A signal detection and processing method suitable for distributed fiber optic acoustic sensing

Abstract

The application discloses a signal detection and processing method suitable for distributed optical fiber acoustic sensing. Different types of distributed optical fiber acoustic sensing signals are collected and preprocessed first, and an original sensing data set is constructed. Then, a DAS signal detection and processing network based on a weighted frequency domain neural operator and an adaptive sparse attention mechanism is constructed. Finally, the DAS signal detection and processing network is trained, and signal detection is performed based on the trained DAS signal detection and processing network to obtain classification and positioning results of events. The application dynamically integrates global spectral features and local spatial details based on the weighted frequency domain neural operator and the adaptive sparse attention mechanism, so that sub-meter spatial positioning and multi-class identification of vibration / acoustic events in a strong noise environment are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of distributed optical fiber acoustic sensors, in particular to a signal detection and processing method suitable for distributed optical fiber acoustic sensing. BACKGROUND

[0002] The optical fiber distributed acoustic sensor (DAS) based on the phase-sensitive optical time domain reflectometer (OTDR) technology is a high-density and low-cost sensor. It has attracted close attention and been widely used due to its advantages of long-distance, high-resolution acoustic signal monitoring, continuous large-range detection and positioning of acoustic signals acting on the sensing optical fiber, high sensitivity and fast response. The DAS system can detect and locate events such as intrusion, vibration or temperature change along the optical fiber path by analyzing the changes of the optical signals propagating in the optical fiber. The widely laid underground communication optical fiber can be used to sense and locate various vibration sources with high sensitivity and high accuracy, so the DAS system has been applied to various underground and ground monitoring applications, such as natural disaster prediction of submarine seismic activity, volcanic events and earthquakes, energy exploration in the oil and gas industry, and safety monitoring of civil infrastructure, railways, perimeter security and safety monitoring of urban underground structures. However, in practical applications, due to the one-dimensional spatiotemporal distribution characteristics of the DAS signal, and the problems of low signal-to-noise ratio and sparse event categories, the traditional detection method is difficult to maintain high accuracy while considering real-time performance and generalization ability.

[0003] In recent years, with the continuous development of artificial intelligence technology, security threat event detection technologies based on deep learning have also emerged. A large number of studies have shown that the model constructed based on the deep learning method can achieve good performance in security threat event detection problems. SUMMARY

[0004] The technical problem to be solved by the present application is to provide a signal detection and processing method suitable for distributed optical fiber acoustic sensing, which dynamically integrates global spectral features and local spatial details based on a weighted frequency domain neural operator and an adaptive sparse attention mechanism, so as to realize sub-meter spatial positioning and multi-class recognition of vibration / acoustic events in a strong noise environment.

[0005] The technical scheme of the present application is as follows:

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

[0007] (1) Collecting distributed optical fiber acoustic sensing signals of different categories through a distributed optical fiber sensing data acquisition platform, and constructing an original sensing data set after preprocessing;

[0008] (2) constructing a DAS signal detection and processing network, the sensing signal detection and processing network comprising a DAS signal feature extraction network, a gated DAS signal feature fusion network, and a DAS signal event category and positioning identification network;

[0009] The DAS signal feature extraction network adopts a diffusion frequency domain neural operator to refine a low-quality signal into a high-quality signal, and capture a complex conditional distribution of the high-quality signal under a low-quality input;

[0010] The gated DAS signal feature fusion network is used to integrate the global dependency relationship captured by the weighted frequency domain neural operator and the local information provided by the adaptive sparse attention module through spatial gating;

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

[0012] (3) constructing a loss function to train the DAS signal detection and processing network, inputting the original sensing signal to be detected in the original sensing data set into the trained DAS signal detection and processing network for detection, and obtaining the classification and positioning results of the event.

[0013] The gated DAS signal feature fusion network comprises a weighted frequency domain neural operator, an adaptive sparse attention module and a gated fusion module, the weighted frequency domain neural operator models the global dependency relationship through spectral convolution, effectively capturing long-range interaction and overall structure, the adaptive sparse attention module adjusts the attention weight in the attention mechanism by using a soft threshold function to obtain the attention-enhanced feature, and the gated fusion module is used to gate and fuse the features processed by the adaptive sparse attention module and the features processed by the weighted frequency domain neural operator to obtain the fused features.

