehensiveness of information acquisition from different dimensions. From the perspective of DAS information extraction and utilization, most existing methods only focus on time features, time-frequency features (TF), spatiotemporal features (ST), or spatial-frequency features (SF), and still cannot simultaneously and automatically utilize the most comprehensive information: namely, the three dimensions of time, frequency, and space. Extracting information from all possible dimensions to establish a multidimensional DAS signal recognition method remains the most promising approach to improve the fast and accurate recognition performance of DAS. Summary of the Invention

[0004] The purpose of this invention is to provide a DAS multidimensional collaborative recognition method based on three-dimensional attention CNN. This method utilizes an end-to-end approach to synchronously and automatically extract time, frequency, and spatial multidimensional depth features of DAS signals. It also combines a three-dimensional attention mechanism to fully utilize the extracted effective information and obtain good recognition results. At the same time, the end-to-end network improves the algorithm's execution efficiency and timeliness.

[0005] To achieve the above objectives, the present invention specifically adopts the following technical solution:

[0006] A method for multidimensional deep feature extraction and recognition of DAS signals, including the following steps:

[0007] Step 1: DAS data acquisition and construction of a two-dimensional spatiotemporal sample database.

[0008] Distributed fiber optic sensing signals along pipelines in multiple scenarios are collected to obtain raw spatiotemporal signals. The collected raw spatiotemporal signals are then divided into matrix blocks according to the spatiotemporal width of the signal samples to obtain two-dimensional spatiotemporal signal samples. This allows the construction of a two-dimensional spatiotemporal sample database for different types of events.

[0009] Step 2: 3D input preparation and signal preprocessing.

[0010] The time signal of each spatial point in the two-dimensional spatiotemporal signal sample is converted into a time-frequency image, and the time-frequency image is preprocessed to construct three-dimensional time-frequency spatial input samples for different types of events. The above conversion is performed on all two-dimensional spatiotemporal signal samples in the database to obtain a dataset of three-dimensional time-frequency spatial input samples. The dataset is divided into training set, validation set and test set in a ratio of 6:2:2.

[0011] Step 3: Construct and optimize a three-dimensional attention convolutional neural network model to perform synchronous deep learning and attention on the time-frequency-space three-dimensional information of typical event signals of different types;

[0012] First, we construct a basic 3D convolutional neural network model, abbreviated as 3D-CNN.

[0013] Then, two-dimensional attention modules and three-dimensional attention modules adapted to three-dimensional convolution are added to the 3D-CNN model respectively to obtain 3D-CNN+2D-CBAM and 3D-CNN+3D-CBAM. The signals of the training set in step 2 are used as input to train the constructed 3D-CNN+2D-CBAM and 3D-CNN+3D-CBAM offline. The performance of each network is optimized through training iteration, adjustment and verification. The two attention mechanisms are compared and the model with the best performance is taken as the three-dimensional attention convolutional neural network model, abbreviated as 3D-ACNN.

[0014] Step 4: Online testing and performance evaluation.

[0015] As a preferred technical solution, the system hardware used for DAS data acquisition in step 1 is a distributed fiber optic sound and vibration sensing system based on phase demodulation phase-sensitive optical time-domain reflectometry technology.

[0016] The system hardware consists of three parts: a detection optical cable, an optical signal demodulation device, and a signal processing host.

[0017] The system returns the original signal trajectory distributed along the space at each moment. By continuously accumulating N original signal trajectories on the time axis, a spatiotemporal signal matrix with T dimensions in time and S dimensions in space is constructed.

[0018] {XX=x ts (t=1,2,…,N; s=1,2,…,M)} (1)

[0019] Where: XX represents the spacetime matrix, x ts The matrix represents a point, where t represents a specific moment, and s represents a specific spatial point; N is the time sampling length, and M is the spatial sampling length, both in units of sampling points. The spatiotemporal response signal of the distributed fiber optic sound and vibration sensing system is accumulated over time. The horizontal axis represents the spatial points along the pipeline, with the distance between two spatial points being ΔS; the vertical axis represents the time axis, with the sampling interval between two sampling points being ΔT = 1 / f. s f s This refers to the pulse trigger frequency on the time axis, which is also the time sampling frequency.

[0020] The collected spatiotemporal matrix is ​​divided into spatiotemporal regions to construct a two-dimensional spatiotemporal signal sample for each data point in the dataset.

[0021] As a preferred technical solution, in step 2, the short-time Fourier transform is used to convert the time signal of each spatial point in the two-dimensional spatiotemporal signal sample into a time-frequency diagram. Let the signal be f(t), the required window function be h(t), and the short-time Fourier transform G of f(t) be... f (w, t) is defined as:

[0022] G f (w,t)=∫f(τ)h(τ-t)e jwτ dτ (2)

[0023] Among them: G f (w, t) represents the short-time Fourier transform operation, w represents the calculated spectrum, t represents a certain moment, and h(τ-t) represents the local time shift t of the window function. The position of the local time window is located using t. In actual calculations, τ is a discrete time point, and the determined short-time signal is:

[0024] S(τ)=f(τ)h(τ-t) (3)

[0025] Then, a fast Fourier transform is performed on S(τ) to calculate its frequency components, thereby reducing the computational load. In order to ensure the resolution of the spectrum in STFT, a rectangular window with a data length of 95 is selected, the number of jumps is one sampling point, and the FFT size is equal to the window length.

[0026] Thus, an original spatiotemporal sample can be converted into a three-dimensional sample with a data size of 129×17×100, representing a time length of 2s, a frequency range of 250Hz, and a spatial range of 129m for each sample.

[0027] The vibration event signals in the typical event database include: background noise, traffic interference, manual excavation, and mechanical excavation.

[0028] As a preferred technical solution, the 3D-ACNN in step 3 includes network architecture design, network structure and parameter tuning, cost function design, and network update process.

[0029] As a preferred technical solution, the network architecture design includes the design and construction of a basic 3D-CNN model, a two-dimensional attention module, and a three-dimensional attention module.

[0030] As a preferred technical solution, a basic 3D-CNN network model is first constructed. This model consists of an input layer, a convolutional layer C1, a pooling layer P1, a residual block 1, a convolutional layer C2, a pooling layer P2, a residual block 2, a fully connected layer FC1, a fully connected layer FC2, and an output layer. In the convolutional block, the three-dimensional convolutional kernel encodes the three-dimensional input and simultaneously extracts the TFS information of the DAS signal from the three directions (x, y, z). In each residual block, there are two convolutional layers and a nonlinear activation function ReLU layer. The specific calculation operations are shown in equations (4) and (5).

[0031] y l =h(x l )+r(x l W l (4)

[0032] x l+1 =f(y l (5)

[0033] Where: x l and x l+1 This represents the input and output of the l-th (l = 1, 2) residual block; y l This is an intermediate result; r(x) l W l h(x) is the residual function, representing the learned residual term; l h(x) is the identity mapping of the input, i.e., h(x) l )=x l ;f(y l The residual block is the ReLU function. Through skip connections and identity mapping, the residual block speeds up the convergence, enhances the nonlinearity, and avoids network degradation.

