Signal detection and processing method suitable for distributed optical fiber acoustic sensing

Abstract

The invention discloses a signal detection and processing method suitable for distributed optical fiber acoustic sensing, which comprises the following steps of: firstly, acquiring different types of distributed optical fiber acoustic sensing signals and preprocessing to construct an original sensing data set; the method comprises the following steps: firstly, constructing a DAS signal detection and processing network based on a weighted frequency domain neural operator and an adaptive sparse attention mechanism, finally, training the DAS signal detection and processing network, and carrying out signal detection based on the trained DAS signal detection and processing network to obtain a classification and positioning result of an event. Based on a weighted frequency domain neural operator and a self-adaptive sparse attention mechanism, spectrum global features and spatial local details are dynamically integrated, and sub-meter spatial positioning and multi-category recognition of vibration / acoustic events in a strong noise environment are realized.

Description

Technical Field

[0001] This invention relates to the field of distributed fiber optic acoustic sensor technology, specifically a signal detection and processing method applicable to distributed fiber optic acoustic sensing. Background Technology

[0002] Fiber-optic distributed acoustic sensors (DAS) based on phase-sensitive optical time-domain reflectometry (OTDR) technology are high-density, low-cost sensors. Their advantages, such as providing long-distance, high-resolution acoustic signal monitoring, continuous large-scale detection and localization of acoustic signals acting on the sensing fiber, high sensitivity, and fast response, have attracted close attention and led to their widespread application. DAS systems can detect and locate events occurring along the fiber path, such as intrusions, vibrations, or temperature changes, by analyzing changes in the optical signal propagating in the fiber. Utilizing widely deployed underground communication fibers, various vibration sources can be sensed and located with high sensitivity and accuracy. Therefore, DAS systems have been applied to various underground and surface monitoring applications, such as natural disaster prediction of submarine seismic activity, volcanic events and earthquakes, energy exploration in the oil and gas industry, safety monitoring of civil infrastructure in energy transmission pipelines, railways, perimeter security, and safety monitoring of urban underground structures. However, in practical applications, due to the one-dimensional spatiotemporal distribution characteristics of DAS signals, and the often accompanied by low signal-to-noise ratios and sparse event categories, traditional detection methods struggle to maintain high accuracy while simultaneously ensuring real-time performance and generalization capabilities.

[0003] In recent years, with the continuous development of artificial intelligence technology, deep learning-based security threat event detection technologies have emerged. Numerous studies have shown that models built based on deep learning methods can achieve good performance in security threat event detection. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a signal detection and processing method suitable for distributed fiber optic acoustic sensing. Based on weighted frequency domain neural operators and adaptive sparse attention mechanism, it dynamically integrates global spectral features and local spatial details to achieve sub-meter level spatial localization and multi-class identification of vibration / acoustic events in strong noise environment.

[0005] The technical solution of this invention is as follows:

[0006] A signal detection and processing method suitable for distributed fiber optic acoustic sensing specifically includes the following steps:

[0007] (1) Collect different types of distributed fiber acoustic sensing signals through a distributed fiber sensing data acquisition platform, and construct the original sensing dataset after preprocessing.

[0008] (2) Construct a DAS signal detection and processing network. The sensor signal detection and processing network includes a DAS signal feature extraction network, a gated DAS signal feature fusion network, and a DAS signal event category and location recognition network.

[0009] The DAS signal feature extraction network uses a spread frequency domain neural operator to refine low-quality signals into high-quality signals, capturing the complex conditional distribution of high-quality signals under low-quality input.

[0010] A gated DAS signal feature fusion network is used to integrate global dependencies captured by weighted frequency domain neural operators with local information provided by an adaptive sparse attention module through spatial gating.

[0011] The DAS signal event category and location identification network is used to obtain the final event classification and location results;

[0012] (3) Construct a loss function to train the DAS signal detection and processing network. Input the original sensor signals to be detected in the original sensor dataset into the trained DAS signal detection and processing network for detection, and obtain the classification and localization results of the events.

[0013] The gated DAS signal feature fusion network includes a weighted frequency domain neural operator, an adaptive sparse attention module, and a gated fusion module. The weighted frequency domain neural operator models global dependencies through spectral convolution, effectively capturing long-range interactions and overall structure. The adaptive sparse attention module uses a soft threshold function to adjust the attention weights in the attention mechanism to obtain attention-enhanced features. The gated fusion module is used to perform gated fusion of the features processed by the adaptive sparse attention module and the features processed by the weighted frequency domain neural operator to obtain fused features.