[0014] The DAS signal feature extraction network is based on a diffusion frequency domain neural operator, the diffusion frequency domain neural operator comprising a forward diffusion process, a reverse diffusion process and a weighted frequency domain neural operator, the forward diffusion process being set in continuous time to simulate down-sampling, The forward diffusion process is represented by a variance-preserving stochastic differential equation, and the specific formula (1) is as follows:

[0015]

[0016] In formula (2), x t represents the input high-quality DAS signal at time t; dx t represents the output DAS signal after forward diffusion at time t; represents a down-sampling operator; θ(t) represents a noise sequence; and dW represents a standard Wiener process; is a drift term, used to simulate the process of high-frequency details loss in the down-sampling process; is a diffusion term, used to add Gaussian noise; θ max represents the maximum noise intensity, θ min represents the minimum noise intensity;

[0017] The weighted frequency domain neural operator is applied to each diffusion time step in the forward diffusion process;

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

[0019] The learning function of the weighted frequency domain neural operator in the weighted frequency domain neural operator and the diffusion frequency domain neural operator in the gated DAS signal feature fusion network is as follows (2):

[0020] x h (p)=Θ[x l (p)] (2);

[0021] In formula (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, and the specific formula is as follows (3):

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

[0024] In formula (3), w i (p) represents the feature representation of the i-th layer at the spatial position p; w i+1 (p) represents the feature representation of the i+1-th layer after updating at the spatial position p; X represents a linear transformation; Γ i represents the integral operator of the i-th layer, used to convert the feature to the frequency domain, and the specific conversion process is as follows (4); Σ() represents a nonlinear activation function;

[0025]

[0026] In formula (4), x represents the inverse frequency domain transformation; ω i (γ) represents a weight coefficient; λ iβi represents a learnable scalar parameter of the i-th layer, used to control the strength of the weight coefficient; β represents a hyperparameter, used to determine how the weight coefficient is scaled with the frequency amplitude ||γ||; F i (γ) is a complex tensor composed of learnable parameters, representing a frequency domain filter; w represents the i-th layer feature i The frequency domain transform takes values at frequency γ.

[0027] The adaptive sparse attention module includes a local exploration layer, an adaptive sparse attention layer, and a feature refining layer. The local exploration layer is composed of multiple consecutive convolution residual blocks, i.e., the output of a convolution residual block is used as the input of the next convolution residual block, and the feature output by the last convolution residual block is used as the output of the local exploration layer.

[0028] The feature x i in the adaptive sparse attention layer i is represented by the following formula (5):

[0029]

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

[0031]

[0032] In formula (6), represents the j-th component of the similarity vector , and the calculation formula of the similarity vector is shown in the following formula (7): is a soft threshold function, which satisfies the following formula (8):

[0033]

[0034] In formula (7), η q () and η k () represent a query vector generation function and a key vector generation function, respectively, both of which adopt a 1×1 convolution operation, T represents transposition, and δ(x i , x j ) represents features x i and x jDot product similarity between them;

[0035] To derive the soft threshold function First, sort the similarity vector And make the sorted Satisfy Let Ω = {1, …, L}, and define M as follows (9):

[0036]

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

[0038] Then, the soft threshold function Derived from equation (8) and equation (9), see equation (10) below:

[0039]

[0040] According to equation (10), the threshold result of the soft threshold function After calculating the threshold result of the soft threshold function

[0041] The feature refining layer is processed by 7x1 convolution operation, and is connected with the input in residual, to obtain the attention enhanced feature.

[0042] The processing process of the gating fusion module is: first, calculate the gating vector See equation (11) below:

[0043]

[0044] In equation (11), Represents the feature vector corresponding to the output high-quality DAS signal after weighted frequency domain neural operator processing; Represents the feature vector corresponding to the output high-quality DAS signal after adaptive sparse attention module processing; Θ[x] and X in and are the feature vectors corresponding to the output high-quality DAS signal of the DAS signal feature extraction network; [,] represents splicing along the channel dimension; Conv 1×1 Represents a 1x1 convolution layer, which is used to reduce the spliced features to a single channel gating vector; ∑() represents a nonlinear activation function;

[0045] Then, the element-level weighted sum of the two feature vectors is calculated, as shown in equation (12) below:

[0046]

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

[0048] The DAS signal event category and positioning identification network comprises a positioning regression branch and a classification branch, which respectively obtain the classification and positioning results of the event.