[0034] Then, an attention mechanism is added after the convolutional output of each residual block in 3D-CNN. The entire network model is referred to as 3D-ACNN, and its network structure is as follows: Input layer - Convolutional layer C1 - Pooling layer P1 - Residual block + 3D-CBAM1 - Convolutional layer C2 - Pooling layer P2 - Residual block + 3D-CBAM2 - Fully connected layer FC1 - Fully connected layer FC2 - Output.

[0035] As a preferred technical solution, the design and construction process of the two-dimensional attention module is as follows:

[0036] An attention module is used to select key information extracted from the 3D deep learning network. The attention mechanism employs a soft attention mechanism, where the input feature map is multiplied by an attention mask, calculated as follows:

[0037] f′ c (x i )=A(x i )·f c (x i (6)

[0038] Where: f c (x i A(x) is the feature map at the output of the previous feature extractor. i ) is the attention mask, f′ c (x i ) is the output of the attention module;

[0039] The specific implementation of the two-dimensional attention module is as follows: given an intermediate feature map f∈R T×F×S×C Subsequently, the 2D-CBAM module derives the attention mapping sequentially along the two dimensions of time-frequency structure and channel. Let the attention mapping of the TF structure be M. TF (f)∈R T×F Spatial splicing is M con (f)∈R T×F×S Channel attention mapping is M C (f)∈R C×1×1×1 In each channel, its time-frequency slice f ij Multiplied by the local two-dimensional attention mask M in the TF structure TF (f) Obtain the attention feature map for each spatial point; all TF feature maps along the fiber space are stitched together to form a three-dimensional feature M. con (f); then pay attention to channel M C (f) Apply to all channels;

[0040] The final feature refinement output f″′∈R T×F×S×C The calculation process is as follows:

[0041]

[0042] f ij " = M con (f ij ′) (8)

[0043]

[0044] M TF (f)=σ(f 7×7 ([AvgPool(f);MaxPool(f)])) (10)

[0046] M C (f)=σ(MLP(AvgPool(f))+MLP(MaxPool(f))) (11)

[0047] Where: f ij f represents the time-frequency slice of each channel. ij ′、f ij "and f ij "″" and "′" are the output feature maps obtained by local TF structure attention, spatial splicing, and channel attention, respectively. Represents element-wise multiplication, ReLU is the linear correction unit function, σ represents the sigmoid function, and f 7×7 It is a convolution operation with a 2D filter of kernel size 7×7. The semicolon ";" indicates the concatenation of set vectors. MLP stands for Multilayer Perceptron, which is similar to a fully connected layer and is used to reduce the dimensionality of the input. AvgPool and MaxPool represent average pooling and max pooling, respectively.

[0048] As a preferred technical solution, the design and construction process of the three-dimensional attention module is as follows:

[0049] Given an intermediate feature map f∈R T×F×S×C Subsequently, the 3D-CBAM module derives the attention mapping sequentially along the three dimensions of time-frequency structure, space, and channel. Let the attention mapping of the TF structure be M. TF (f)∈R T×F Spatial attention mapping is M S (f)∈R T×F×S Channel attention mapping is M C (f)∈R C×1×1×1 The final feature refinement output f″′∈R T×F×S×C The calculation process is as follows:

[0050]

[0051]

[0052]

[0053] M TF (f) = ReLU(f) 3×3×3 (f)) (15)

[0054] M S (f)=σ(f) (16)

[0055] M C (f)=f 3×3×3 (f) (17)

[0056] Where: f ij f represents the time-frequency slice of each channel. ij ′、f ij "and f ij "″′ are the output feature maps obtained by local TF structural attention, spatial attention, and channel attention, respectively. Represents element-wise multiplication, ReLU is the linear correction unit function, σ represents the sigmoid function, and f 3×3×3 It is a convolution operation with a three-dimensional filter of kernel size 3×3×3.

[0057] As a preferred technical solution, the cost function design and network update process includes: after the 3D-ACNN is initialized with parameters, the data in the training set is input into it, the predicted class probability distribution is obtained through forward propagation of the network, and the loss value between the class probability output by the fully connected layer and the true probability is calculated through the cross-information entropy loss function. The gradient of each learning parameter is calculated by backpropagation using the loss value, and the model parameters θ are updated according to the gradient descent method with a specified learning rate. The model parameters θ include matrix weights W and biases b.

[0058] The Xavier initialization method is used for parameter initialization. To ensure consistent variance across layers during forward and backward propagation, the random parameter initialization distribution ranges from the number of input parameters n passing through that layer. in Number of output parameters n out The formula for calculating the distribution range U of the obtained uniform distribution is:

[0059]

[0060] The time-frequency spatial feature sample signals in the training set are input into the 3D-ACNN network model for forward propagation to obtain the predicted labels of the time-frequency spatial signal samples;

[0061] The cost function chosen is the cross-information entropy loss function, which is used to calculate the loss value between the predicted label and the true label: the distance between the predicted event label and the true label is calculated based on the cross-information entropy loss function to obtain the loss value L, and the calculation formula is as follows:

[0062]

[0063] Where: n, a, and y represent the total number of samples, the predicted label of the sample, and the true label of the sample, respectively;

[0064] The parameter gradients of the 3D-ACNN network model are calculated in reverse using the loss value, and the 3D-ACNN deep learning model is updated using the parameter gradients.

[0065] The Adam algorithm is used to optimize the network. The calculation steps are as follows:

[0066] g t =▽ θ L t (θ t-1 (20)

[0067] m t =u*m t-1 +(1-u)*g t (twenty one)

[0068]

[0069] Wherein: g t For calculating the target gradient, m t n t , which are the first and second moment estimates of the gradient, u,v∈[0,1), and , which are the exponential decay rates of the first and second moments of the gradient, respectively;

[0070] Let the t-th iteration be... The network parameter update formula is:

[0071]

[0072] Where: α is the learning rate, and ε is a very small number to prevent division by zero;

[0073] After updating the 3D-ACNN network model with model parameters θ, the training loss value is used to determine whether the 3D-ACNN network model has converged. If it has converged, the training process ends; otherwise, the training is repeated until the maximum number of iterations is reached. After multiple iterations, the model is considered to have converged when the loss function value is less than a certain threshold or the iteration exceeds a certain threshold, and the above iteration process is stopped.

[0074] After iterative convergence, the model is validated using a validation set, and the model with the best results is saved as the final event recognition model.

[0075] The beneficial effects of this invention are as follows:

[0076] 1. This study is the first to introduce 3D CNN into DAS signal recognition, enabling the simultaneous and automatic extraction of comprehensive information from three dimensions: time, frequency, and space. In the designed 3D CNN model, a residual block structure is used to accelerate the convergence speed of the 3D network and improve the model's robustness.