[0014] The DAS signal feature extraction network described above is based on a diffusion frequency domain neural operator, which includes a forward diffusion process, a reverse diffusion process, and a weighted frequency domain neural operator. The forward diffusion process is set to be continuous time. Upsampling and downsampling The forward diffusion process is represented by a variance-preserving stochastic differential equation, as shown in equation (1) below:

[0015]

[0016] In equation (2), x t dx represents the high-quality DAS signal input at time t. t This represents the DAS signal output after forward diffusion at time t; dW represents the downsampling operator; θ(t) represents the noise sequence; dW represents the standard Wiener process. This is the drift term, used to simulate the loss of high-frequency details during downsampling. The diffusion term is used to add Gaussian noise; θ max Represents the maximum noise intensity, θ min Represents the minimum noise intensity;

[0017] Weighted frequency domain neural operators are applied to each diffusion time step in the forward diffusion process;

[0018] The reverse diffusion module is used to learn the inverse operation of the forward diffusion process of the forward diffusion module, and is used to reverse diffuse the low-quality DAS signal input into the DAS signal feature extraction network into a high-quality DAS signal.

[0019] The learning functions of the weighted frequency domain neural operator in the gated DAS signal feature fusion network and the weighted frequency domain neural operator in the spread frequency domain neural operator are both given by the following equation (2):

[0020] x h (p)=Θ[x l (p)] (2);

[0021] In equation (2), x l (p) represents the high-quality DAS signal output by the DAS signal feature extraction network, x h (p) represents the high-quality DAS signal output after processing by the weighted frequency domain neural operator, and Θ represents the weighted frequency domain neural operator;

[0022] The weighted frequency domain neural operator Θ is represented by stacking, as shown in equation (3) below:

[0023] w i+1 (p)=Σ(Xw i (p)+Γ i w i (p)) (3);

[0024] In equation (3), w i (p) represents the feature representation of the i-th layer at spatial location p; w i+1 (p) represents the feature representation at spatial location p after the (i+1)th layer update; X represents the linear transformation; Γ i Σ represents the integral operator of the i-th layer, used to transform the features to the frequency domain. The specific transformation process is shown in equation (4); Σ() represents the nonlinear activation function.

[0025]

[0026] In equation (4), Represents inverse frequency domain transform; ω i (γ) represents the weighting coefficient; λ iβ represents the learnable scalar parameter of the i-th layer, used to control the strength of the weight coefficients; β represents the hyperparameter, used to determine how the weight coefficients scale with frequency amplitude ||γ||; F i (γ) is a complex tensor composed of learnable parameters, representing a frequency domain filter; w represents the feature of the i-th layer i Frequency domain transformation of values ​​taken at frequency γ.

[0027] The adaptive sparse attention module includes a local exploration layer, an adaptive sparse attention layer, and a feature refinement layer. The local exploration layer consists of multiple consecutive convolutional residual blocks, where the output of the previous convolutional residual block is used as the input of the next convolutional residual block, and the feature output of the last convolutional residual block is used as the output of the local exploration layer.

[0028] Feature x in adaptive sparse attention layer i The response Φ(x) i ) is expressed as the following formula (5):

[0029]

[0030] In equation (5), the input features of the adaptive sparse attention layer are X∈R L×C x j It is the j-th feature on X; L represents the length of the feature vector of the input feature X; η v () represents the value vector generation function, which uses a 1×1 convolution operation; the j-th feature x j Allocation is performed using a soft threshold operation. It is the j-th component of the soft thresholding result, and the calculation process is shown in the following formula (6):

[0031]

[0032] In equation (6), Represents similarity vectors The j-th component, the similarity vector The calculation formula is shown in equation (7) below; It is a soft threshold function that satisfies the following equation (8);

[0033]

[0034] In equation (7), η q () and η k () represent the query vector generation function and the key vector generation function, respectively, both using a 1×1 convolution operation. T represents transpose, and δ(x) i ,x j ) represents the characteristic x i and feature x jDot product similarity between them;

[0035] To derive the soft threshold function First, consider the similarity vectors. Sort, and make the sorted satisfy Let Ω = {1,…,L}, and define M as the following equation (9):

[0036]

[0037] In equation (9), M represents the maximum index that satisfies the above conditions;

[0038] Then, the soft threshold function It is derived from equations (8) and (9), as shown in equation (10) below:

[0039]

[0040] The soft threshold function is calculated according to equation (10). After obtaining the threshold result, according to Equation (8), the similarity vector components below the threshold result are replaced with zero, and the remaining similarity vector components are replaced with the difference between the threshold result and the value of the threshold result, thereby achieving threshold shift.

