A pipeline safety monitoring method, device and system

Abstract

The application relates to a pipeline safety monitoring method, device and system, comprising: acquiring a to-be-identified vibration signal output by a collection demodulator; inputting the to-be-identified vibration signal into a trained physical information sparse structure learning model, extracting space-time features by using a feature encoder and projecting the space-time features to a hyperbolic geometric space to obtain hyperbolic latent features; calling a pre-stored physical control equation corresponding to each type of event in the model, which is obtained by pre-distillation of a physical law discovery module, and calculating geometric distances between the hyperbolic latent features and each type of theoretical state as physical residuals; inputting the to-be-identified vibration signal into a classifier to obtain a classification confidence; determining a final event category according to the classification confidence and the physical residuals, and synchronously outputting a corresponding physical control equation and physical parameters. The application realizes decision explainability, cross-scene generalization and high robustness of pipeline safety monitoring.

Description

Technical Field

[0001] This application belongs to the field of distributed optical fiber sensing technology, specifically relating to a pipeline safety monitoring method and device. Background Technology

[0002] Distributed Acoustic Sensing (DAS) systems achieve continuous distributed monitoring of vibration signals along optical fibers by injecting laser pulses into the sensing fibers and analyzing the phase changes of Rayleigh backscattered signals. They have been widely used in fields such as oil and gas pipeline monitoring and perimeter security.

[0003] Existing DAS event identification mainly uses deep learning models to classify events by learning statistical patterns in labeled data, but it has the following shortcomings: (1) The model decision-making process lacks physical basis, and the identification results are uninterpretable; (2) The model is highly sensitive to environmental conditions such as soil type and fiber optic laying method, and has poor cross-scene generalization ability; (3) Some schemes attempt to introduce physical information neural networks (PINNs), but they rely on preset simplified equations and are difficult to characterize complex nonlinear dynamic processes such as soil layer coupling and strain hysteresis; (4) In long-distance and strong interference environments, the signal is weak and accompanied by photoelectric attenuation, and the pure data-driven model is prone to false alarms or missed alarms. Therefore, there is an urgent need for a physically interpretable DAS monitoring method with strong generalization ability. Summary of the Invention

[0004] The purpose of this application is to propose a pipeline safety monitoring method based on DAS, which can screen fault sections and determine the location of fault points by monitoring the power frequency current on both sides of each cable segment.

[0005] To achieve the above objectives, embodiments of this application provide a pipeline safety monitoring method, including: Acquire the vibration signal to be identified output by the acquisition and demodulation instrument; The vibration signal to be identified is input into a trained physical information sparse structure learning model. The spatiotemporal features of the vibration signal to be identified are extracted using the feature encoder in the physical information sparse structure learning model. The spatiotemporal features are then projected onto hyperbolic geometric space to obtain hyperbolic latent features. The physical control equations corresponding to each category of events are pre-stored in the physical information sparse structure learning model and obtained offline by the physical law discovery module. Based on the hyperbolic latent features and the physical control equations corresponding to each category, the theoretical states corresponding to each category are predicted respectively. The geometric distance between the hyperbolic latent features and each theoretical state in the hyperbolic geometric space is calculated, and the geometric distance is used as the physical residual corresponding to each category. The hyperbolic latent features are input into the classifier of the physical information sparse structure learning model to obtain the classification confidence of each category; The final event category is determined based on the classification confidence level and the physical residual, and the physical control equation and its physical parameters corresponding to the final event category are output synchronously as the physical basis for safety monitoring.

[0006] In one specific implementation, the physical information sparse structure learning model is trained in the following way: Obtain a training sample set, which includes distributed acoustic sensing vibration signals and their corresponding real event category labels; The training samples are input into the physical information sparse structure learning model to be trained, the spatiotemporal features of the training samples are extracted using the feature encoder, and the spatiotemporal features are projected onto the hyperbolic geometric space to obtain the training hyperbolic latent features. The trained hyperbolic latent features are input into the classifier to obtain the classification prediction result, and the cross-entropy loss is calculated based on the classification prediction result and the real event category label. The physical control equations corresponding to the real event category, which are pre-stored in the physical information sparse structure learning model and obtained offline by the physical law discovery module, are called. The physical residual loss is calculated based on the training hyperbolic latent features and the theoretical state predicted by the physical control equations. The parameters of the physical information sparse structure learning model are updated based on the weighted sum of the cross-entropy loss and the physical residual loss.