[0077] 2. To quickly focus on the most useful information across different dimensions and effectively utilize multidimensional features, a comparative design and study of two-dimensional and three-dimensional attention mechanisms were conducted. Results show that attention-assisted 3D-CNN outperforms basic 3D-CNN, and the 3D attention module in the network outperforms the 2D attention module. Furthermore, it is shown that the information contribution of different dimensions is different. Compared to TS and SF dimensions, TF dimensions contain richer identifiable information.

[0078] 3. Simultaneously, the network consciously incorporates lightweight structures, such as skip connections in residual blocks, the number of residual blocks, the location of attention modules, and global average pooling (GAP) in the output layer. These structures reduce network redundancy and mitigate the increasing computational complexity as the model size grows. Furthermore, the input sample size of the first convolutional layer is carefully chosen to control computational costs.

[0079] 4. Finally, in the field monitoring of underground communication cables in two Chinese cities, the proposed 3D-ACNN outperformed three other commonly used 2D networks in recognizing four event types, increasing the accuracy from 95.47% to 99.33%. Furthermore, the computational complexity was significantly reduced; the testing time for a single spatial sensing point was only about 1 / 171 of that of a TF-input 2D network. This end-to-end network not only improved recognition accuracy but also enhanced algorithm execution efficiency and timeliness. Attached Figure Description

[0080] Figure 1 This is a typical setup for the DAS system based on a phase-sensitive optical time-domain reflectometer (Ф-OTDR) according to the present invention;

[0081] Figure 2 This is a flowchart of the multidimensional information extraction and collaborative recognition algorithm based on 3D-ACNN configuration proposed in this invention;

[0082] Figure 3 The four typical signals obtained on-site in this invention are: (a) background noise; (b) traffic interference; (c) manual excavation; and (d) mechanical excavation.

[0083] Figure 4 This is a detailed structural diagram of the 2D-CBAM customized in 3D-CNN according to the present invention;

[0084] Figure 5 This is a detailed structural diagram of the 3D-CBAM designed in 3D-CNN according to the present invention;

[0085] Figure 6 For the training and validation of network models with different numbers of residual blocks during the optimization process of this invention: (a) training time; (b) validation accuracy;

[0086] Figure 7 To facilitate the optimization process of this invention, attention positions were compared in three different cases: (a) training time; (b) verification accuracy.

[0087] Figure 8 This invention verifies the accuracy of different reduction numbers in CBAM during the optimization process.

[0088] Figure 9 This is the optimal 3D attention CNN network structure diagram in this invention;

[0089] Figure 10 These are two-dimensional and three-dimensional attention diagrams of the four event signals in this invention;

[0090] Figure 11 The confusion matrix of the six models in the testing phase of this invention is shown in (a) for TS-2D-CNN, (b) for TF-2D-CNN, (c) for SF-DPN, (d) for basic 3D-CNN, (e) for 3D-CNN assisted by 2D CBAM, and (f) for 3D-ACNN.

[0091] Figure 12 This is a bar chart showing the time cost and model size of the six models during the testing phase of this invention. Detailed Implementation

[0092] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0093] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0094] Example 1

[0095] Taking a security monitoring application based on optical fiber communication cables as an example, see [link / reference]. Figure 1 The typical setup of a DAS system based on a phase-sensitive optical time domain reflectometer (Ф-OTDR) provided in this embodiment requires first connecting the DAS system to the communication optical cable in the monitoring area, and collecting spatiotemporal signals of different types of events occurring on the ground, i.e., distributed optical fiber sensing signals.

[0096] A DAS multidimensional information extraction and collaborative recognition method based on 3D attention CNN (3D-ACNN) is described, with the entire signal processing flow as follows: Figure 2 As shown, it mainly consists of four steps:

[0097] Step 1: DAS Data Acquisition and Construction of a Two-Dimensional Spatiotemporal Sample Database. Using a distributed optical fiber acoustic / vibration sensing system (DAS) hardware based on a phase-sensitive optical time-domain reflectometer, sound or vibration signals (i.e., distributed optical fiber sensing signals) along pipelines in complex background environments across multiple scenarios are acquired, resulting in the original spatiotemporal signal matrix. The acquired spatiotemporal signals are then divided into matrix blocks according to an appropriate signal sample spatiotemporal width to obtain two-dimensional spatiotemporal signal samples, thereby constructing a database of different types of events.

[0098] Step 2: 3D Input Preparation and Signal Preprocessing. Perform a Short-Time Fourier Transform (STFT) on the time signal of each spatial point in the 2D spatiotemporal sample to obtain its time-frequency image. Preprocess this time-frequency image, including cropping and grayscale conversion, to construct 3D spatiotemporal input samples for different types of events. Perform the above transformations on all 2D spatiotemporal samples in the database to obtain the 3D spatiotemporal input sample set. Randomly divide the dataset into training, validation, and test sets in a 6:2:2 ratio.

[0099] Step 3: Construct and optimize a 3D attention convolutional neural network (3D-ACNN) to perform synchronous deep learning and attention on the spatiotemporal 3D information of typical event signals of different types. First, a basic 3D convolutional neural network model, abbreviated as 3D-CNN, is constructed. Its network structure is: input layer - convolutional layer C1 - pooling layer P1 - residual block - convolutional layer C2 - pooling layer P2 - residual block - fully connected layer FC1 - fully connected layer FC2 - output layer. Then, a 2D attention module (2D-CBAM) adapted to 3D convolution and a customized 3D attention module (3D-CBAM) are designed in 3D-CNN. Subsequently, the 2D-CBAM and the customized 3D-CBAM designed in this invention are added to the basic 3D-CNN model respectively, and the two attention mechanisms are compared to construct a 3D attention convolutional neural network (3D-ACNN). The signals in the training set obtained in Step 2 are used as input to train the constructed networks offline. Through training iterations, adjustments, and validation optimizations, the performance of each network is optimized.

[0100] Step 4: Online Testing and Performance Evaluation. The test set obtained in Step 2 is input into the optimal 3D-ACNN model to identify the time-frequency-space dataset of the test events. Common 2D networks with only time-frequency (TF), time-space (TS), or space-frequency (SF) inputs are compared with the proposed basic 3D-CNN, 3D-CNN+2D-CBAM, and 3D-ACNN, including confusion matrix, precision, recall, F1-score, and average accuracy. The performance differences brought by different attention modules are also compared.

[0101] Example 2

[0102] This embodiment is based on Embodiment 1, and further optimizes the present invention.