[0041] The feature refinement layer is processed by 7×1 convolution operation and then residually connected to the input to obtain attention-enhanced features.

[0042] The processing procedure of the gated fusion module is as follows: First, the gated vector is calculated. See the following formula (11) for details:

[0043]

[0044] In equation (11), This represents the feature vector corresponding to the high-quality DAS signal output after processing by the weighted frequency domain neural operator. The eigenvectors representing the high-quality DAS signals output after processing by the adaptive sparse attention module; Θ[x] and In the diagram, x represents the feature vector corresponding to the high-quality DAS signal output by the DAS signal feature extraction network; [,] represents concatenation along the channel dimension; Conv 1×1 ∑ represents a 1×1 convolutional layer, used to reduce the concatenated features to a single-channel gated vector; ∑() represents a non-linear activation function;

[0045] Then, the element-wise weighted sum of the two eigenvectors is calculated as shown in equation (12):

[0046]

[0047] In equation (12), m 融合 ⊙ represents the fused feature vector, and ⊙ represents element-wise multiplication.

[0048] The DAS signal event category and location identification network includes a location regression branch and a classification branch, which respectively obtain the classification and location results of the event;

[0049] The localization regression branch includes three consecutive convolutional residual blocks, IoU loss and distributed focus loss DFL. The IoU loss and distributed focus loss DFL are calculated based on the localization result output by the last convolutional residual block. The calculation of IoU loss is shown in Equation (13), and the calculation of distributed focus loss DFL is shown in Equation (14).

[0050]

[0051] In equation (13), I pre I represents the prediction interval for positioning. gt `Distance()` represents the actual interval corresponding to the event; `Distance()` represents the length of the interval.

[0052]

[0053] In equation (14), y di p represents the actual distance label di Represents the predicted probability;

[0054] The classification branch includes a first convolutional residual block, a second convolutional residual block, a third convolutional residual block, a first depthwise separable convolutional block, a second depthwise separable convolutional block, and a binary cross-entropy loss (BCE). The output of the first depthwise separable convolutional block is used as the input of the first convolutional residual block, the output of the first convolutional residual block is used as the input of the second depthwise separable convolutional block, the output of the second depthwise separable convolutional block is used as the input of the second convolutional residual block, and the output of the second convolutional residual block is used as the input of the third convolutional residual block. The third convolutional residual block outputs the classification result. The binary cross-entropy loss (BCE) is calculated based on the classification result output by the third convolutional residual block. The calculation of the binary cross-entropy loss (BCE) is shown in the following formula (15):

[0055]

[0056] In equation (15), N represents the number of samples, y ci p represents the true classification label of sample i. ci It is the probability that sample i is predicted to be a positive sample.

[0057] The loss function used for training the DAS signal detection and processing network is obtained by weighted summation of IoU loss, distributed focus loss (DFL), and binary cross-entropy loss (BCE). IoU loss and distributed focus loss (DFL) are used for localization regression branch, and binary cross-entropy loss (BCE) is used for classification branch.

[0058] Advantages of this invention:

[0059] (1) The neural operator is an efficient variant of the neural operator, designed to learn the mapping relationships between function spaces. It directly acts on input data of arbitrary quality, achieving signal quality reconstruction by mapping low-quality input to high-quality output. Specifically, it first transforms the input data to the frequency domain, applies a learned filter, and then transforms the data back to the spatial domain. Frequency domain convolution and mode truncation significantly improve computational efficiency and accuracy. Therefore, the neural operator is well-suited for DAS vibration signal resolution tasks due to its inherent resolution invariance and ability to efficiently model global dependencies. However, mode truncation underestimates the high-frequency components that are crucial for the quality reconstruction of real-world vibration signals. To address this limitation, this invention employs a weighted frequency domain neural operator to capture global dependencies, effectively capturing key frequency components, significantly improving the reconstruction effect of high-frequency details crucial in signal quality reconstruction tasks, enhancing the ability to capture key frequency components of weak events, and significantly improving the signal-to-noise ratio, achieving a dual breakthrough in accuracy and computational efficiency.