[0049] The positioning regression branch comprises three consecutive convolution residual blocks, an IoU loss, and a distribution focal loss DFL, the IoU loss and the distribution focal loss DFL are respectively based on the positioning result output by the last convolution residual block to calculate the loss, the calculation of the IoU loss is shown in formula (13), and the calculation of the distribution focal loss DFL is shown in formula (14).

[0050]

[0051] In formula (13), I pre represents the predicted interval of positioning, I gt represents the real interval corresponding to the event; and Distance() represents the length value of the interval.

[0052]

[0053] In formula (14), y di represents the real distance label, p di represents the predicted probability.

[0054] The classification branch comprises a first convolution residual block, a second convolution residual block, a third convolution residual block, a first depth separable convolution block, a second depth separable convolution block, and a binary cross-entropy loss BCE, the output of the first depth separable convolution block is taken as the input of the first convolution residual block, the output of the first convolution residual block is taken as the input of the second depth separable convolution block, the output of the second depth separable convolution block is taken as the input of the second convolution residual block, the output of the second convolution residual block is taken as the input of the third convolution residual block, the third convolution residual block outputs the classification result, and the binary cross-entropy loss BCE is based on the classification result output by the third convolution residual block to calculate the loss, the calculation of the binary cross-entropy loss BCE is shown in formula (15):

[0055]

[0056] In formula (15), N represents the number of samples, y ci represents the real classification label of sample i, p ci is the probability that sample i is predicted as a positive sample.

[0057] The loss function for training of the DAS signal detection and processing network is obtained by weighted summation of an IoU loss, a distribution focal loss DFL and a binary cross-entropy loss BCE, the IoU loss and the distribution focal loss DFL are used for a positioning regression branch, and the binary cross-entropy loss BCE is used for a classification branch.

[0058] Advantages of the present application:

[0059] (1) The neural operator is a high-efficiency neural operator variant, which aims to learn the mapping relationship between function spaces, and directly acts on input data of any quality, and realizes signal quality reconstruction processing by mapping low-quality input to high-quality output. Specifically, it first converts the input data to the frequency domain, applies the learned filter, and then converts the data back to the spatial domain, significantly improving the computational efficiency and accuracy through frequency domain convolution and mode truncation. Therefore, the neural operator is very suitable for the DAS vibration signal resolution task, has inherent resolution invariance, and can efficiently model global dependencies. However, mode truncation can underestimate the high-frequency components that are crucial for quality reconstruction of real-world vibration signals. In order to solve this limitation, the present application uses a weighted frequency domain neural operator to capture global dependencies, effectively captures key frequency components, significantly improves the reconstruction effect of high-frequency details that are crucial for signal quality reconstruction tasks, enhances the ability to capture key frequency components of weak events, and significantly improves the signal-to-noise ratio, achieving a double breakthrough in accuracy and computational efficiency.

[0060] (2) The neural operator is very suitable for the DAS vibration signal resolution task, has inherent resolution invariance, and can efficiently model global dependencies; the diffusion model can iteratively refine low-quality signals into high-quality signals, and capture the complex conditional distribution of high-quality signals under low-quality input. The present application combines the above two frameworks to construct a diffusion frequency domain neural operator for feature extraction of DAS signals, applies the weighted frequency domain neural operator to each diffusion time step in the forward diffusion process, converts the current degraded signal in the forward diffusion process to the frequency domain, applies frequency domain convolution and a learnable frequency domain filter to enhance high-frequency information, and then inversely transforms back to the space-time domain. And through the learnable weight coefficient, the high-frequency components are given greater weight, so as to enhance the key details of the signal, so that the diffusion model can refine low-quality signals into high-quality signals, and capture the complex conditional distribution of high-quality signals under low-quality input.

[0061] (3), the adaptive sparse attention module of the application is aimed at the defect that the soft maximum transformation in the traditional attention mechanism is sensitive to irrelevant information, the soft threshold operation is performed on the attention weight to sparsify the attention weight, suppress the low similarity position, and only keep the information of the high similarity area, while keeping the basic properties of the soft maximum transformation, so that all probability values are greater than zero and the sum is one, and higher weights are assigned to the feature vectors with higher similarity.

