Space-time diagram learning optical fiber vibration signal anomaly detection method of space-time interference mechanism

Abstract

The invention relates to the field of optical fiber vibration signal processing, and discloses a space-time diagram learning optical fiber vibration signal anomaly detection method of a space-time interference mechanism, which comprises the following steps: S1, acquiring an optical fiber vibration signal through an optical fiber vibration sensor, and converting the optical fiber vibration signal into discrete space-time data; s2, a space-time diagram is constructed based on the space-time data, nodes represent vibration signals of different space positions or time points, and edges represent space-time dependency relationships among the nodes; s3, applying an adaptive space-time interference mechanism, performing random disturbance on nodes or edges of the space-time diagram, and simulating space-time change in the environment; and S4, introducing an embedding method to maintain the relative positions of the nodes. Through the adaptive space-time interference mechanism and the space-time diagram convolutional network, the robustness and adaptability of the fiber vibration signal anomaly detection in a dynamic environment are enhanced, the scenes of node loss, edge weight fluctuation, time sequence abrupt change and the like can be effectively coped with, and high detection precision is still kept under complex noise.

Description

Technical Field

[0001] The present invention relates to the field of optical fiber vibration signal processing, and in particular to an optical fiber vibration signal anomaly detection method based on spatiotemporal graph learning of a spatiotemporal interference mechanism. Background Art

[0002] Fiber optic vibration signal anomaly detection is an important technology in the fields of structural health monitoring and security monitoring. Relying on the high sensitivity of fiber optic sensing technology, it can monitor tiny vibration signals in real time. However, in actual scenarios, how to accurately identify abnormal signals from massive spatiotemporal data remains a major technical challenge. Traditional anomaly detection methods usually rely on time series analysis or Fourier transform, which show certain effectiveness in static environments, but have poor adaptability to dynamic environments, especially when the spatiotemporal distribution changes, the detection performance decreases significantly. At the same time, traditional methods find it difficult to capture the complex dependencies of vibration signals in spatial and temporal dimensions, which limits their application in large-scale dynamic scenarios.

[0003] In recent years, with the rapid development of graph representation learning, spatiotemporal graph convolutional networks have been gradually applied to optical fiber vibration signal anomaly detection. By constructing vibration signals as graph structured data, spatiotemporal graph convolutional networks can effectively model the spatiotemporal dependencies between nodes. However, existing spatiotemporal graph learning methods usually assume that the training data and test data come from the same fixed environment. In dynamic spatiotemporal scenarios, this assumption cannot cope with changes in spatial structure or interference from temporal noise, and exhibits poor robustness and generalization capabilities. In addition, problems such as sensor node loss and time series mutation in real scenarios further aggravate the performance degradation of existing methods in noisy environments.

[0004] To solve these problems, the present invention proposes a method for detecting anomaly of optical fiber vibration signals by using spatiotemporal graph learning based on spatiotemporal interference mechanism. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention provides a method for detecting anomalies of optical fiber vibration signals by spatiotemporal graph learning based on spatiotemporal interference mechanism, which solves the problems of insufficient robustness and generalization ability of anomaly detection of optical fiber vibration signals in a dynamic spatiotemporal environment due to noise and distribution changes.

[0006] To achieve the above objectives, the present invention is implemented by the following technical scheme: a method for detecting abnormality of optical fiber vibration signals by learning a spatiotemporal graph of a spatiotemporal interference mechanism, comprising the following steps:

[0007] S1, collecting optical fiber vibration signals through optical fiber vibration sensors and converting them into discrete spatiotemporal data;

[0008] S2. constructing a spatiotemporal graph based on the spatiotemporal data, wherein nodes represent vibration signals at different spatial positions or time points, and edges represent spatiotemporal dependencies between nodes;

[0009] S3, applying an adaptive spatiotemporal interference mechanism to randomly perturb the nodes or edges of the spatiotemporal graph to simulate spatiotemporal changes in the environment;

[0010] S4, introduce embedding method to maintain the relative position of nodes;

[0011] S5. Using a graph convolutional network to extract features from the space-time graph after interference, extracting the space-time features of the vibration signal;

[0012] S6. Implement anomaly detection of optical fiber vibration signals based on the extracted features.

[0013] Preferably, the optical fiber vibration signal in step S1 is collected using Φ-OTDR technology, and the collected signal includes vibration position data in the spatial dimension and signal amplitude data in the time dimension.

[0014] Preferably, the step S1 is discretized in the following manner to convert it into discrete spatiotemporal data:

[0015] Spatial discretization: The fiber optic sensors are arranged as discrete sampling points along the transmission path. The fiber length L is divided into N equidistant spatial intervals, each of which represents a spatial node V. i , using the light pulse delay to determine the spatial position d i ;

[0016] Time discretization: according to the sampling frequency f s The vibration signal is divided into T time steps to form an N×T space-time matrix.

[0017] Preferably, constructing the space-time graph in step S2 includes:

[0018] S2.1. Define each node as a data point v in the space-time matrix i,k , indicating the i-th spatial position at time step t k The vibration signal value of

[0019] S2.2. Construct spatial edges and temporal edges, where:

[0020] Spatial edges represent the physical distance correlation between adjacent nodes;

[0021] Temporal edges represent the signal correlation of the same node at consecutive time steps.

[0022] Preferably, the interference process of the adaptive spatiotemporal interference mechanism in step S3 includes the following steps:

[0023] S3.1, apply random noise to the nodes of the space-time graph;

[0024] S3.2. Apply random perturbations to the edges of the space-time graph.