[0007] In one specific implementation, the feature encoder is any one of a Transformer architecture encoder, an InceptionTime architecture encoder, or a hyperbolic Transformer architecture encoder; the hyperbolic geometric space is a Poincaré sphere model or a Lorentz model; and the spatiotemporal features are projected onto the hyperbolic geometric space through an exponential mapping of the hyperbolic geometric model.

[0008] In one specific implementation, the geometric distance is any one of geodesic distance, Klein model distance, or Riemannian metric distance.

[0009] In one specific implementation, the physical law discovery module is any one of a physical-symbolic consistency constraint symbolic regression module, a physical information neural network, or a sparse recognition nonlinear dynamics algorithm, and the physical control equation is obtained by the physical law discovery module from the training data.

[0010] In one specific implementation, determining the final event category based on the classification confidence level and the physical residual includes: The category corresponding to the maximum classification confidence score is determined as the first candidate category; The category corresponding to the minimum physical residual is determined as the second candidate category; When the first candidate category matches the second candidate category, the first candidate category or the second candidate category is determined as the final event category; When the first candidate category and the second candidate category are inconsistent, the second candidate category is determined as the final event category.

[0011] In one specific implementation, when the first candidate category is inconsistent with the second candidate category, the classification confidence of the first candidate category and the physical residual of the second candidate category are output simultaneously as evidence for physical-attribution consistency analysis.

[0012] In one specific implementation, before determining the final event category based on the classification confidence level and the physical residual, the method further includes: Determine whether the physical residuals corresponding to each category are all greater than a preset threshold. If so, determine that the current vibration signal to be identified is unknown interference or noise, and do not output the event category. Otherwise, the final event category is determined based on the classification confidence level and the physical residual.

[0013] This application also provides a pipeline safety monitoring device, including: The acquisition module is used to acquire the vibration signal to be identified output by the acquisition and demodulation instrument; The feature extraction and embedding module is used to input the vibration signal to be identified into the trained physical information sparse structure learning model, use the feature encoder in the physical information sparse structure learning model to extract the spatiotemporal features of the vibration signal to be identified, and project the spatiotemporal features onto the hyperbolic geometric space to obtain hyperbolic latent features. The physical residual calculation module is used to call the physical control equations corresponding to each category of events that are pre-stored in the physical information sparse structure learning model and obtained offline by the physical law discovery module. Based on the hyperbolic latent features and the physical control equations corresponding to each category, it predicts the theoretical state corresponding to each category and calculates the geometric distance between the hyperbolic latent features and each theoretical state in the hyperbolic geometric space. The geometric distance is used as the physical residual corresponding to each category. The classification and decision output module is used to input the hyperbolic latent features into the classifier of the physical information sparse structure learning model to obtain the classification confidence of each category, determine the final event category based on the classification confidence and the physical residual, and simultaneously output the physical control equation and its physical parameters corresponding to the final event category as the physical basis for safety monitoring.

[0014] This application also provides a pipeline safety monitoring system, including: The sensing fiber is laid along the pipeline to sense vibration signals along the pipeline. The acquisition and demodulation instrument is connected to the sensing optical fiber and is used to transmit probe light pulses to the sensing optical fiber and receive Rayleigh backscattered light signals, and demodulate and output digitized vibration signals to be identified. An edge computing module is connected to the acquisition and demodulation device, and the edge computing module includes the pipeline safety monitoring device as described above; The monitoring terminal is communicatively connected to the edge computing module and is used to display the final event category and the corresponding physical control equation and physical parameters.