[0062] (4), the gating DAS signal feature fusion network of the application is used for integrating the global dependence relationship captured by the weighted frequency domain neural operator through spatial gating with the local information provided by the adaptive sparse attention module, the weighted frequency domain neural operator models the global dependence relationship through spectral convolution, effectively captures long-range interaction and overall structure, the adaptive sparse attention module quickly captures local dependence relationship and fine-grained details through the attention mechanism, and the gating fusion mechanism adaptively combines the feature maps of the weighted frequency domain neural operator and the adaptive sparse attention module, dynamically balances global and local information, and thus improves the quality of high DAS signal reconstruction.

[0063] (5), the DAS signal event category and positioning identification network of the application includes two branches, which are respectively used for obtaining the final event classification result and positioning result, and different loss functions are used for training, so that the sub-meter spatial positioning and multi-category detection and identification of vibration / acoustic events in a strong noise environment are realized. BRIEF DESCRIPTION OF DRAWINGS

[0064] Figure 1 is a flowchart of the application.

[0065] Figure 2 is a structural schematic diagram of the DAS signal detection and processing network of the application.

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

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

[0068] Figure 5 is a network architecture diagram of the DAS signal event category and positioning identification network of the application.

[0069] Figure 6 is a network architecture diagram of the convolution residual block of the application. DETAILED DESCRIPTION

[0070] With reference to the drawings of the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of the present application.

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

[0072] (1) Collecting different types of distributed optical fiber acoustic sensing signals through a distributed optical fiber sensing data acquisition platform, specifically including DAS signals of events such as mechanical excavation, manual hoeing, vehicle driving, etc., and after preprocessing, obtaining an original sensing data set, and dividing it into a training data set and a test data set;

[0073] (2) Constructing a DAS signal detection and processing network, the sensing signal detection and processing network including a DAS signal feature extraction network, a gated DAS signal feature fusion network, and a DAS signal event category and positioning identification network;

[0074] The DAS signal feature extraction network refines low-quality signals into high-quality signals using a diffusion frequency domain neural operator, which 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, obtaining attention-enhanced features. The gated fusion module is used to gate and fuse the features processed by the adaptive sparse attention module and the features processed by the weighted frequency domain neural operator, obtaining 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 positioning identification network is used to obtain the final event classification result and positioning result;

[0077] (3) Constructing a loss function, inputting the training data set into the DAS signal detection and processing network for training, inputting the test data set into the trained DAS signal detection and processing network for detection test, and obtaining the classification and positioning results of the event.

[0078] wherein the forward diffusion process in the diffusion frequency domain neural operator is set in continuous time the up-sampling is simulated, The forward diffusion process is represented by a variance-preserving stochastic differential equation, as shown in the following formula (1):

[0079]

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

[0081] The weighted frequency domain neural operator is 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 reversely diffuse the low-quality DAS signal input into the DAS signal feature extraction network into a high-quality DAS signal.

[0083] The learning function of the weighted frequency domain neural operator in the weighted frequency domain neural operator and the diffusion frequency domain neural operator in the gated DAS signal feature fusion network is shown in the following formula (2):

[0084] x h (p) = Θ[x l (p)] (2).

[0085] In formula (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. represent spatial coordinates;

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

[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 position p; w i+1 (p) represents the updated feature representation of the i+1-th layer at spatial position p; X represents a linear transformation; Γ i represents the integral operator of the i-th layer, which is used to convert the feature to the frequency domain, and the specific conversion process is shown in equation (4) below; ∑() represents a nonlinear activation function.

[0089]

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

[0091] As shown in Figure 3 , the adaptive sparse attention module includes a local exploration layer, an adaptive sparse attention layer, and a feature refining layer.

[0092] The local exploration layer is composed of three consecutive convolution residual blocks CRB, that is, the output of the previous convolution residual block CRB is taken as the input of the next convolution residual block CRB, and the feature output by the last convolution residual block CRB is taken as the output of the local exploration layer.

[0093] As shown in Figure 6 , each convolution residual block includes two 7x1 convolutions, a ReLU function, and a residual block. The output of the first 7x1 convolution is input to the second 7x1 convolution after ReLU function normalization processing, and the output of the second 7x1 convolution is output after residual connection with the input.