[0025] Preferably, the following formula is used for noise disturbance in step S3.1:

[0026]

[0027] Among them, J is the original node feature matrix, α is the interference intensity coefficient, σ 2 is the noise variance, ∈ is the standard normal distribution noise;

[0028] In step S3.2, the adjacency matrix is ​​adjusted using the following formula:

[0029] A′=AM,M ij ~B(p)

[0030] Where M is a binary matrix with the same dimension as A, representing the random masking or adjustment of edges, B(p) is a random variable generated from a Bernoulli distribution, and p is the probability of edge retention.

[0031] Preferably, the feature extraction of the graph convolution network in step S5 mainly uses graph convolution to extract features on the spatiotemporal graph structure, and aggregates the features of the node neighborhood through the following formula:

[0032]

[0033] in, is the node feature of the lth layer, σ is the activation function, W (l) is the convolution weight matrix, c ij is the normalization factor between adjacent nodes.

[0034] Preferably, the abnormality detection in step S6 is completed by the following steps:

[0035] S6.1. Using the extracted high-dimensional features Perform signal prediction;

[0036] S6.2. Determine whether an abnormality occurs based on the error calculation formula;

[0037] S6.3, when e t >∈, an abnormal alarm is triggered, where ∈ is the preset abnormal threshold.

[0038] Preferably, the error calculation formula in step S6.2 is as follows:

[0039] e t =|y t -y′ t|

[0040] Among them, y t is the actual value, y′ t is the predicted value.

[0041] Preferably, the S4 step specifically includes:

[0042] The embedding feature vector f of each node i (u) is defined by the following formula:

[0043] f i (u) = α i ·d V (u,V i )

[0044] Among them, α i To control the parameters of the embedding distribution, d V (u,V i ) is the distance from node u to anchor point V i distance.

[0045] The present invention provides a method for detecting abnormality of optical fiber vibration signals by using a spatiotemporal graph learning method based on a spatiotemporal interference mechanism.

[0046] It has the following beneficial effects:

[0047] 1. By introducing an adaptive spatiotemporal interference mechanism and a spatiotemporal graph convolutional network, the present invention can model and learn the dynamic spatiotemporal distribution of optical fiber vibration signals. The spatiotemporal graph convolutional network uses a graph structure to capture the complex spatiotemporal dependencies between nodes, and combines the adaptive spatiotemporal interference mechanism to simulate spatiotemporal disturbances, which significantly enhances the robustness of the model to dynamic environments, enabling it to maintain a high anomaly detection accuracy in scenarios with spatiotemporal changes and noise.

[0048] 2. The present invention applies spatial interference and temporal interference to simulate scenarios such as node loss, edge weight fluctuation, and time series mutation, thereby enhancing the adaptability to environmental changes during model training. The spatial interference matrix and the temporal interference matrix adjust the spatial connection and time series respectively, which significantly improves the robustness of spatiotemporal graph learning, enabling the algorithm to adapt to different dynamic environments and meet a wide range of application needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is the overall structure diagram of the present invention;

[0050] Figure 2 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the specification of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] Please see attached Figure 1 -Attached Figure 2 The embodiment of the present invention provides a method for detecting abnormal optical fiber vibration signals by using a spatiotemporal graph learning method based on a spatiotemporal interference mechanism, comprising the following steps:

[0053] S1, collecting optical fiber vibration signals through optical fiber vibration sensors and converting them into discrete spatiotemporal data;

[0054] S2, constructing a spatiotemporal graph based on spatiotemporal data, where nodes represent vibration signals at different spatial locations or time points, and edges represent spatiotemporal dependencies between nodes;

[0055] S3, apply an adaptive spatiotemporal interference mechanism to randomly perturb the nodes or edges of the spatiotemporal graph to simulate the spatiotemporal changes in the environment;

[0056] S4, introduce embedding method to maintain the relative position of nodes;

[0057] S5. Using a graph convolutional network to extract features from the space-time graph after interference, extracting the space-time features of the vibration signal;

[0058] S6. Implement anomaly detection of optical fiber vibration signals based on the extracted features.

[0059] Specifically, the present invention provides a method for detecting anomalies of optical fiber vibration signals by spatiotemporal graph learning based on spatiotemporal interference mechanism, which can effectively meet the needs of detecting anomalies of optical fiber vibration signals in complex dynamic environments. By introducing spatiotemporal interference mechanism and embedding method, combined with graph convolutional network to extract spatiotemporal features, the robustness and detection accuracy of the model under spatiotemporal noise conditions are greatly improved.

[0060] The implementation of this method starts with the acquisition of optical fiber signal data. Vibration signals in dynamic environments are collected through optical fiber vibration sensors, and the signals are converted into time-space discrete data with Φ-OTDR technology as the core. Through the Φ-OTDR system, laser pulses are injected into the optical fiber and receive Rayleigh scattering signals to capture tiny vibration changes. To ensure the time-space resolution of the signal, the optical fiber sensors are evenly arranged as several sampling points along the monitoring area. These sampling points form discrete data in the spatial dimension, and the time series is segmented by a high-sampling-rate analog-to-digital conversion module to generate discrete data in the time dimension.

[0061] After organizing the collected discrete signals into a spatiotemporal matrix, a spatiotemporal graph is constructed to represent the spatiotemporal dependency of the signals. The spatiotemporal graph consists of nodes and edges, where each node represents a spatiotemporal point and stores the vibration signal value at the spatial position and time step. The nodes are connected by edges to reflect their spatial or temporal dependencies. The weight of the spatial edge is determined by the physical distance between the nodes, while the temporal edge is determined based on the correlation between the time steps, and the weights are calculated using exponential decay functions. The finally constructed spatiotemporal graph can effectively capture the complex spatiotemporal correlation structure and provide data support for the subsequent graph learning process.