[0015] The embodiments of this application have the following beneficial effects: (1) By calling the physical control equations pre-stored in the model, the corresponding physical equations and physical parameters are output simultaneously while the event category is output, so that the decision has a clear physical basis and overcomes the defect of the unreliability of the "black box" of traditional deep learning models; (2) By changing the judgment basis from "statistical pattern matching" to "physical law verification", the model can automatically adapt to environmental changes such as different soil types and different paving methods, and can still maintain high-precision recognition in unseen scenarios; (3) In addition to the classification confidence, physical residuals are introduced as an independent judgment basis. Noise artifacts have large physical residuals because they do not conform to the real physical laws and are effectively suppressed, avoiding false alarms and false negatives under low signal-to-noise ratio; (4) The pre-stored physical equations are directly called in the inference stage for residual comparison, without complex iterative calculations, and the computational efficiency is high, which is suitable for real-time deployment of edge computing modules. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings required in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic flowchart of a pipeline safety monitoring method according to an embodiment of this application. Detailed Implementation

[0018] The detailed description of the accompanying drawings is intended to illustrate some embodiments of this application and is not intended to represent only the forms in which this application can be implemented. It should be understood that the same or equivalent functions can be accomplished by different embodiments intended to be included within the spirit and scope of this application.

[0019] like Figure 1 As shown, one embodiment of this application provides a pipeline safety monitoring method, including the following steps: Step S10: Acquire the vibration signal to be identified output by the acquisition and demodulation instrument.

[0020] Specifically, in this step, the vibration signal to be identified is a sequence of digital signals output by the acquisition and demodulation instrument after real-time detection and demodulation along the sensing fiber using phase-sensitive optical time-domain reflectometry (OTDR) technology. Specifically, the narrow-linewidth laser in the acquisition and demodulation instrument emits a probe light pulse, which generates Rayleigh backscattered light as it propagates in the sensing fiber. When an external vibration event acts on the fiber, it causes a periodic, minute deformation, thereby changing the phase of the backscattered light. The acquisition and demodulation instrument extracts the phase change information characterizing the vibration intensity through interference and demodulation circuits and quantifies it into a digital signal, which is the vibration signal to be identified acquired in this step.

[0021] Step S20: Input the vibration signal to be identified into the trained physical information sparse structure learning model, use the feature encoder in the physical information sparse structure learning model to extract the spatiotemporal features of the vibration signal to be identified, and project the spatiotemporal features onto hyperbolic geometric space to obtain hyperbolic latent features.

[0022] Specifically, the physical information sparse structure learning model is a neural network model trained jointly based on data-driven principles and physical laws. Its core function is to convert the original time-domain vibration signal into a low-dimensional feature representation that conforms to physical dynamic constraints. In this step, the feature encoder first extracts spatiotemporal features from the input vibration signal to be identified. This encoder can capture the temporal evolution of the vibration signal and its spatial distribution characteristics along the fiber optic distance. Subsequently, using the exponential mapping operation of hyperbolic geometric space, the extracted spatiotemporal feature vector is projected onto a hyperbolic manifold space. This hyperbolic geometric space has negative curvature characteristics, and its geometric structure is suitable for characterizing the hierarchical relationships and complex dynamic features present in distributed acoustic sensing signals. Through this projection, the hyperbolic latent features are obtained. These hyperbolic latent features are no longer purely statistical abstractions, but rather geometric points located on the physical manifold.

[0023] Step S30: Call the physical control equations corresponding to each category of events that are pre-stored in the physical information sparse structure learning model and obtained offline by the physical law discovery module. Based on the hyperbolic latent features and the physical control equations corresponding to each category, predict the theoretical state corresponding to each category, and calculate the geometric distance between the hyperbolic latent features and each theoretical state in the hyperbolic geometric space. Use the geometric distance as the physical residual corresponding to each category.

[0024] Specifically, the physical law discovery module is used to autonomously extract physical control equations that describe the dynamic characteristics of each type of vibration event from historically collected distributed acoustic sensing signals. This process is completed offline before model training and does not depend on the subsequent online inference process; hence, it is called offline distillation. Specifically, for each type of known event (such as mechanical excavation, drilling operations, vehicle passage, etc.), state trajectories are extracted from the corresponding historical signal samples. Then, the physical law discovery module performs sparse regression analysis, searching and combining symbolic expressions from a pre-set symbol set to best describe the dynamic characteristics of this type of trajectory, serving as the physical control equations for that type of event. After distillation, these equations are pre-stored in the model in parametric form (i.e., the specific values ​​of each coefficient term in the equation are determined), for direct use during the inference phase. No further modification or refitting of the equations is required during inference.