[0094] As shown in Figure 4 , in the adaptive sparse attention layer, the response Φ(x i is calculated by the following equation:i ) is represented as the following formula (5):

[0095]

[0096] In formula (5), the input feature of the adaptive sparse attention layer is X ∈ R L×C , x j is the jth feature on X; L represents the feature vector length of the input feature X; η v () represents the generation function of the value vector V, which adopts 1 × 1 convolution operation; the jth feature x j is allocated through a soft threshold operation, is the jth component of the soft threshold operation result, and the calculation process is shown in the following formula (6):

[0097]

[0098] In formula (6), represents the jth component of the similarity vector , and the calculation formula of the similarity vector is shown in the following formula (7); is a soft threshold function, which satisfies the following formula (8);

[0099]

[0100] In formula (7), η q () and η k () represent the generation functions of the query vector Q and the key vector K respectively, both of which adopt 1 × 1 convolution operation, T represents transposition, and δ (x i , x j ) represents the dot product similarity between features x i and x j .

[0101] In order to derive the soft threshold function , first, the similarity vector is sorted, and the sorted satisfies Let Ω = {1, …, L}, and define M as the following formula (9):

[0102]

[0103] In formula (9), M represents the maximum index satisfying the above condition, so 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 by zero;

[0104] Then, the soft threshold function The soft threshold function is calculated according to formula (10)

[0105]

[0106] The soft threshold function is calculated according to formula (10) After the threshold result of the soft threshold function is calculated according to formula (10), the similarity vector components lower than the threshold result are replaced by zero, and the remaining similarity vector components are replaced by the difference minus the threshold result, so as to realize threshold offsetting according to formula (8).

[0107] The feature refining layer is processed by 7x1 convolution operation, and is connected with the input in residual connection to obtain the attention-enhanced features.

[0108] The processing process of the gate fusion module is as follows: first, the gate vector is calculated Specifically, formula (11) is as follows:

[0109]

[0110] In formula (11), represents the feature vector corresponding to the output high-quality DAS signal after the weighted frequency domain neural operator processing; represents the feature vector corresponding to the output high-quality DAS signal after the adaptive sparse attention module processing; and x in and x in are the feature vectors corresponding to the output high-quality DAS signal of the DAS signal feature extraction network; [,] represents splicing along the channel dimension; Conv 1×1 represents a 1x1 convolution layer, which is used to reduce the spliced features to a single-channel gate vector; and

[0111] Then, the element-level weighted sum of the two feature vectors is calculated, and formula (12) is as follows:

[0112]

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

[0114] As shown in Figure 5 The DAS signal event category and positioning recognition network includes a positioning regression branch and a classification branch, which respectively obtain the classification and positioning results of the event.

[0115] The positioning regression branch includes three consecutive convolution residual blocks CRB (see Figure 6IoU loss and distribution focal loss DFL, the IoU loss and the distribution focal loss DFL are based on the last convolution residual block CRB (the output of the positioning result is used for loss calculation, the calculation of the IoU loss is shown in formula (13), and the calculation of the distribution focal loss DFL is shown in formula (14);

[0116]

[0117] In formula (13), I pre represents the predicted interval of positioning, I gt represents the real interval corresponding to the event; and Diistance() represents the length value of the interval.

[0118]

[0119] In formula (14), y di represents the real distance label, p di represents the predicted probability.

[0120] The classification branch includes a first convolution residual block CRB, a second convolution residual block CRB, a third convolution residual block CRB, a first depth separable convolution block DWConv, a second depth separable convolution block DWConv, and a binary cross-entropy loss BCE, the structures of the first convolution residual block CRB, the second convolution residual block CRB, and the third convolution residual block CRB are consistent, and are shown in Figure 6 The output of the first depth separable convolution block DWConv is used as the input of the first convolution residual block CRB, the output of the first convolution residual block CRB is used as the input of the second depth separable convolution block DWConv, the output of the second depth separable convolution block DWConv is used as the input of the second convolution residual block CRB, the output of the second convolution residual block CRB is used as the input of the third convolution residual block CRB, the third convolution residual block CRB outputs a classification result, and the binary cross-entropy loss BCE is based on the classification result output by the third convolution residual block to calculate the loss, and the calculation of the binary cross-entropy loss BCE is shown in formula (15):

[0121]

[0122] In formula (15), N represents the number of samples, y ci represents the real classification label of sample i, p ci 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 a supervised loss obtained by weighted summation of the IoU loss, the distribution focal loss DFL, and the binary cross-entropy loss BCE IoU loss and distribution focal loss DFL are used for the positioning regression branch, and binary cross-entropy loss BCE is used for the classification branch; in the training stage, the DAS signal with real labels is input into the DAS signal detection and processing network after being weakly enhanced, supervised training is performed, and supervised loss is calculated

[0124]

[0125] In formula (16), x i represents the DAS signal data of sample i, y i represents the real label of sample i, represents the classification loss (binary cross-entropy loss) of the DAS event classification label, represents the positioning regression loss (including IoU loss and distribution focal loss DFL) of the DAS event interval label.