[0062] In order to improve the adaptability of the model to dynamically changing scenarios, the present invention proposes an adaptive spatiotemporal interference mechanism. After the spatiotemporal graph is constructed, interference is applied to the nodes and edges of the graph to simulate random changes in the environment. For example, random noise is applied to node features to simulate sensor failure or signal loss; random perturbations are introduced to edge weights to reflect dynamic changes in spatial or temporal dependencies. These interference operations control the intensity of interference by setting specific parameters, so that the model can generate simulated scenarios of multiple dynamic environments during training, enhancing its robustness to different noise conditions.

[0063] In order to maintain the relative position relationship between nodes, an embedding method is further proposed. In the spatiotemporal graph, the embedded feature vector of each node is determined by calculating the distance from the node to the anchor point. The anchor point can be dynamically selected as the geometric center of the node or the central node generated based on clustering. This embedding method can not only capture the local spatial distribution characteristics, but also provide a global spatiotemporal constraint, effectively supplementing the uncertainty introduced by the interference mechanism.

[0064] On the basis of the above, the graph convolutional network is used to extract features from the space-time graph after interference. The graph convolutional network extracts high-order features in the space-time graph layer by layer by aggregating the feature information of the node neighborhood. In the first layer of convolution, the dependencies of the spatial dimension are captured; in the second layer of convolution, the dynamic features of the time series are extracted; in the third layer of convolution, the spatial and temporal features are fused to generate a high-dimensional feature representation of the node. Through the cascade operation of multiple layers of convolution, the model can gradually capture the complex space-time patterns contained in the optical fiber vibration signal.

[0065] Finally, anomaly detection is performed through the extracted high-dimensional features. Specifically, the model uses node features to predict the normal value of the vibration signal and compares the error with the actual value. After setting the threshold, if the error exceeds the threshold, it is judged as an anomaly and an alarm is triggered. Through comprehensive learning of spatiotemporal graph features, the model can effectively identify abnormal situations such as sensor failure and sudden noise.

[0066] In step S1, the optical fiber vibration signal is collected using the Φ-OTDR technology, and the collected signal includes vibration position data in the spatial dimension and signal amplitude data in the time dimension.

[0067] Step S1 is discretized into discrete spatiotemporal data by the following method:

[0068] Spatial discretization: The fiber optic sensors are arranged as discrete sampling points along the transmission path. The fiber length L is divided into N equidistant spatial intervals, each of which represents a spatial node V. i , using the light pulse delay to determine the spatial position d i ;

[0069] Time discretization: according to the sampling frequency f s The vibration signal is divided into T time steps to form an N×T space-time matrix.

[0070] Specifically, in view of the above content, the collection and discretization of optical fiber vibration signals are first generally described. The collection and processing of optical fiber vibration signals is the basic part of the space-time graph learning method of the entire space-time interference mechanism, which directly determines the effectiveness of subsequent space-time graph construction and model training. The Φ-OTDR technology is used in the present invention to provide a complete space-time signal structure by converting the optical fiber vibration signal into discrete space-time data, laying the foundation for the construction of the space-time graph. The discretization of the signal involves the processing of two dimensions, space and time, which can fully capture the space-time distribution characteristics of the vibration signal.

[0071] In this embodiment, the optical fiber vibration signal is collected using Φ-OTDR technology. Φ-OTDR is an optical fiber sensing technology based on light pulses and Rayleigh scattering effect, which can sensitively capture phase changes caused by tiny vibrations in the optical fiber. The vibration response data of the optical fiber can be obtained by propagating laser pulses along the optical fiber transmission path and receiving Rayleigh scattering signals.

[0072] The Φ-OTDR system consists of a narrow pulse laser, a photodetector, an analog-to-digital conversion module, and a data storage device. After the laser pulse is emitted into the optical fiber, the Rayleigh signal generated by the optical fiber scattering is transmitted back to the detector. The phase change in the signal reflects the vibration intensity and vibration position of the optical fiber. In this embodiment, the optical fiber sensors are evenly arranged along the monitoring area to achieve continuous monitoring of spatial vibration. The collected signal includes two dimensions:

[0073] Vibration position data in the spatial dimension: corresponds to a specific spatial position on the optical fiber, indicating where the vibration occurs;

[0074] Signal amplitude data in the time dimension: corresponds to the change in vibration intensity in the time series.

[0075] In order to facilitate the subsequent construction of space-time graphs and model learning, the continuous fiber vibration signal needs to be converted into discrete space-time data. The discretization process is divided into two steps: spatial discretization and temporal discretization.

[0076] In this embodiment, the optical fiber sensors are evenly arranged as discrete sampling points along the transmission path, and the optical fiber length L is divided into N equidistant spatial intervals, each of which represents a spatial node V. i The node position d i It is determined by the light pulse delay and is calculated as:

[0077]

[0078] Among them, d i represents the spatial position of the ith node, L is the total length of the optical fiber, and N is the total number of sampling points.

[0079] The spatial resolution of the optical fiber is determined by the optical pulse width and the sampling rate of the photodetector. In order to ensure the spatial discretization accuracy of the vibration signal, the sampling interval Δd and the number of sampling points N are optimized in this embodiment to ensure that the nodes are evenly distributed and can cover all vibration positions in the monitoring area.

[0080] Time discretization is to divide the continuous time series signal into several time steps T, each time step represents the amplitude of the vibration signal within a fixed time interval. Set the sampling frequency to f s , then the time interval is Δ The calculation formula for the time step number T is:

[0081]

[0082] Among them, T total Indicates the total sampling duration of the vibration signal.