[0060] (2) Neural operators are well-suited for DAS vibration signal resolution tasks, possessing inherent resolution invariance and the ability to efficiently model global dependencies. Diffusion models can iteratively refine low-quality signals into high-quality signals, capturing the complex conditional distribution of high-quality signals under low-quality input. This invention combines the above two frameworks to construct a diffusion frequency domain neural operator for feature extraction of DAS signals. The weighted frequency domain neural operator is applied to each diffusion time step in the forward diffusion process, transforming the current degraded signal in the forward diffusion process to the frequency domain. Frequency domain convolution and learnable frequency domain filters are applied to enhance high-frequency information, and then the transformation is reversed back to the spatiotemporal domain. Learnable weight coefficients are used to assign greater weight to high-frequency components, thereby enhancing key signal details. This allows the diffusion model to refine low-quality signals into high-quality signals, capturing the complex conditional distribution of high-quality signals under low-quality input.

[0061] (3) The adaptive sparse attention module of the present invention addresses the defect of the soft maximization transformation in the traditional attention mechanism being sensitive to irrelevant information. By using soft thresholding, it sparsifies the attention weights, suppresses low similarity positions, and retains only the information of high similarity regions. While retaining the basic properties of the soft maximization transformation, it makes all probability values ​​greater than zero and sums to one, and assigns higher weights to feature vectors with higher similarity.

[0062] (4) The gated DAS signal feature fusion network of the present invention is used to integrate the global dependencies captured by the weighted frequency domain neural operator with the local information provided by the adaptive sparse attention module through spatial gating. The weighted frequency domain neural operator models the global dependencies through spectral convolution, effectively capturing long-range interactions and overall structure. The adaptive sparse attention module quickly captures local dependencies and fine-grained details through the attention mechanism. The gated fusion mechanism adaptively combines the feature maps of the weighted frequency domain neural operator and the adaptive sparse attention module, dynamically balancing global and local information, thereby improving the quality of high DAS signal reconstruction.

[0063] (5) The DAS signal event category and location recognition network of the present invention includes two branches, which are used to obtain the final event classification result and location result respectively, and are trained with different loss functions, so as to realize sub-meter level spatial positioning and multi-category detection and recognition of vibration / acoustic events in a strong noise environment. Attached Figure Description

[0064] Figure 1 This is a flowchart of the present invention.

[0065] Figure 2 This is a schematic diagram of the structure of the DAS signal detection and processing network of the present invention.

[0066] Figure 3 This is a network architecture diagram of the adaptive sparse attention module of the present invention.

[0067] Figure 4 This is a network architecture diagram of the adaptive sparse attention layer of this invention.

[0068] Figure 5 This is a network architecture diagram of the DAS signal event category and location identification network of the present invention.

[0069] Figure 6 This is a network architecture diagram of the convolutional residual block of the present invention. Detailed Implementation

[0070] 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] See Figure 1 and Figure 2 A signal detection and processing method suitable for distributed fiber optic acoustic sensing, specifically including the following steps:

[0072] (1) Collect different types of distributed optical fiber acoustic sensor signals through the distributed optical fiber sensor data acquisition platform, including DAS signals of events such as mechanical digging, manual hoeing, and vehicle driving. After preprocessing, construct the original sensor dataset and divide it into training dataset and test dataset.

[0073] (2) Construct a DAS signal detection and processing network. The sensor signal detection and processing network includes a DAS signal feature extraction network, a gated DAS signal feature fusion network, and a DAS signal event category and location recognition network.

[0074] The DAS signal feature extraction network uses a diffusion frequency domain neural operator to refine low-quality signals into high-quality signals. The diffusion frequency domain neural operator includes a forward diffusion process, a reverse diffusion process, and a weighted frequency domain neural operator. The weighted frequency domain neural operator is applied to each diffusion time step in the forward diffusion process, thereby capturing the complex conditional distribution of high-quality signals under low-quality input.

[0075] The gated DAS signal feature fusion network includes a weighted frequency domain neural operator, an adaptive sparse attention module, and a gated fusion module. The weighted frequency domain neural operator models global dependencies through spectral convolution, effectively capturing long-range interactions and overall structure. The adaptive sparse attention module uses a soft threshold function to adjust the attention weights in the attention mechanism to obtain attention-enhanced features. The gated fusion module is used to perform gated fusion of the features processed by the adaptive sparse attention module and the features processed by the weighted frequency domain neural operator to obtain fused features, thereby integrating the global dependencies captured by the weighted frequency domain neural operator with the local information provided by the adaptive sparse attention module.

[0076] The DAS signal event category and location identification network is used to obtain the final event classification and location results;

[0077] (3) Construct a loss function, input the training dataset into the DAS signal detection and processing network for training, input the test dataset into the trained DAS signal detection and processing network for detection testing, and obtain the classification and localization results of the event.