[0103] The system hardware used for signal acquisition in Example 1 is a distributed fiber optic sound and vibration sensing system based on phase-modulated phase-sensitive optical time-domain reflectometry (Φ-OTDR) technology. The system structure and its working principle are as follows: Figure 1As shown, the system hardware consists of three parts: a detection optical cable, an optical signal demodulation device, and a signal processing host. The detection optical cable typically uses ordinary single-mode communication optical fiber, usually laid underground along pipelines, power transmission cables, and urban roads. It can also directly utilize spare fiber cores from communication optical cables laid along pipelines or roads. The optical signal demodulation device is the core of the system, and its internal components mainly include optical and electrical components. A continuous coherent optical signal is generated by an ultra-narrow linewidth laser. This signal is modulated into an optical pulse signal by an acousto-optic or electro-optic modulator. The optical pulse signal is then amplified by an erbium-doped fiber amplifier (EDFA). The amplified optical pulse signal is injected into the detection optical cable sequentially through an isolator and ports 1 and 2 of a circulator. Rayleigh scattering occurs as the optical pulse signal travels along the optical cable. The Rayleigh-scattered optical signal returns along the optical cable and is received by ports 2 and 3 of the circulator. After noise is filtered out by an optical filter, the signal is coupled by a first coupler and then injected into an unbalanced Mach-Zehnder or Michelson interferometer, the specific method of which is determined by the demodulation method. The phase change information introduced by external disturbances with a phase difference of 120 degrees is output through a 3*3 second coupler. This allows the acquisition of the signal of sound waves and vibrations on the optical fiber. The demodulated optical signal is converted into an electrical signal by a photodetector. The signal is then synchronously acquired by a synchronously triggered analog-to-digital converter controlled by a waveform generator. Finally, the digital electrical signal is transmitted to the signal processing host in real time through a network interface. The signal processing host is a regular computer host (PC) or an FPGA / DSP embedded motherboard. It is used for the analysis and processing of fiber optic detection signals. It obtains event information that causes sound waves, vibrations, etc. through specific signal processing algorithms, determines the location of the event by the principle of optical time domain reflection, and performs intelligent analysis, processing, identification and classification of the sensed events.

[0104] The system returns the original signal trajectory distributed along the space at each moment. By continuously accumulating N original signal trajectories on the time axis, a spatiotemporal signal matrix with T dimensions in time and S dimensions in space is constructed.

[0105] {XX=x ts (t=1,2,…,N; s=1,2,…,M)} (1)

[0106] Where: XX represents the spacetime matrix, x ts The matrix represents a point, where t represents a specific moment, and s represents a specific spatial point; N is the time sampling length, and M is the spatial sampling length, both in units of sampling points. The spatiotemporal response signal of the distributed fiber optic sound and vibration sensing system is accumulated over time. The horizontal axis represents the spatial points along the pipeline, with the distance between two spatial points being ΔS; the vertical axis represents the time axis, with the sampling interval between two sampling points being ΔT = 1 / f. s f sThis refers to the pulse trigger frequency on the time axis, i.e., the time sampling frequency. The acquired spatiotemporal matrix is ​​divided into spatiotemporal regions to construct each spatiotemporal signal sample in the dataset.

[0107] Example 3

[0108] This embodiment is based on Embodiment 2, and further optimizes the present invention.

[0109] By converting the time signal of a spatial point in each spatiotemporal sample in the dataset of Example 2 into a time-frequency graph, a three-dimensional time-frequency spatial input sample can be constructed. There are many ways to convert time signals into time-frequency information, such as Short-Time Fourier Transform (STFT), Mel Frequency Cepstral Coefficients (MFCC), and Gram Angular Field (GAF). This invention converts the time signals of each spatial point in the spatiotemporal graph into a time-frequency graph using Short-Time Fourier Transform (STFT). To determine the frequency domain components of a non-stationary signal within a certain time window, the time window needs to be localized to determine the frequency components within that time window. Let the signal be f(t), the required window function be h(t), and the Short-Time Fourier Transform of f(t) be defined as:

[0110] G f (w,t)=∫f(τ)h(τ-t)e jwτ dτ (2)

[0111] Among them: G f (w, t) represents the short-time Fourier transform operation, w represents the calculated spectrum, t represents a certain moment, and h(τ-t) represents the local time shift t of the window function. The position of the local time window is located using t. In actual calculations, τ is a discrete time point, and the determined short-time signal S(τ) is:

[0112] S(τ)=f(τ)h(τ-t) (3)

[0113] Then, a Fast Fourier Transform (FFT) is performed on S(τ) to calculate its frequency components, thereby reducing the computational load. To ensure the resolution of the spectrum in the STFT, a rectangular window with a data length of 95 (approximately 0.2 s) is selected, with a jump size equal to one sampling point (0.2 ms). The FFT size is equal to the window length. Thus, an original spatiotemporal sample can be converted into a three-dimensional sample with a data size of 129 × 17 × 100, representing a time length of 2 s, a frequency range of 250 Hz, and a spatial range of 129 m for each sample. The vibration event signals in the typical event database include: background noise, traffic interference, manual excavation, and mechanical excavation. The on-site photos, one-dimensional time series, and time-spectrum diagrams corresponding to each spatial point for the above events are shown below. Figure 3As shown in Table 1, the data in the typical event database is randomly divided into training, validation, and test sets in a 6:2:2 ratio.

[0114] Table 1 Typical Event Database

[0115]

[0116] Example 4

[0117] This embodiment is based on Embodiment 3, and further optimizes the present invention.

[0118] Based on the DAS time-frequency spatial feature dataset obtained in Example 3, a three-dimensional attention convolutional neural network (3D-ACNN) is constructed and optimized to perform synchronous deep learning and attention on the time-frequency spatial three-dimensional information of typical event signals of different types. Specifically, this includes three steps: 3D-ACNN network architecture design, network structure and parameter tuning, and cost function design and network update process. The specific method is as follows:

[0119] 1.3D-ACNN Network Architecture Design

[0120] This process mainly includes the following three steps:

[0121] (1) The design and construction process of 3D-CNN;

[0122] (2) Design and construction process of two-dimensional attention module (2D-CBAM);

[0123] (3) Design and construction process of 3D-CBAM.

[0124] 2. Network structure and parameter optimization

[0125] This step mainly includes the following two steps:

[0126] (1) Optimization of basic network structure: Starting from the network structure, consider the basic model of the residual network and the number of residual blocks;

[0127] (2) Optimization of attention mechanism: Finding the optimal position of attention mechanism in the model, and the optimal selection of hyperparameters in attention mechanism, etc.

[0128] 3. Cost Function Design and Network Update Process

[0129] This step mainly includes: cost function formulation, network selection optimization algorithm, and network update process.

[0130] Example 5

[0131] This embodiment is based on Embodiment 4, and further optimizes the present invention.

[0132] For the 3D-ACNN network architecture design, a basic 3D-CNN network model is first constructed, which consists of an input layer, a convolutional layer C1, a pooling layer P1, a residual block 1, a convolutional layer C2, a pooling layer P2, a residual block 2, a fully connected layer FC1, a fully connected layer FC2, and an output layer. Essentially, it involves two three-dimensional convolutional blocks, each followed by a residual block. In the convolutional blocks, the three-dimensional convolutional kernels encode the three-dimensional input, simultaneously extracting the TFS information of the DAS signal from the three directions (x, y, z). The residual blocks in the network are connected by skip connections, which accelerates the network convergence speed and greatly avoids the gradient vanishing and exploding problems. Specifically, in each residual block, there are two convolutional layers and a non-linear activation function ReLU layer. The specific calculation operations are shown in equations (4) and (5):

[0133] y l =h(x l )+r(x l W l (4)

[0134] x l+1 =f(y l (5)

[0135] Where: x l and x l+1 This represents the input and output of the l-th (l = 1, 2) residual block; y l This is an intermediate result; r(x) l W l h(x) is the residual function, representing the learned residual term; l h(x) is the identity mapping of the input, i.e., h(x) l )=x i ;f(y l () is the ReLU function. Through skip connections and identity mapping, residual blocks accelerate convergence, enhance nonlinearity, and prevent network degradation.