[0126] Although the embodiments of the present application have been shown and described, it can be understood by those of ordinary skill in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application 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 is based on the diffusion frequency domain neural operator, which includes forward diffusion, reverse diffusion, and weighted frequency domain neural operators. The forward diffusion process is set to continuous time... Up-simulation downsampling, The forward diffusion process is represented by a variance-preserving stochastic differential equation, as shown in equation (1) below: （1）； In equation (1), This represents the high-quality DAS signal input at time t; This represents the DAS signal output after forward diffusion at time t; Represents the downsampling operator; Represents a noise sequence; 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; Represents the maximum noise intensity. 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. 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 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 shown in the following equation (2): （2）； In equation (2), High-quality DAS signal output by the DAS signal feature extraction network. This represents the high-quality DAS signal output after weighted frequency domain neural operator processing. Represents a weighted frequency domain neural operator; The weighted frequency domain neural operator This can be represented by stacking, as shown in equation (3) below: （3）； In equation (3), Representing the Layer in spatial location Feature representation of the location; Representing the After the layer is updated, its spatial location Feature representation of the location; Represents a linear transformation; Indicates the first The layer integral operator is used to transform the features to the frequency domain. The specific transformation process is shown in the following equation (4). Represents a non-linear activation function; （4）； In equation (4), Represents inverse frequency domain transform; Represents the weighting coefficient; Representing the The layer has learnable scalar parameters used to control the strength of the weight coefficients; This represents a hyperparameter used to determine how the weighting coefficients change with frequency amplitude. Scaling; It is a complex tensor composed of learnable parameters, representing a frequency domain filter; Representing the Layer features In frequency Frequency domain transformation of the value at that location.

4. The signal detection and processing method for distributed fiber optic acoustic sensing according to claim 3, 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. Features in adaptive sparse attention layer response It can be expressed as the following formula (5): （5）； In equation (5), the input features of the adaptive sparse attention layer are: , yes The first One feature; Representative input features The length of the feature vector; The representative value vector generating function uses a 1×1 convolution operation; the... Features Allocation is performed using a soft threshold operation. It is the first result of the soft threshold calculation Each component is calculated as shown in the following formula (6): （6）； In equation (6), Represents similarity vectors The Each component, similarity vector The calculation formula is shown in the following formula (7); It is a soft threshold function that satisfies the following equation (8); （7）； （8）； In equation (7), and These represent the query vector generation function and the key vector generation function, respectively, both using a 1×1 convolution operation. Represents transpose. Representative characteristics and characteristics Dot product similarity between them; To derive the soft threshold function First, for similarity vectors Sort, and make the sorted satisfy ;make and define The following formula (9) is given: （9）； In equation (9), This represents the largest 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: （10）； 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 similarity vector components, 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.

5. The signal detection and processing method for distributed fiber optic acoustic sensing according to claim 4, 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: （11）； In equation (11), , representing the feature vector corresponding to the high-quality DAS signal output after processing by the weighted frequency domain neural operator; , representing the feature vector corresponding to the high-quality DAS signal output after processing by the adaptive sparse attention module; and In All of these are feature vectors corresponding to high-quality DAS signals output by the DAS signal feature extraction network. This represents splicing along the channel dimension; This represents a 1×1 convolutional layer, used to reduce the concatenated features into a single-channel gated vector; Represents a non-linear activation function; Then, the two eigenvectors are summed element-wise with weights, as shown in equation (12) below: （12）； In equation (12), This represents the fused feature vector. This represents element-wise multiplication.

6. 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 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 DFL is shown in Equation (14). （13）； In equation (13), The prediction interval representing the location. The actual interval corresponding to the event; This represents the length of the interval. （14）； In equation (14), Represents the actual distance label. 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): （15）； In equation (15), Represents the number of samples. Representative sample Real category tags, It is a sample The probability of predicting a positive sample.

7. The signal detection and processing method for distributed fiber optic acoustic sensing according to claim 6, 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