[0078] In this context, the forward diffusion process in the diffusion frequency domain neural operator is set in continuous time. Up-simulation downsampling, The forward diffusion process is represented by a variance-preserving stochastic differential equation, as shown in equation (1) below:

[0079]

[0080] In equation (2), x t dx represents the high-quality DAS signal input at time t. t This represents the DAS signal output after forward diffusion at time t; represents the downsampling operator, used to reduce the spatial dimension of a signal by a specified scaling factor, specifically using bicubic interpolation downsampling; θ(t) represents the noise sequence; dW represents the standard Wiener process; This is the drift term, used to simulate the loss of high-frequency details during downsampling. The diffusion term is used to add Gaussian noise; θ max Represents the maximum noise intensity, θ max =20; θ min Representing the minimum noise intensity, θ min =0;

[0081] Weighted frequency domain neural operators are applied to each diffusion time step in the forward diffusion process;

[0082] The reverse diffusion module is used to learn the inverse operation of the forward diffusion process of the forward diffusion module, and is used to reverse diffuse the low-quality DAS signal input into the DAS signal feature extraction network into a high-quality DAS signal.

[0083] The learning functions of the weighted frequency domain neural operator in the gated DAS signal feature fusion network and the weighted frequency domain neural operator in the spread frequency domain neural operator are both given by the following equation (2):

[0084] x h (p)=Θ[x l (p)] (2);

[0085] In equation (2), x l (p) represents the high-quality DAS signal output by the DAS signal feature extraction network, x h (p) represents the high-quality DAS signal output after processing by the weighted frequency domain neural operator, and Θ represents the weighted frequency domain neural operator; Represents spatial coordinates;

[0086] The weighted frequency domain neural operator Θ is represented by stacking, as shown in equation (3) below:

[0087] w i+1 (p)=∑(Xw i (p)+Γ i w i (p)) (3);

[0088] In equation (3), w i (p) represents the feature representation of the i-th layer at spatial location p; w i+1 (p) represents the feature representation at spatial location p after the (i+1)th layer update; X represents the linear transformation; Γ i ∑() represents the integral operator of the i-th layer, used to transform the features to the frequency domain. The specific transformation process is shown in equation (4) below; ∑() represents the nonlinear activation function;

[0089]

[0090] In equation (4), Represents inverse frequency domain transform; ω i (γ) represents the weighting coefficient; λ i β represents the learnable scalar parameter of the i-th layer, used to control the strength of the weight coefficients; β represents the hyperparameter, used to determine how the weight coefficients scale with frequency amplitude ||γ||; F i (γ) is a complex tensor composed of learnable parameters, representing a frequency domain filter; w represents the feature of the i-th layer i Frequency domain transformation of values ​​taken at frequency γ.

[0091] See Figure 3 The adaptive sparse attention module includes a local exploration layer, an adaptive sparse attention layer, and a feature refinement layer.

[0092] The local exploration layer consists of three consecutive convolutional residual blocks (CRBs). The output of the previous CRB is used as the input of the next CRB, and the feature output of the last CRB is used as the output of the local exploration layer.

[0093] See Figure 6 Each convolutional residual block includes two 7×1 convolutions, a ReLU function, and a residual block. The output of the first 7×1 convolution is normalized by the ReLU function and then input into the second 7×1 convolution. The output of the second 7×1 convolution is then output after being concatenated with the residual block.

[0094] See Figure 4 Feature x in the adaptive sparse attention layer i The response Φ(x)i ) is expressed as the following formula (5):

[0095]

[0096] In equation (5), the input features of the adaptive sparse attention layer are X∈R L×C x j It is the j-th feature on X; L represents the length of the feature vector of the input feature X; η v () represents the generating function of the value vector V, which uses a 1×1 convolution operation; the j-th feature x j Allocation is performed using a soft threshold operation. It is the j-th component of the soft thresholding result, and the calculation process is shown in the following formula (6):

[0097]

[0098] In equation (6), Represents similarity vectors The j-th component, the similarity vector The calculation formula is shown in equation (7) below; It is a soft threshold function that satisfies the following equation (8);

[0099]

[0100] In equation (7), η q () and η k () represent the generating functions of the query vector Q and the key vector K, respectively, both using a 1×1 convolution operation. T represents the transpose, and δ(x) i ,x j ) represents the characteristic x i and feature x j Dot product similarity between them;

[0101] To derive the soft threshold function First, consider the similarity vectors. Sort, and make the sorted satisfy Let Ω = {1,…,L}, and define M as the following equation (9):

[0102]

[0103] In equation (9), M represents the maximum index that satisfies the above conditions, such that the sum of the first M sorted similarity vector components after threshold offset is equal to 1, and the remaining similarity vector components are replaced with zero.