[0136] Then, an attention mechanism is added after the convolutional output of each residual block in the 3D-CNN. The entire network model is referred to as 3D-ACNN, and its network structure is as follows: Input layer - Convolutional layer C1 - Pooling layer P1 - Residual block + 3D-CBAM1 - Convolutional layer C2 - Pooling layer P2 - Residual block + 3D-CBAM2 - Fully connected layer FC1 - Fully connected layer FC2 - Output layer, as shown below. Figure 2 As shown.

[0137] Example 6

[0138] This embodiment is based on Embodiment 5 and further optimizes the present invention, mainly introducing the design process of the two-dimensional attention module (2D-CBAM).

[0139] After preparing the basic 3D features and the basic 3D network, an attention module is used to select key information extracted from the 3D deep learning network. The attention mechanisms involved in this invention all employ soft attention, i.e., calculating a weighted average of N input information instead of selecting just one. Soft attention mechanisms can be obtained through backpropagation and are easier to update than hard attention mechanisms. Basically, the input feature map is multiplied by the attention mask, calculated as follows:

[0140] f c ′(x)=A(x i )·f c (x i (6)

[0141] Where: f c (x i A(x) is the feature map at the output of the previous feature extractor. i ) is the attention mask, f′ c (x i The output of the attention module is shown below. This invention designs a commonly used 2D-CBAM and incorporates it into a 3D CNN. The specific structure of this attention module is as follows: Figure 4 As shown.

[0142] The specific implementation of 2D-CBAM is as follows: given an intermediate feature map f∈R T×F×S×C After representing multiple channels of three-dimensional features (time, frequency, and space), the 2D-CBAM module derives the attention mapping sequentially along the two dimensions of time-frequency structure (TF) and channels. Let the attention mapping of the TF structure be M. TF (f)∈R T×F Spatial splicing is M con (f)∈R T×F×S Channel attention mapping is M C (f)∈R C×1×1×1 In each channel, its time-frequency slice f ij Multiplied by the local two-dimensional attention mask M in the TF structure TF (f) Obtain the attention feature map for each spatial point; all TF feature maps along the fiber space are stitched together to form a three-dimensional feature M. con (f); then pay attention to channel M C (f) Apply to all channels. The final feature refinement output f″′∈R T×F×S×C The calculation process is as follows:

[0143]

[0144] f ij " = M con (f ij ′) (8)

[0145]

[0146] M TF (f)=σ(f 7×7 ([AvgPool(f);M axPool(f)])) (10)

[0147] M C (f)=σ(MLP(AvgPool(f))+MLP(MaxPool(f))) (11)

[0148] Where: f ij f represents the time-frequency slice of each channel. ij ′、f ij "and f ij "″" and "′" are the output feature maps obtained by local TF structure attention, spatial splicing, and channel attention, respectively. Represents element-wise multiplication, ReLU is the linear correction unit function, σ represents the sigmoid function, and f 7×7 It is a convolution operation with a 2D filter of kernel size 7×7. The semicolon ";" indicates the concatenation of set vectors. MLP stands for Multilayer Perceptron, which is similar to a fully connected layer and is used to reduce the dimensionality of the input. AvgPool and MaxPool represent average pooling and max pooling, respectively.

[0149] Example 7

[0150] This embodiment is based on Embodiment 6 and further optimizes the present invention, mainly introducing the design process of the three-dimensional attention module (3D-CBAM).

[0151] In the 2D attention module of Example 6, it only uses splicing operations in space instead of a specific attention mechanism. To fully capture the characteristics of DAS in the temporal, frequency, and spatial dimensions, this invention also customizes a three-dimensional attention module (3D-CBAM), the specific structure of which is as follows... Figure 5 As shown.

[0152] Given an intermediate feature map f∈R T×F×S×C After representing multiple channels of three-dimensional features (time, frequency, and space), the 3D-CBAM module sequentially derives the attention mapping along the three dimensions of time-frequency structure (TF), space, and channels. Let the attention mapping of the TF structure be M. TF (f)∈R T×FSpatial attention mapping is M S (f)∈R T×F×S Channel attention mapping is M C (f)∈R C×1×1×1 The final feature refinement output f″′∈R T×F×S×C The calculation process is as follows:

[0153]

[0154]

[0155]

[0156] M TF (f) = ReLU(f) 3×3×3 (f)) (15)

[0157] M S (f)=σ(f) (16)

[0158] M C (f)=f 3×3×3 (f) (17)

[0159] Where: f ij f represents the time-frequency slice of each channel. ij ′、f ij "and f ij "″′ are the output feature maps obtained by local TF structural attention, spatial attention, and channel attention, respectively. Represents element-wise multiplication, ReLU is the linear correction unit function, σ represents the sigmoid function, and f 3×3×3 It is a convolution operation with a three-dimensional filter of kernel size 3×3×3.

[0160] Example 8

[0161] This embodiment further optimizes the present invention based on Embodiment 4, mainly introducing the implementation and optimization process of two attention-based 3D-CNNs. The training and validation datasets in Table 1 are used to implement and optimize the network. The optimization mainly includes determining the optimal architecture of the basic 3D-CNN and the optimal position of the CBAM module.

[0162] First, for the basic residual network (ResNet) used in 3D-CNN, there are some common architectures, such as ResNet18, ResNet34, ResNet50, and ResNet101. DAS signal recognition differs from image recognition; image recognition has relatively simple data scenarios but higher real-time requirements. Therefore, this invention does not consider deep networks and selects ResNet18, which has better real-time performance, as the basic architecture in the residual block.

[0163] In DAS signal recognition, the first step is to select an appropriate number of residual blocks in the ResNet network. Then, under the same parameters, training and validation are performed with different numbers of residual blocks for comparison. Figure 6 As shown, the error bars for the mean and standard deviation (3σ) of training time and validation accuracy were obtained. The Epoch was set to 10 during training. It can be seen that as the number of residual blocks increases, both training time and validation accuracy continuously increase; however, when the number of residual blocks exceeds 2, the rate of increase in validation accuracy decreases. Therefore, in Figure 2 The model proposed uses two residual blocks.