[0083] After discretization, we get an N×T space-time matrix X, whose element x i,k represents the i-th spatial position at time step t k The vibration signal amplitude is:

[0084]

[0085] Each row in the matrix represents a signal sequence of a spatial node changing over time, and each column represents the spatial vibration distribution at a certain time step. Through this discretization method, the vibration characteristics of space and time are completely preserved, providing a data basis for the subsequent construction of space-time graphs.

[0086] In order to ensure the integrity of data discretization, a noise filtering mechanism is introduced in this embodiment. During the signal acquisition process, it may be affected by sensor thermal noise or external environmental interference, so the following method is used for noise processing:

[0087] Baseline correction: eliminate the static background noise of the optical fiber itself;

[0088] Low-pass filtering: filter out high-frequency noise components in the signal and retain the effective vibration signal;

[0089] Average filtering: The signal of the spatial node is sampled multiple times at each time step to take the average value to reduce random errors.

[0090] Through the above steps, high-quality discrete space-time data are generated, which can accurately describe the spatial distribution and time evolution characteristics of the optical fiber vibration signal, laying the foundation for the next step of constructing a space-time map.

[0091] Through the high-precision acquisition of Φ-OTDR technology and the discretization method of space and time dimensions, this embodiment realizes the transformation of optical fiber vibration signals into spatiotemporal data. The discretized data not only retains the spatiotemporal distribution characteristics of the signal, but also enhances the data quality through noise processing. This process provides reliable data support for the subsequent construction of spatiotemporal graphs and the introduction of spatiotemporal interference mechanisms.

[0092] The construction of the spatiotemporal graph in step S2 includes:

[0093] S2.1. Define each node as a data point v in the space-time matrix i,k , indicating the i-th spatial position at time step t k The vibration signal value of

[0094] S2.2. Construct spatial edges and temporal edges, where:

[0095] Spatial edges represent the physical distance correlation between adjacent nodes;

[0096] Temporal edges represent the signal correlation of the same node at consecutive time steps.

[0097] Specifically, for the above content, constructing a space-time graph is to further structure the discrete space-time data to capture the complex space-time dependency of the vibration signal. The space-time graph can express the multi-dimensional connection between spatial nodes and time steps. By constructing a graph structure of space-time association, it lays the foundation for subsequent graph convolutional network learning. The following describes the construction method of the space-time graph in detail from two aspects: node definition and edge construction.

[0098] In step S2 of the present invention, by constructing a space-time graph, the discretized space-time matrix is ​​represented as a graph structure model, wherein each node represents the vibration signal value of a specific spatial position at a certain time step, and the edge connection between nodes is used to express the dependency in space or time. Specifically, a node is defined as a data point in the space-time matrix, and an edge is generated according to the dynamic correlation of spatial adjacency or time series, and is represented by a spatial edge and a temporal edge, respectively. The construction of the space-time graph can fully capture the changing characteristics of the vibration signal in the space-time dimension.

[0099] Node definition: mapping of spatiotemporal matrix to graph structure

[0100] In this embodiment, each node v i,k It is defined as a data point in the space-time matrix, specifically representing the i-th spatial position at time step t k The vibration signal value x i,k . Node v i,k The properties include the following:

[0101] Spatial attribute: indicates the spatial location of the vibration signal, corresponding to the sampling point position d of the optical fiber sensor i ;

[0102] Time attribute: represents the current time step t k , reflecting the temporal dynamics of the vibration signal;

[0103] Signal value attribute: amplitude x of the vibration signal i,k .

[0104] The mathematical definition of a node is:

[0105] v i,k =(d i ,t k ,x i,k )

[0106] Among them, d i is the spatial position corresponding to the node, which is determined by spatial discretization; t k is the time step, generated by time discretization; x i,k is the vibration signal amplitude.

[0107] Edge construction: spatial edges and temporal edges

[0108] The edge definition of the space-time graph is used to connect the dependencies between nodes, which includes two types: spatial edge and temporal edge:

[0109] (1) Spatial edge: The spatial edge connects adjacent spatial nodes and is used to represent the correlation of vibration signals in the spatial dimension. Whether a spatial edge is established between two nodes and its edge weight is determined by the physical distance d. ij Decide.

[0110] Edge generation rule: If node v i,k and v j,k Located at spatial position d i and d j , and the physical distance d ij Satisfy ij ≤τ s , then in v i,k and v j,k A spatial edge is established between them, where τs is the spatial adjacency threshold.

[0111] Edge weight calculation: weight w of spatial edge ij The distance relationship between nodes is reflected by the Gaussian attenuation function calculation. The formula is:

[0112]

[0113] Where: d ij =|d i -d j | represents node v i and v j The physical distance; β is the distance decay coefficient, which controls the rate of change of edge weight.

[0114] (2) Time edge: The time edge connects nodes at the same spatial position at consecutive time steps and is used to represent the dynamic dependency of vibration signals in the time dimension.

[0115] Edge generation rule: If node v i,k and v i,k+1 Indicates the same spatial position d i At adjacent time steps t k and t k+1 The vibration signal value is i,k and v i,k+1 Create a time edge between them.

[0116] Edge weight calculation: the weight w of the time edge kl Calculated by exponential decay function to reflect the correlation between time steps, the formula is:

[0117]

[0118] Among them, γ is the time correlation decay parameter; τ t is the temporal adjacency threshold, which indicates the condition for establishing a temporal edge.