[0104] Then, the soft threshold function It is derived from equations (8) and (9), as shown in equation (10) below:

[0105]

[0106] The soft threshold function is calculated according to equation (10). After obtaining the threshold result, according to Equation (8), the similarity vector components below the threshold result are replaced with zero, and the remaining similarity vector components are replaced with the difference between the threshold result and the value of the threshold result, thereby achieving threshold shift.

[0107] The feature refinement layer is processed by 7×1 convolution operation and then residually connected to the input to obtain attention-enhanced features.

[0108] The gating fusion module processes the following steps: First, it calculates the gating vector. See the following formula (11) for details:

[0109]

[0110] In equation (11), This represents the feature vector corresponding to the high-quality DAS signal output after processing by the weighted frequency domain neural operator. The eigenvectors representing the high-quality DAS signals output after processing by the adaptive sparse attention module; Θ[x] and In the diagram, x represents the feature vector corresponding to the high-quality DAS signal output by the DAS signal feature extraction network; [,] represents concatenation along the channel dimension; Conv 1×1 ∑ represents a 1×1 convolutional layer, used to reduce the concatenated features to a single-channel gated vector; ∑() represents a non-linear activation function;

[0111] Then, the element-wise weighted sum of the two eigenvectors is calculated as shown in equation (12):

[0112]

[0113] In equation (12), m 融合 ⊙ represents the fused feature vector, and ⊙ represents element-wise multiplication.

[0114] See Figure 5 The DAS signal event category and location identification network includes a location regression branch and a classification branch, which respectively obtain the classification and location results of the event;

[0115] The localization regression branch includes three consecutive convolutional residual blocks (CRB) (see...). Figure 6IoU loss and distributed focus loss DFL are calculated based on the localization results output by the last convolutional residual block CRB. The calculation of IoU loss is shown in Equation (13), and the calculation of distributed focus loss DFL is shown in Equation (14).

[0116]

[0117] In equation (13), I pre I represents the prediction interval for positioning. gt The Diistance() value represents the actual interval corresponding to the event; Diistance() represents the length of the interval.

[0118]

[0119] In equation (14), y di p represents the actual distance label di Represents the predicted probability;

[0120] The classification branches include the first convolutional residual block (CRB), the second convolutional residual block (CRB), the third convolutional residual block (CRB), the first depthwise separable convolutional block (DWConv), the second depthwise separable convolutional block (DWConv), and the binary cross-entropy loss (BCE). The structures of the first, second, and third convolutional residual blocks (CRB) are consistent and can be found in [the relevant documentation]. Figure 6 The output of the first depthwise separable convolutional block DWConv is used as the input of the first convolutional residual block CRB. The output of the first convolutional residual block CRB is used as the input of the second depthwise separable convolutional block DWConv. The output of the second depthwise separable convolutional block DWConv is used as the input of the second convolutional residual block CRB. The output of the second convolutional residual block CRB is used as the input of the third convolutional residual block CRB. The third convolutional residual block CRB outputs the classification result. The binary cross-entropy loss BCE is calculated based on the classification result output by the third convolutional residual block. The calculation of the binary cross-entropy loss BCE is shown in the following formula (15):

[0121]

[0122] In equation (15), N represents the number of samples, y ci p represents the true classification label of sample i. ci It is the probability that sample i is predicted as a positive sample.

[0123] The loss function used for training the DAS signal detection and processing network is the supervision loss obtained by weighted summation of IoU loss, distribution focus loss (DFL), and binary cross-entropy loss (BCE). IoU loss and distribution focus loss (DFL) are used to locate the regression branch, while binary cross-entropy loss (BCE) is used for the classification branch. During the training phase, DAS signals with real labels are weakly enhanced and then input into the DAS signal detection and processing network for supervised training, and the supervised loss is calculated.

[0124]

[0125] In equation (16), x i y represents the DAS signal data of sample i. i The true label of sample i. This represents the classification loss (binary cross-entropy loss) for the DAS event classification labels. This represents the localization regression loss (including IoU loss and distribution focus loss, DFL) for DAS event interval labels.