[0164] Once the basic structure of the residual blocks is fixed, the basic ResNet model is ready. The next step is to find the optimal location for the attention module (CBAM). The basic model consists of two convolutional layers, each followed by a residual block. Within each residual block, there are two convolutional layers. The attention module is always used to enhance the deep feature extraction obtained by the convolutional network and then inserted after the convolutional layers. Therefore, three possible attention module locations were compared, including two classes: outside the residual block (Case I) and inside the residual block (Case II and Case III). Specifically, in Case I, the CBAM is added outside the residual block, after Conv1 and Conv2 respectively; in Case II, the CBAM is inserted after the first Conv of each residual block; and in Case III, the CBAM is added after the second Conv of each residual block. Figure 7 The mean and standard deviation (3σ) error bars of training time and validation accuracy for three attention networks were compared. The results show that the CBAM inserted inside the residual block performs better than the CBAM inserted outside the residual block, exhibiting higher validation accuracy and shorter training time. Case III achieved the highest validation accuracy of 98.90%, a 0.55% improvement over Case II. This suggests that the two convolutional layers within the residual block may play different roles, with the first layer potentially mining more common features and the second layer mining more specific features. In Case III, the attention module focuses more on specific features, thus performing better.

[0165] At the same time, combined Figure 6 and Figure 7It can be seen that the training time of ResNet without CBAM is 40.29ms, while the training time of ResNet with CBAM is approximately 44.22ms, indicating that the training time with CBAM is less than 4ms. This means that CBAM is a lightweight module with minimal impact on the overall network training time. Therefore, the Case III architecture was chosen.

[0166] Furthermore, the reduction number in CBAM is a crucial hyperparameter that affects the ratio of input to output channels and needs to be carefully selected experimentally. This paper tests and compares values ​​of 16, 8, 4, 2, and 1 in the validation set. Figure 8 As shown, the validation accuracies are 97.92%, 98.12%, 98.90%, 98.75%, and 98.36% when the reduction numbers are 16, 8, 4, 2, and 1, respectively. Therefore, setting the optimal reduction number to 4 is more suitable for DAS applications. Finally, the optimized 3D attention CNN configuration is as follows. Figure 9 As shown.

[0167] Example 9

[0168] This embodiment is based on embodiment 4 and further optimizes the present invention, mainly including the design of the cost function and the network update process.

[0169] After parameter initialization, 3D-ACNN inputs data from the training set. The predicted class probability distribution is obtained through forward propagation of the network. The loss value between the class probability output by the fully connected layer and the true probability is calculated using the cross-information entropy loss function. The gradient of each learning parameter is calculated using backpropagation based on the loss value. The model parameters θ are updated according to the gradient descent method with a specified learning rate. The model parameters θ include matrix weights W and biases b.

[0170] Specifically, the parameters of the 3D-ACNN network model are initialized, including matrix weights W and biases b. Initialized network parameters are the starting point for model training; good initialization parameters make the model easier to learn and converge faster. This invention uses the Xavier initialization method for parameter initialization. To ensure consistent variance in each layer during forward and backward propagation, the random initialization of parameters is distributed within the range of the number of input parameters n passing through that layer. in Number of output parameters n out The formula for calculating the distribution range U of the obtained uniform distribution is:

[0171]

[0172] The time-frequency spatial feature sample signals in the training set are input into the 3D-ACNN network model for forward propagation to obtain the predicted labels of the time-frequency spatial signal samples;

[0173] The cost function chosen is the cross-information entropy loss function, which is used to calculate the loss value between the predicted label and the true label: the distance between the predicted event label and the true label is calculated based on the cross-information entropy loss function to obtain the loss value L, and the calculation formula is as follows:

[0174]

[0175] Where n, a, and y represent the total number of samples, the predicted label of the sample, and the true label of the sample, respectively.

[0176] The loss value is used to inversely calculate the parameter gradients of the 3D-ACNN network model, and these parameter gradients are used to update the 3D-ACNN deep learning model. This invention uses the Adam algorithm to optimize the network; the calculation steps are as follows:

[0177] g t =▽ θ L t (θ t-1 (20)

[0178] m t =u*m t-1 +(1-u)*g t (twenty one)

[0179]

[0180] Wherein: g t For calculating the target gradient, m t n t , , are the first and second moment estimates of the gradient, respectively, u,v∈[0,1), and , are the exponential decay rates of the first and second moments of the gradient, respectively.

[0181] Let the t-th iteration be... The network parameter update formula is:

[0182]

[0183] Where: α is the learning rate, and ε is a very small number to prevent division by zero.

[0184] After updating the 3D-ACNN network model with model parameters θ, the training loss value is used to determine whether the 3D-ACNN network model has converged. If it has converged, the training process ends; otherwise, training is repeated until the maximum number of iterations is reached. After multiple iterations, the model is considered converged when the loss function value is less than a certain threshold or the iteration exceeds a certain threshold, and the above iteration process stops. After convergence, the model is validated using a validation set, and the model with the best result is saved as the final event recognition model. Finally, the optimal structural parameters of the proposed 3D-ACNN and 2D-CBAM-assisted 3D-CNN (3D-CNN+2D-CBAM) are shown in Tables 2 and 3, respectively.

[0185] Table 2. Structural parameters of the 3D-ACNN network

[0186]

[0187] Table 3. Structural parameters of the 3D-CNN+2D-CBAM network

[0188]

[0189]

[0190] Example 10

[0191] This embodiment is based on Embodiment 9, and further optimizes the present invention.

[0192] The optimal 3D-ACNN model obtained in Example 9 was used as the event recognition model for online testing and performance evaluation of events to be tested along the optical cable. It was compared with five well-trained and validated network models under the same conditions (same input conditions): TF-2D-CNN, TS-2D-CNN, SF-DPN, the basic 3D-CNN proposed in this invention, and 3D-CNN+2D-CBAM. The comparison included confusion matrix, precision, recall, F1-score, average accuracy, and online recognition time (Testing time per sample). The network structures of the basic 3D-CNN, 3D-CNN+2D-CBAM, and 3D-ACNN are as follows: Figure 9 As shown (will) Figure 9 Removing the attention module from the diagram gives the network structure of a 3D-CNN.

[0193] The recognition results of TF-2D-CNN, TS-2D-CNN, SF-DPN, basic 3D-CNN, 3D-CNN+2D-CBAM, and the 3D-ACNN of this invention are compared. The confusion matrices of the six models are as follows: Figure 11 As shown in the figure, the corresponding performance metrics such as precision, recall, F1-score, and average accuracy are shown in Table 4.

[0194] Table 4 Comparison of recognition results of six models

[0195]

[0196]

[0197] For the first four models without attention assistance, specifically, from Table 4 and Figure 11 It can be seen that the 3D-CNN model proposed in this invention achieves a recognition accuracy of 98.67%; TF-2D-CNN, which uses time-frequency features, has the second highest accuracy at 97.73%; TS-2D-CNN, which utilizes spatiotemporal features, and SF-DPN, which utilizes spatial-frequency features, are 2.75% and 3.2% lower than the 3D recognition model, respectively, ranking third and fourth. The results show that TS-2D-CNN and SF-DPN models, due to the lack of frequency or time features, have much lower recognition rates for traffic interference (label 2), even below 90%. This indicates that the 3D-CNN model with 3D input has better feature extraction capabilities, thus outperforming the recognition performance of other 2D input models. This is because it contains the richest spatiotemporal dimension information in 3D-CNN. Furthermore, it shows that the information contribution of different dimensions is different: the identifiable information contained in the TF dimension is richer than that contained in the TS and SF dimensions.