[0119] Space-time graph representation: Through the construction of the above nodes and edges, this embodiment finally generates a complete space-time graph G = (V, E), where:

[0120] Node set V: contains all v i,k node;

[0121] Edge set E: includes all spatial edges and temporal edges that meet the requirements.

[0122] The adjacency matrix of the space-time graph can be divided into the spatial adjacency matrix A s and the time adjacency matrix A t The two parts are combined to form a complete spatiotemporal adjacency matrix A, expressed as: A = A s+A t

[0123] Among them A s and A t The specific calculation rules of are determined by the above formulas.

[0124] To ensure the integrity of the spatiotemporal graph, this embodiment cleans up noise nodes or isolated nodes and removes nodes that do not meet the adjacency conditions. In addition, the weights of the adjacency matrix A are processed using a normalization method to enhance the stability of the calculation. The normalization operation is as follows:

[0125]

[0126] Where D is the degree matrix of the node, defined as:

[0127]

[0128] Through node definition and the construction of spatial and temporal edges, this embodiment successfully transforms the space-time matrix into a space-time graph structure, fully capturing the complex correlation between optical fiber vibration signals in spatial and temporal dimensions. This graph structure provides a complete data representation for subsequent graph convolutional network learning, ensuring that the model can effectively extract space-time features for anomaly detection.

[0129] The interference process of the adaptive spatiotemporal interference mechanism in step S3 includes the following steps:

[0130] S3.1, apply random noise to the nodes of the space-time graph;

[0131] S3.2. Apply random perturbations to the edges of the space-time graph.

[0132] In step S3.1, the following formula is used for noise perturbation:

[0133]

[0134] Among them, J is the original node feature matrix, α is the interference intensity coefficient, σ 2 is the noise variance, ∈ is the standard normal distribution noise;

[0135] In step S3.2, the adjacency matrix is ​​adjusted using the following formula:

[0136] A′=AM,M ij ~B(p)

[0137] Where M is a binary matrix with the same dimension as A, representing the random masking or adjustment of edges, B(p) is a random variable generated from a Bernoulli distribution, and p is the probability of edge retention.

[0138] Specifically, in the spatiotemporal graph learning of the present invention, an adaptive spatiotemporal interference mechanism is proposed, which simulates the spatiotemporal changes that may occur in actual scenarios by randomly perturbing the node feature matrix and the adjacency matrix, including node signal fluctuations, loss, and edge shielding or adjustment. This interference mechanism can enhance the robustness of the model in a dynamic spatiotemporal environment, thereby improving the performance of the model in anomaly detection.

[0139] Specifically, node perturbation is achieved by adding random noise to node features, while edge perturbation is achieved by introducing a mask matrix to randomly mask or adjust the adjacency matrix. Furthermore, spatial interference and temporal interference use different mask matrices M s and M t , used to simulate dynamic changes in spatial and temporal dimensions.

[0140] In this embodiment, the node feature matrix J represents the attribute values ​​(such as vibration signal amplitude) of all nodes. In order to simulate the random fluctuation or loss of node signals in actual scenarios, random noise is applied to the node feature matrix.

[0141] The formula for node perturbation is defined as:

[0142]

[0143] Where: J is the original node feature matrix; J′ is the node feature matrix after the disturbance; α is the interference intensity parameter, which controls the noise amplitude; ∈ is the standard normal distribution N(0,σ 2 ) noise; σ 2 is the variance of the noise, which is used to adjust the randomness of the disturbance.

[0144] Implementation process:

[0145] A random noise matrix ∈ is generated according to the dimension of the node feature matrix J, where each element is sampled from a standard normal distribution.

[0146] Scale the noise matrix by a factor of α to get the final noise value.

[0147] The scaled noise matrix is ​​added to the original node feature matrix J to obtain the perturbed node feature matrix J′.

[0148] In order to simulate the random masking or adjustment of edges, edge weight perturbations are applied to the adjacency matrix A. The perturbation is achieved by introducing a mask matrix M.

[0149] The formula for edge weight perturbation is defined as:

[0150] A′=AM,M ij ~B(p)

[0151] Where: A is the original adjacency matrix; A′ is the adjacency matrix after perturbation; M is a randomly generated mask matrix with the same dimension as A; B(p) is a random variable generated by Bernoulli distribution, and the parameter p represents the probability of edge preservation.

[0152] Implementation process: Generate a random mask matrix M according to the dimension of the adjacency matrix A, where each element M ij The value of is determined by the Bernoulli distribution B(p).

[0153] The mask matrix M and the original adjacency matrix A are multiplied element by element to obtain the perturbed adjacency matrix A′.

[0154] In order to simulate the dynamic changes of spatial dimension and time dimension respectively, different mask matrices M are used. s and M t .

[0155] Spatial interference matrix M s : Used to simulate changes in spatial connectivity between nodes, such as the addition or loss of edges.

[0156] M s [i,j]=B(1,p(u))

[0157] Where: p(u) represents the probability of keeping edge (i, j), which depends on the spatial distance or connection strength; u is the spatial attribute between nodes (such as physical distance).

[0158] Generation rule: If M s [i,j]=1, indicating that the spatial edge (i,j) is retained;

[0159] If M s [i,j]=0, indicating that the spatial edge (i,j) is deleted.

[0160] Time interference matrix M t : Used to simulate sudden changes or signal loss in the time dimension.

[0161] M t [k,l]=B(1,q(v))

[0162] Where: q(v) represents the probability of keeping the time edge (k,l), which depends on the time step difference or signal dynamics; ν is the property of the time dimension (such as the time step difference).

[0163] Generation rule: If M t [k,l]=1, indicating that the time edge (k,l) is retained;

[0164] If M t [k,l]=0, indicating that the time edge (k,l) is shielded.