[0126] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A signal detection and processing method suitable for distributed fiber optic acoustic sensing, characterized in that: Specifically, it includes the following steps: (1) Collect different types of distributed fiber acoustic sensing signals through a distributed fiber sensing data acquisition platform, and construct the original sensing dataset after preprocessing. (2) Construct a DAS signal detection and processing network. The sensor signal detection and processing network includes a DAS signal feature extraction network, a gated DAS signal feature fusion network, and a DAS signal event category and location recognition network. The DAS signal feature extraction network uses a spread frequency domain neural operator to refine low-quality signals into high-quality signals, capturing the complex conditional distribution of high-quality signals under low-quality input. A gated DAS signal feature fusion network is used to integrate global dependencies captured by weighted frequency domain neural operators with local information provided by an adaptive sparse attention module through spatial gating. The DAS signal event category and location identification network is used to obtain the final event classification and location results; (3) Construct a loss function to train the DAS signal detection and processing network. Input the original sensor signals to be detected in the original sensor dataset into the trained DAS signal detection and processing network for detection, and obtain the classification and localization results of the events.

2. The signal detection and processing method for distributed fiber optic acoustic sensing according to claim 1, characterized in that: The gated DAS signal feature fusion network includes a weighted frequency domain neural operator, an adaptive sparse attention module, and a gated fusion module. The weighted frequency domain neural operator models global dependencies through spectral convolution, effectively capturing long-range interactions and overall structure. The adaptive sparse attention module uses a soft threshold function to adjust the attention weights in the attention mechanism to obtain attention-enhanced features. The gated fusion module is used to perform gated fusion of the features processed by the adaptive sparse attention module and the features processed by the weighted frequency domain neural operator to obtain fused features.

3. The signal detection and processing method for distributed fiber optic acoustic sensing according to claim 2, characterized in that: The DAS signal feature extraction network described above is based on a diffusion frequency domain neural operator, which includes a forward diffusion process, a reverse diffusion process, and a weighted frequency domain neural operator. The forward diffusion process is set to be continuous time. Up-simulation downsampling, The forward diffusion process is represented by a variance-preserving stochastic differential equation, as shown in equation (1) below: In equation (2), x t dx represents the high-quality DAS signal input at time t. t This represents the DAS signal output after forward diffusion at time t; dW represents the downsampling operator; θ(t) represents the noise sequence; dW represents the standard Wiener process. This is the drift term, used to simulate the loss of high-frequency details during downsampling. This is the diffusion term, used to add Gaussian noise; θ max Represents the maximum noise intensity, θ min Represents the minimum noise intensity; Weighted frequency domain neural operators are applied to each diffusion time step in the forward diffusion process; The reverse diffusion module is used to learn the inverse operation of the forward diffusion process of the forward diffusion module, and is used to reverse diffuse the low-quality DAS signal input into the DAS signal feature extraction network into a high-quality DAS signal.

4. The signal detection and processing method for distributed fiber optic acoustic sensing according to claim 3, characterized in that: The learning functions of the weighted frequency domain neural operator in the gated DAS signal feature fusion network and the weighted frequency domain neural operator in the spread frequency domain neural operator are both given by the following equation (2): x h (p)=Θ[x l (p)] (2); In equation (2), x l (p) represents the high-quality DAS signal output by the DAS signal feature extraction network, x h (p) represents the high-quality DAS signal output after processing by the weighted frequency domain neural operator, and Θ represents the weighted frequency domain neural operator; The weighted frequency domain neural operator Θ is represented by stacking, as shown in equation (3) below: w i+1 (p)=∑(Xw i (p)+Γ i w i (p)) (3); In equation (3), w i (p) represents the feature representation of the i-th layer at spatial location p; w i+1 (p) represents the feature representation at spatial location p after the (i+1)th layer update; X represents a linear transformation; Γ i ∑() represents the integral operator of the i-th layer, used to transform the features to the frequency domain. The specific transformation process is shown in equation (4) below; ∑() represents the nonlinear activation function; In equation (4), Represents inverse frequency domain transform; ω i (γ) represents the weighting coefficient; λ i Represents the learnable scalar parameters of the i-th layer, used to control the strength of the weight coefficients; β represents a hyperparameter used to determine how the weighting coefficients scale with the frequency amplitude ||γ||. F i (γ) is a complex tensor composed of learnable parameters, representing a frequency domain filter; w represents the feature of the i-th layer i Frequency domain transformation of values ​​taken at frequency γ.