[0198] With attention assistance, the attention maps obtained by 2D-CBAM and 3D-CBAM for the four typical DAS signals mentioned above are as follows: Figure 10As shown, compared to the original TF spectra in the first column, the TF spectra in the second and third columns contain much richer information, and the attention maps obtained in the second and third columns are enhanced in each relevant event. The results show that the attention map obtained by the 3D-CBAM in the third column focuses on the details of different events more accurately than the attention map obtained by the 2D-CBAM in the second column. For example, for traffic interference events, 3D-CBAM can more accurately capture the details of vehicle passage; for manual excavation, 3D-CBAM is better at capturing regular excavation moments than 2D-CBAM; for mechanical excavation, the information obtained by the 3D attention map is more detailed than that obtained by the 2D attention map; for background noise without events, the accuracy of key information in the 3D attention map is slightly higher than that of the broad and local attention regions in the 2D attention map. The results indicate that in recognition networks, the feature extraction and representation capabilities of 3D-CBAM are generally superior to those of 2D-CBAM. Therefore, this invention incorporates a 3D attention module to construct the final 3D-ACNN model.

[0199] For the latter three 3D recognition models, specifically, from Table 4 and Figure 11 It can be seen that 3D-ACNN has the best recognition performance for all four types of events, outperforming 3D-CNN+2D-CBAM and the basic 3D-CNN in that order. The average recognition accuracy of the 3D-ACNN model reaches 99.33%, which is 0.66% and 0.21% higher than the previous basic 3D CNN and 3D-CNN+2D-CBAM, respectively, on the same test set. We can conclude that: ① After adding the attention mechanism, the recognition performance can be further improved due to the more accurate attention to key features; ② The 3D attention module is superior to the 2D attention module and is more suitable for the 3D CNN model proposed in this invention; ③ The 3D-ACNN model has the best recognition capability to date, utilizing the richest multidimensional information and using this information in the most efficient way through a customized 3D attention mechanism.

[0200] In addition, the average time cost and number of model parameters for a test sample of the above six models were compared, as shown in the bar chart below. Figure 12As shown, the model sizes of the three 3D CNN models are all smaller than those of the other 2D CNN models. In terms of model size and number of parameters, the three 3D CNN models are basically the same, but 3D-ACNN has the lowest time cost, less than the basic 3D-CNN and 3D-CNN+2D-CBAM, and almost half the time cost of other 2D networks. 3D attention is a necessary step to improve recognition accuracy while accelerating processing speed, and 3D-ACNN maintains the best recognition performance while also being the fastest model. In online testing, the time cost of a spatiotemporal matrix (equivalent to 129m of fiber optic length and 1.29m spatial resolution) based on 100 sensor nodes using a commercial GPU was 22.48ms. This means that each spatial perception point only requires 0.22ms, only 1 / 171 of the time required for a 2D network with TF input. Therefore, 3D-ACNN also provides a good lightweight network, which helps DAS make more accurate decisions more effectively with excellent recognition performance and online real-time processing capabilities.

[0201] In summary, this invention proposes a fast and accurate multidimensional information extraction and collaborative recognition method for DAS systems. Utilizing a customized 3D attention module, it emphasizes the completeness and effectiveness of multidimensional feature extraction, achieving a lightweight 3D attention CNN (3D-ACNN). In real-world field tests, its recognition performance and timeliness are superior compared to other widely used 2D-CNN networks. Therefore, it can easily achieve accurate online recognition of large or ultra-large-scale arrays in many DAS applications.

[0202] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for multidimensional depth feature extraction and recognition of DAS signals, characterized in that, Including the following steps: Step 1: DAS data acquisition and construction of a two-dimensional spatiotemporal sample database. Distributed fiber optic sensing signals along pipelines in multiple scenarios are collected to obtain raw spatiotemporal signals. The collected raw spatiotemporal signals are then divided into matrix blocks according to the spatiotemporal width of the signal samples to obtain two-dimensional spatiotemporal signal samples. This allows the construction of a two-dimensional spatiotemporal sample database for different types of events. Step 2: 3D input preparation and signal preprocessing. The time signal of each spatial point in the two-dimensional spatiotemporal signal sample is converted into a time-frequency image, and the time-frequency image is preprocessed to construct three-dimensional time-frequency spatial input samples for different types of events. The above conversion is performed on all two-dimensional spatiotemporal signal samples in the database to obtain a dataset of three-dimensional time-frequency spatial input samples. The dataset is divided into training set, validation set and test set in a ratio of 6:2:

2. Step 3: Construct and optimize a three-dimensional attention convolutional neural network model to perform synchronous deep learning and attention on the time-frequency-space three-dimensional information of typical event signals of different types; First, we construct a basic 3D convolutional neural network model, abbreviated as 3D-CNN. Then, two-dimensional attention modules and three-dimensional attention modules adapted to three-dimensional convolution are added to the 3D-CNN model respectively to obtain 3D-CNN+2D-CBAM and 3D-CNN+3D-CBAM. The signals of the training set in step 2 are used as input to train the constructed 3D-CNN+2D-CBAM and 3D-CNN+3D-CBAM offline. The performance of each network is optimized through training iteration, adjustment and verification. The two attention mechanisms are compared and the model with the best performance is taken as the three-dimensional attention convolutional neural network model, abbreviated as 3D-ACNN. Step 4: Online testing and performance evaluation; In step 2, a short-time Fourier transform is used to convert the time signal of each spatial point in the two-dimensional spatiotemporal signal sample into a time-frequency diagram. Let the signal be... The required window function is ， Short-time Fourier Transform Defined as: （2） in: This indicates the short-time Fourier transform operation. This represents the calculated spectrum. Indicates a certain moment, This indicates a local time shift of the window function. ,use This allows us to locate the position of the local time window. In actual calculations, For discrete time points, a defined short-time signal for: （3） Again Performing a Fast Fourier Transform (FFT) to calculate its frequency components reduces the computational load. To ensure the resolution of the spectrum in the STFT, a rectangular window with a data length of 95 is selected, the jump count is one sampling point, and the FFT size is equal to the window length. Therefore, a raw spatiotemporal sample can be converted into a data sample of size [data size]. The three-dimensional samples represent a time length of 2 seconds, a frequency range of 250 Hz, and a spatial range of 129 m for each sample. The vibration event signals in the typical event database include: background noise, traffic interference, manual excavation, and mechanical excavation; The network architecture design of 3D-ACNN includes the design and construction of the basic 3D-CNN model, two-dimensional attention module, and three-dimensional attention module; First, a basic 3D-CNN network model is constructed. This model consists of an input layer, a convolutional layer C1, a pooling layer P1, a residual block 1, a convolutional layer C2, a pooling layer P2, a residual block 2, a fully connected layer FC1, a fully connected layer FC2, and an output layer. In the convolutional blocks, 3D convolutional kernels encode the 3D input. The TFS information of the DAS signal is extracted synchronously from three directions; in each residual block, there are two convolutional layers and one nonlinear activation function ReLU layer. The specific calculation operations are shown in equations (4) and (5): (4) (5) in: and Indicates the first The input and output of each residual block; This is an intermediate result; Let be the residual function, representing the learned residual terms; The identity mapping for the input, i.e. ; Using the ReLU function, the residual block accelerates the convergence speed, enhances nonlinearity, and avoids network degradation through skip connections and identity mapping. Then, an attention mechanism is added after the convolutional output of each residual block in the 3D-CNN. The entire network model is referred to as 3D-ACNN, and its network structure is as follows: Input layer - Convolutional layer C1 - Pooling layer P1 - Residual block + 3D-CBAM1 - Convolutional layer C2 - Pooling layer P2 - Residual block + 3D-CBAM2 - Fully connected layer FC1 - Fully connected layer FC2 - Output; The design and construction process of the three-dimensional attention module is as follows: Given an intermediate feature mapping Subsequently, the 3D-CBAM module derives the attention mapping sequentially along the three dimensions of time-frequency TF structure, space, and channels. Let the TF structure attention mapping be... Spatial attention mapping Channel attention mapping is The final feature refinement output The calculation process is as follows: （12） （13） （14） （15） （16） （17） in: express Time-frequency slices for each channel, , and These are the output feature maps obtained from local TF structure attention, spatial attention, and channel attention, respectively. This represents element-wise multiplication. It is a linear correction unit function. This represents the Sigmoid function. It has a core size of Convolution operation of a three-dimensional filter.