[0165] The adjacency matrix after perturbation: The spatial adjacency matrix after perturbation: A′ s =A s ⊙M s

[0166] After the temporal adjacency matrix is ​​perturbed: A′ t =A t ⊙M t

[0167] After integration, the final adjacency matrix is ​​formed: A′=A′ s +A′ t

[0168] To ensure the numerical stability of the perturbed adjacency matrix in graph learning, the perturbed adjacency matrix A′ is normalized:

[0169]

[0170] Where: D is the node degree matrix, and its diagonal elements are the rows and D of the adjacency matrix. ii =∑ j A′ ij ; is the normalized adjacency matrix.

[0171] Through the above interference mechanism, the random noise of nodes simulates the dynamic changes of signals, and the random perturbations of edges reflect the dynamic adjustments of spatial and temporal connections. s and M t The interference mechanism can flexibly adapt to the changing scenarios of time and space. The space-time graph G′=(V,E′) after interference is further normalized to ensure the numerical stability of the calculation, providing highly robust data support for subsequent model training.

[0172] The S4 step specifically includes:

[0173] The embedding feature vector f of each node i (u) is defined by the following formula:

[0174] f i (u) = α i ·d V (u,V i )

[0175] Among them, α i To control the parameters of the embedding distribution, d V (u,V i ) is the distance from node u to anchor point V i distance.

[0176] Specifically, in the spatiotemporal graph learning method of the present invention, an embedding mechanism is introduced to maintain the relative position relationship between nodes. The embedding method generates a feature vector for each node and quantitatively expresses the position of the node using the distance information between the node and the anchor point. This method can not only enhance the local representation ability of the node in the graph structure, but also provide additional spatial geometric constraints for the subsequent graph learning process.

[0177] In this embodiment, the embedded feature vector f of each node i (u) is defined by the following formula:

[0178] f i (u) = α i ·d V (u,V i )

[0179] Where: α i To control the parameters of the embedding distribution, d V (u,V i ) is the distance from node u to anchor point V i The distance between nodes reflects the spatial or geometric relationship between nodes.

[0180] Implementation of the embed method:

[0181] Anchor point selection: Anchor point V i It is one or more key positions in the graph structure, used to determine the relative positions between nodes. The methods for selecting anchor points include:

[0182] Geometric center: Select the center point in the graph structure as the anchor point. The geometric center can be obtained by calculating the average position of the nodes.

[0183] Cluster center: cluster the nodes and use the center point of each cluster as the anchor point;

[0184] Specify points: Specify specific nodes as anchor points based on prior knowledge.

[0185] Distance calculation: node u to anchor point V i The distance d V (u,V i ) is calculated as:

[0186] d V (u,V i )=||pos(u)-pos(V i )||

[0187] Among them: pos(u) and pos(V i ) are node u and anchor point V respectively iis the position vector of ; ‖·‖ is the Euclidean distance, which is used to measure the geometric distance between nodes.

[0188] Generation of embedded features:

[0189] The embedding feature f of each node u i (u) is determined by the influence factor α of the anchor point i and the distance d from the node to the anchor point V (u,V i ) decide jointly.

[0190] By adjusting the parameter α i , the contribution of different anchor points to the node embedding features can be controlled.

[0191] Mathematical properties of embedding:

[0192] Local constraints: The embedding method captures the local geometric characteristics of nodes through distance calculation, so that the embedded features of adjacent nodes have strong similarity.

[0193] Global expression: When multiple anchor points are selected, the final embedding feature of each node can be composed of a weighted combination of multiple anchor points, and the formula is:

[0194]

[0195] Where k is the number of anchor points.

[0196] Application of embedding methods in spatiotemporal graphs:

[0197] Node embedding combined with feature learning:

[0198] Embedding feature f i (u) can be combined with the original node feature J as the input feature of graph learning:

[0199] J′=J+f i (u)

[0200] This method can supplement spatial geometric information based on the original signal characteristics and enhance the graph convolutional network's understanding of the relative position relationship between nodes.

[0201] Robustness improvement: The embedding method dynamically adjusts the anchor point distribution and parameter α i , which can adapt to the graph learning needs in different spatial change scenarios and help improve the generalization ability of the model in complex environments.

[0202] By introducing embedded feature vectors for nodes, this embodiment realizes the quantitative expression of the relative positions between nodes. The embedding method combines anchor point selection and distance calculation to provide spatial geometric constraint information for spatiotemporal graph learning, which not only enhances the expression ability of node features, but also provides a richer data foundation for subsequent anomaly detection.

[0203] The feature extraction of the graph convolutional network in step S5 mainly uses graph convolution to extract features on the spatiotemporal graph structure, and aggregates the features of the node neighborhood through the following formula:

[0204]

[0205] in, is the node feature of the lth layer, σ is the activation function, W (l) is the convolution weight matrix, c ij is the normalization factor between adjacent nodes.

[0206] Specifically, in the spatiotemporal graph learning of the present invention, the graph convolutional network is a key part, which extracts the spatiotemporal features of nodes and their neighborhoods by performing convolution operations on the spatiotemporal graph structure. The core idea of ​​the graph convolutional network is to express the characteristics of each node and the characteristics of the neighboring nodes through feature aggregation, so as to extract high-order features layer by layer. In each layer of convolution, the feature update of the node is jointly acted by the activation function and the weight matrix, and the neighborhood information is normalized and aggregated using the adjacency matrix to ensure the stability of learning.