5. The signal detection and processing method for distributed fiber optic acoustic sensing according to claim 4, characterized in that: The adaptive sparse attention module includes a local exploration layer, an adaptive sparse attention layer, and a feature refinement layer. The local exploration layer consists of multiple consecutive convolutional residual blocks, where the output of the previous convolutional residual block is used as the input of the next convolutional residual block, and the feature output of the last convolutional residual block is used as the output of the local exploration layer. Feature x in adaptive sparse attention layer i The response Φ(x) i ) is expressed as the following formula (5): In equation (5), the input features of the adaptive sparse attention layer are X∈R L×C x j It is the j-th feature on X; L represents the length of the feature vector of the input feature X; η v () represents the value vector generation function, which uses a 1×1 convolution operation; the j-th feature x j Allocation is performed using a soft threshold operation. It is the j-th component of the soft thresholding result, and the calculation process is shown in the following formula (6): In equation (6), Represents similarity vectors The j-th component, the similarity vector The calculation formula is shown in equation (7) below; It is a soft threshold function that satisfies the following equation (8); In equation (7), η q () and η k () represent the query vector generation function and the key vector generation function, respectively, both using a 1×1 convolution operation. T represents transpose, and δ(x) i ,x j ) represents the characteristic x i and feature x j Dot product similarity between them; To derive the soft threshold function First, consider the similarity vectors. Sort, and make the sorted satisfy Let Ω = {1,…,L}, and define M as the following equation (9): In equation (9), M represents the maximum index that satisfies the above conditions; Then, the soft threshold function It is derived from equations (8) and (9), as shown in equation (10) below: The soft threshold function is calculated according to equation (10). After obtaining the threshold result, according to Equation (8), the similarity vector components below the threshold result are replaced with zero, and the remaining similarity vector components are replaced with the difference between the threshold result and the value of the threshold result, thereby achieving threshold shift. The feature refinement layer is processed by 7×1 convolution operation and then residually connected to the input to obtain attention-enhanced features.

6. The signal detection and processing method for distributed fiber optic acoustic sensing according to claim 5, characterized in that: The processing procedure of the gated fusion module is as follows: First, the gated vector is calculated. See the following formula (11) for details: In equation (11), This represents the feature vector corresponding to the high-quality DAS signal output after processing by the weighted frequency domain neural operator. The eigenvectors representing the high-quality DAS signals output after processing by the adaptive sparse attention module; Θ[x] and In the diagram, x represents the feature vector corresponding to the high-quality DAS signal output by the DAS signal feature extraction network; [,] represents concatenation along the channel dimension; Conv 1×1 ∑ represents a 1×1 convolutional layer, used to reduce the concatenated features to a single-channel gated vector; ∑() represents a non-linear activation function; Then, the element-wise weighted sum of the two eigenvectors is calculated as shown in equation (12): In equation (12), m 融合 ⊙ represents the fused feature vector, and ⊙ represents element-wise multiplication.

7. The signal detection and processing method for distributed fiber optic acoustic sensing according to claim 2, characterized in that: The DAS signal event category and location identification network includes a location regression branch and a classification branch, which respectively obtain the classification and location results of the event; The localization regression branch includes three consecutive convolutional residual blocks, IoU loss and distributed focus loss DFL. The IoU loss and distributed focus loss DFL are calculated based on the localization result output by the last convolutional residual block. The calculation of IoU loss is shown in Equation (13), and the calculation of distributed focus loss DFL is shown in Equation (14). In equation (13), I pre I represents the prediction interval for positioning. gt `Distance()` represents the actual interval corresponding to the event; `Distance()` represents the length of the interval. In equation (14), y di p represents the actual distance label di Represents the predicted probability; The classification branch includes a first convolutional residual block, a second convolutional residual block, a third convolutional residual block, a first depthwise separable convolutional block, a second depthwise separable convolutional block, and a binary cross-entropy loss (BCE). The output of the first depthwise separable convolutional block is used as the input of the first convolutional residual block, the output of the first convolutional residual block is used as the input of the second depthwise separable convolutional block, the output of the second depthwise separable convolutional block is used as the input of the second convolutional residual block, and the output of the second convolutional residual block is used as the input of the third convolutional residual block. The third convolutional residual block outputs the classification result. The binary cross-entropy loss (BCE) is calculated based on the classification result output by the third convolutional residual block. The calculation of the binary cross-entropy loss (BCE) is shown in the following formula (15): In equation (15), N represents the number of samples, y ci p represents the true classification label of sample i. ci It is the probability that sample i is predicted as a positive sample.

8. The signal detection and processing method for distributed fiber optic acoustic sensing according to claim 7, characterized in that: The loss function used for training the DAS signal detection and processing network is obtained by weighted summation of IoU loss, distributed focus loss (DFL), and binary cross-entropy loss (BCE). IoU loss and distributed focus loss (DFL) are used for localization regression branch, and binary cross-entropy loss (BCE) is used for classification branch.

Images