2. The method for multidimensional depth feature extraction and recognition of DAS signals according to claim 1, characterized in that, The system hardware used for DAS data acquisition in step 1 is a distributed fiber optic sound and vibration sensing system based on phase demodulation phase-sensitive optical time-domain reflectometry technology. The system hardware consists of three parts: a detection optical cable, an optical signal demodulation device, and a signal processing host. The system returns the original signal trajectory distributed along the space at each moment. By continuously accumulating N original signal trajectories on the time axis, a spatiotemporal signal matrix with T dimensions in time and S dimensions in space is constructed. （1） in: This represents the spacetime matrix. To represent a point in the matrix, t Indicates a specific moment. s It represents a specific point in space; N The time sampling length, M The spatial sampling length, with units of sampling points, represents the spatiotemporal response signal of the distributed fiber optic sound and vibration sensing system accumulated over time. The horizontal axis represents the spatial data acquisition points along the pipeline, and the distance between two spatial points is... The vertical axis represents the time axis, and the sampling interval between two sampling points is the vertical axis. , This refers to the pulse trigger frequency on the time axis, which is also the time sampling frequency. The collected spatiotemporal matrix is ​​divided into spatiotemporal regions to construct a two-dimensional spatiotemporal signal sample for each data point in the dataset.

3. The method for multidimensional depth feature extraction and recognition of DAS signals according to claim 1, characterized in that, Step 3 of the 3D-ACNN also includes network structure and parameter tuning, cost function design, and network update processes.

4. The method for multidimensional depth feature extraction and recognition of DAS signals according to claim 1, characterized in that, The design and construction process of the two-dimensional attention module is as follows: An attention module is used to select key information extracted from the 3D deep learning network. The attention mechanism employs a soft attention mechanism, where the input feature map is multiplied by an attention mask, calculated as follows: (6) in: It is the feature mapping at the output of the previous feature extractor. It is an attention mask. It is the output of the attention module; The specific implementation of the 2D attention module is as follows: given an intermediate feature map Subsequently, the 2D-CBAM module derives the attention mapping sequentially along the two dimensions of time-frequency structure and channel. Let the attention mapping of the TF structure be... Spatial splicing is Channel attention mapping is In each channel, its time-frequency slice Multiply by the local two-dimensional attention mask in the TF structure This yields the attention feature map for each spatial point; all TF feature maps along the fiber space are then stitched together to form a three-dimensional feature map. Then pay attention to the channel. Apply to all channels; Final feature refinement output The calculation process is as follows: （7） （8） （9） (10) (11) in: express Time-frequency slices for each channel, , and These are the output feature maps obtained from local TF structure attention, spatial attention, and channel attention, respectively. This represents element-wise multiplication. It is a linear correction unit function. This represents the Sigmoid function. It has a core size of The convolution operation of a two-dimensional filter, the semicolon ";" indicates the concatenation of set vectors. This represents a multilayer perceptron, similar to a fully connected layer, used to reduce the dimensionality of the input. 𝐴𝑣𝑔𝑃𝑜𝑜l and 𝑀𝑎𝑥𝑃𝑜𝑜l represent average pooling and max pooling, respectively.

5. The method for multidimensional depth feature extraction and recognition of DAS signals according to claim 3, characterized in that, The cost function design and network update process includes: after parameter initialization, the 3D-ACNN is input with data from the training set; the predicted class probability distribution is obtained through forward propagation; the loss value between the class probability output by the fully connected layer and the true probability is calculated using the cross-information entropy loss function; the gradient of each learning parameter is calculated using backpropagation based on the loss value; and the model parameters are updated using gradient descent with a specified learning rate. The model parameters Including matrix weights and bias ; The Xavier initialization method is used for parameter initialization. To ensure consistent variance across layers during forward and backward propagation, the random parameter initialization distribution ranges from the number of input parameters passing through that layer. Number of output parameters The formula for calculating the distribution range U of the obtained uniform distribution is: （18） The time-frequency spatial feature sample signals in the training set are input into the 3D-ACNN network model for forward propagation to obtain the predicted labels of the time-frequency spatial signal samples; The cost function chosen is the cross-information entropy loss function, which is used to calculate the loss value between the predicted label and the true label: the distance between the predicted event label and the true label is calculated based on the cross-information entropy loss function to obtain the loss value L, and the calculation formula is as follows: （19） in: , , These represent the total number of samples, the predicted label of the samples, and the true label of the samples, respectively. The parameter gradients of the 3D-ACNN network model are calculated in reverse using the loss value, and the 3D-ACNN deep learning model is updated using the parameter gradients. The Adam algorithm is used to optimize the network. The calculation steps are as follows: （20） （21） （22） in: To calculate the target gradient, , These are the first and second moment estimates of the gradient, respectively. , which are the exponential decay rates of the first and second moments of the gradient, respectively; Let the first iteration , , The network parameter update formula is: （23） in: For learning rate, For very small numbers, prevent division by zero; Using model parameters After updating the 3D-ACNN network model, the training loss value is used to determine whether the 3D-ACNN network model has converged. If it has converged, the training process ends; otherwise, the training is repeated until the maximum number of iterations is reached. After multiple iterations, the model is considered to have converged when the loss function value is less than a certain threshold or the iteration exceeds a certain threshold, and the above iteration process is stopped. After iterative convergence, the model is validated using a validation set, and the model with the best results is saved as the final event recognition model.

Images