[0207] In this embodiment, the graph convolution operation implements feature aggregation of the node neighborhood through the following formula:

[0208]

[0209] in: is the feature representation of node i in the I+1th layer; σ is a nonlinear activation function (such as ReLU), which is used to introduce nonlinear characteristics; W (l) is the convolution weight matrix of the lth layer, representing the transformation parameters of the node features; is the feature representation of node j in the lth layer, where j is the neighbor node of node i; c ij is a normalization factor used to balance the feature contributions of different neighbor nodes to node i.

[0210] Specific implementation process:

[0211] Initialize node features: The initial features of the node are It can be the original feature of the node (such as signal value) or its embedded feature J;

[0212] The initial features are input into the first layer of graph convolutional network.

[0213] Neighborhood feature aggregation:

[0214] For each node i, the features of all its neighboring nodes To carry out polymerization;

[0215] The aggregation method is through the normalized version of the adjacency matrix A Expression, specifically:

[0216]

[0217] Where: A is the adjacency matrix; D is the node degree matrix, D ii =∑ j A ij ; Normalize the adjacency matrix to be symmetric and ensure numerical stability during aggregation.

[0218] Feature transformation: Neighborhood features are transformed through the weight matrix W (l) Linear transformation of , we get the potential representation of the node:

[0219]

[0220] Where: H (l) is the feature matrix of all nodes in the lth layer; Z (l) is the feature matrix after linear transformation. Activation function: (l) Apply the activation function σ to introduce nonlinear characteristics:

[0221] H (l+1) =σ(Z (l) )

[0222] Among them, the activation function can be ReLU, Sigmoid or Tanh.

[0223] In this embodiment, the graph convolutional network is composed of multiple layers of convolution. The output of each layer serves as the input of the next layer. By extracting high-order features of nodes layer by layer, the gradual aggregation of spatiotemporal features is achieved.

[0224] The first convolution layer:

[0225] Extract local features of nodes and aggregate feature information of direct neighbors;

[0226] The output is the node feature H after the first layer of convolution (1) .

[0227] The second convolution layer:

[0228] Extract the second-order features of the node, that is, the neighbor features of the neighbor nodes;

[0229] The output is the node feature H after the second layer of convolution (2) .

[0230] The third and further layers:

[0231] Continue to stack convolutional layers to extract high-order features and further expand the receptive field of the node;

[0232] Output the final node feature H (L) , as the final feature representation.

[0233] In order to enhance the numerical stability and training effect of the model, this embodiment normalizes the features of each convolution output. The normalization formula is:

[0234]

[0235] Where: μ is the mean of the feature matrix; σ is the standard deviation of the feature matrix.

[0236] The normalized features are used for subsequent anomaly detection or classification tasks.

[0237] Through the above graph convolution operation, this embodiment realizes the layer-by-layer aggregation extraction of spatiotemporal graph nodes and their neighborhood features. The final feature of the node not only contains its own signal information, but also integrates the spatiotemporal characteristics of the neighboring nodes, providing a high-dimensional and highly expressive feature representation for subsequent anomaly detection. The weight matrix W of each layer of convolution is (l) Adaptive learning of spatiotemporal features can be achieved through gradient descent optimization.

[0238] In step S6, anomaly detection is completed through the following steps:

[0239] S6.1. Using the extracted high-dimensional features Perform signal prediction;

[0240] S6.2. Determine whether an abnormality occurs based on the error calculation formula;

[0241] S6.3, when e t >∈, an abnormal alarm is triggered, where ∈ is the preset abnormal threshold.

[0242] The error calculation formula in step S6.2 is as follows:

[0243] e t =|y t -y′ t |

[0244] Among them, y t is the actual value, y′ t is the predicted value.

[0245] Specifically, in the final stage of spatiotemporal graph learning, abnormal behaviors in the optical fiber vibration signal are identified through anomaly detection. Anomaly detection predicts signals based on the extracted high-dimensional features, and determines whether anomalies exist by calculating the error between the predicted value and the actual value. If the error exceeds the preset threshold ∈, an abnormal alarm is triggered. The specific implementation includes three steps: signal prediction, error calculation, and anomaly determination.

[0246] In this embodiment, the high-dimensional features H of nodes extracted by graph convolutional network are used. (L) Predict the node signal. The signal prediction model can be a simple linear mapping, or it can be combined with a nonlinear activation function for complex feature learning. The prediction formula is:

[0247] y′ t =f(H (L) ,W p )

[0248] Where: y′ t is the predicted signal value of node t; f(·) is the prediction model; H (L) is the high-dimensional feature of the node; W p is the parameter matrix of the prediction model, which is obtained through training and learning.

[0249] Signal prediction process:

[0250] Input the high-dimensional features output by the graph convolutional network into the prediction model f(·);

[0251] By parameter W p transformation and mapping, and output the predicted signal value of the node.

[0252] After the signal prediction is completed, the predicted value y′ is calculated t With the actual value y t The error between and is used to determine whether the node is abnormal. The error calculation formula is:

[0253] e t =|y t -y′ t |

[0254] Where: e t is the error value of node t; y t is the actual signal value of node t; y′ t is the predicted signal value of node t.

[0255] The core of error calculation is to measure the deviation between the predicted result and the real signal. A large error may indicate that the node signal is abnormal.

[0256] Error calculation process:

[0257] Get the actual signal value y of the node t and the predicted signal value y′ t ;

[0258] Calculate the node error value e according to the error formula t .

[0259] After the error calculation is completed, a preset abnormal threshold ∈ is set to determine whether the node is abnormal. The abnormality judgment formula is:

[0260] If e t >∈, then alarm.

[0261] Where: ∈ is the preset abnormal threshold, reflecting the allowable signal error range; when e t >∈, it indicates that the node signal is abnormal and triggers the alarm mechanism.

[0262] Abnormal determination process:

[0263] Compare the error value e of each node t with threshold ∈;

[0264] If e t >∈, the node is marked as an abnormal node and an alarm is triggered.

[0265] In this embodiment, by extracting high-dimensional features and combining error calculation, the anomaly detection of optical fiber vibration signals is accurately achieved. The error calculation formula and the setting of the anomaly threshold provide flexibility for the method and can adapt to different application scenarios. The entire anomaly detection process is based on graph learning, combined with signal prediction and error evaluation, providing high-precision detection capabilities for optical fiber vibration monitoring.

[0266] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting abnormality of optical fiber vibration signals by spatiotemporal graph learning based on spatiotemporal interference mechanism, characterized in that: The following steps are involved: S1, collecting optical fiber vibration signals through optical fiber vibration sensors and converting them into discrete spatiotemporal data; S2. constructing a spatiotemporal graph based on the spatiotemporal data, wherein nodes represent vibration signals at different spatial positions or time points, and edges represent spatiotemporal dependencies between nodes; S3, applying an adaptive spatiotemporal interference mechanism to randomly perturb the nodes or edges of the spatiotemporal graph to simulate spatiotemporal changes in the environment; S4, introduce embedding method to maintain the relative position of nodes; S5. Using a graph convolutional network to extract features from the space-time graph after interference, extracting the space-time features of the vibration signal; S6. Implement anomaly detection of optical fiber vibration signals based on the extracted features.

2. The method for detecting abnormality of optical fiber vibration signals by spatiotemporal graph learning based on spatiotemporal interference mechanism according to claim 1, characterized in that: The optical fiber vibration signal in step S1 is collected using Φ-OTDR technology, and the collected signal includes vibration position data in the spatial dimension and signal amplitude data in the time dimension.

3. The method for detecting abnormality of optical fiber vibration signals by spatiotemporal graph learning based on spatiotemporal interference mechanism according to claim 1, characterized in that: The S1 step is discretized in the following way to convert it into discrete spatiotemporal data: Spatial discretization: The fiber optic sensors are arranged as discrete sampling points along the transmission path. The fiber length L is divided into N equidistant spatial intervals, each of which represents a spatial node V. i , using the light pulse delay to determine the spatial position d i ; Time discretization: according to the sampling frequency f s The vibration signal is divided into T time steps to form an N×T space-time matrix.

4. The method for detecting abnormality of optical fiber vibration signals by spatiotemporal graph learning based on spatiotemporal interference mechanism according to claim 1, characterized in that: The construction of the spatiotemporal graph in step S2 includes: S2.

1. Define each node as a data point v in the space-time matrix i,k , indicating the i-th spatial position at time step t k The vibration signal value of S2.

2. Construct spatial edges and temporal edges, where: Spatial edges represent the physical distance correlation between adjacent nodes; Temporal edges represent the signal correlation of the same node at consecutive time steps.

5. The method for detecting abnormality of optical fiber vibration signals by spatiotemporal graph learning based on spatiotemporal interference mechanism according to claim 1, characterized in that: The interference process of the adaptive spatiotemporal interference mechanism in step S3 includes the following steps: S3.1, apply random noise to the nodes of the space-time graph; S3.

2. Apply random perturbations to the edges of the space-time graph.

6. The method for detecting abnormality of optical fiber vibration signals by spatiotemporal graph learning based on spatiotemporal interference mechanism according to claim 5, characterized in that: The following formula is used for noise perturbation in step S3.1: Among them, J is the original node feature matrix, α is the interference intensity coefficient, σ 2 is the noise variance, ∈ is the standard normal distribution noise; In step S3.2, the adjacency matrix is ​​adjusted using the following formula: A ′ =AM,M ij ~B(p Where M is a binary matrix with the same dimension as A, representing the random masking or adjustment of edges, B(p) is a random variable generated from a Bernoulli distribution, and p is the probability of edge retention.

7. The method for detecting abnormality of optical fiber vibration signals by spatiotemporal graph learning based on spatiotemporal interference mechanism according to claim 1, characterized in that: The feature extraction of the graph convolution network in step S5 mainly uses graph convolution to extract features on the spatiotemporal graph structure, and aggregates the features of the node neighborhood through the following formula: in, is the node feature of the lth layer, σ is the activation function, W (l) is the convolution weight matrix, c ij is the normalization factor between adjacent nodes.

8. The method for detecting abnormality of optical fiber vibration signals by spatiotemporal graph learning based on spatiotemporal interference mechanism according to claim 1, characterized in that: The anomaly detection in step S6 is completed by the following steps: S6.

1. Using the extracted high-dimensional features Perform signal prediction; S6.

2. Determine whether an abnormality occurs based on the error calculation formula; S6.3, when e t >∈, an abnormal alarm is triggered, where ∈ is the preset abnormal threshold.

9. The method for detecting abnormality of optical fiber vibration signals by spatiotemporal graph learning based on spatiotemporal interference mechanism according to claim 8, characterized in that: The error calculation formula in step S6.2 is as follows: and t =|and t -and' t | Among them, y t is the actual value, y′ t is the predicted value.

10. The method for detecting abnormality of optical fiber vibration signals by spatiotemporal graph learning based on spatiotemporal interference mechanism according to claim 1, characterized in that: The S4 step specifically includes: The embedding feature vector f of each node i (u) is defined by the following formula: f i (u)=α i ·d V (u,V i ) Among them, α i To control the parameters of the embedding distribution, d V (u,V i ) is the distance from node u to anchor point V i distance.

Images