[0025] In a preferred embodiment, the physical law discovery module employs a physics-symbolic consistency-constrained symbolic regression framework. This framework maintains a symbol pool containing basic mathematical operators (such as addition, multiplication, and exponentiation), elementary functions (such as sine, cosine, exponential, and logarithmic functions), and physical templates (such as the general form of a second-order vibration system). It also includes a set of equivalent transformation rules to ensure that the expressions generated during the search process are mathematically equivalent or mutually convertible. During distillation, the framework performs sparse regression on the input state trajectory data, searching and combining symbolic expressions from the symbol pool that accurately fit the data and satisfy a preset mechanical topology. The optimization objective of this framework comprises two parts: the first part measures the fitting accuracy of the candidate equations to the training data, and the second part measures the structural deviation of the candidate equations from the preset template. A comprehensive evaluation index is obtained by weighted summation of these two parts, thereby distilling out control equations with concise structure and clear physical meaning while ensuring fitting accuracy. In this way, the framework can describe complex nonlinear dynamic processes such as soil coupling and strain hysteresis, breaking through the limitations of traditional physical information neural networks that rely on preset simple equations.

[0026] In this step, for each predefined event category, its corresponding distilled physical control equation is invoked. The hyperbolic latent feature obtained in step S20 is used as the current state and substituted into the physical control equation for each category to calculate the next moment or the theoretical standard state, denoted as the theoretical state. Then, in the hyperbolic geometric space, the geometric distance between the actual hyperbolic latent feature point and the theoretical state point corresponding to each category is calculated. This geometric distance reflects the degree of deviation between the actual vibration signal and the theoretical physical laws of a certain category, and is defined as the physical residual under that category. The smaller the physical residual, the more the current signal conforms to the physical laws of that type of event.

[0027] Step S40: Input the hyperbolic latent features into the classifier of the physical information sparse structure learning model to obtain the classification confidence of each category.

[0028] Specifically, the classifier typically consists of a fully connected layer and a softmax activation function. It takes hyperbolic latent features as input and outputs a probability distribution vector, where each element represents the confidence level of the vibration signal to be identified belonging to a predefined event category. This confidence level is based on data-driven statistical discrimination results.

[0029] Step S50: Determine the final event category based on the classification confidence level and the physical residual, and simultaneously output the physical control equation and its physical parameters corresponding to the final event category as the physical basis for safety monitoring.

[0030] Specifically, in this step, the final event category is not determined solely by classification confidence, but rather by a combination of statistical discrimination results and physical consistency verification results. This dual-path decision-making mechanism ensures that the output results have both high statistical confidence and conform to objective physical laws. Once the final event category is determined, the system outputs the corresponding physical control equation and its key physical parameters. This physical information provides monitoring personnel with clear and traceable decision-making evidence, making the identification results no longer black-box outputs, but interpretable results with physical meaning.

[0031] In some embodiments, the physical information sparse structure learning model is obtained through a two-stage approach, which includes an offline distillation stage and a model training stage.

[0032] Specifically, in the offline distillation stage: a labeled historical vibration signal sample set is acquired, covering various known events. This sample set includes distributed acoustic sensing vibration signals and their corresponding real event category labels. The physical law discovery module independently analyzes the historical signals for each event category, distilling the corresponding physical governing equations for each category and pre-storing these equations in parameterized form. In this stage, the physical law discovery module only performs symbolic regression analysis on the historical data and does not involve subsequent parameter updates to the neural network model.

[0033] Model training phase: A training sample set is acquired and input into the sparse structure learning model of physical information to be trained. The spatiotemporal features of the training samples are extracted using a feature encoder and projected onto a hyperbolic geometric space to obtain training hyperbolic latent features. These training hyperbolic latent features are then input into a classifier to obtain classification prediction results. Cross-entropy loss is calculated based on the classification prediction results and the real event category labels. The physical control equations corresponding to the real event category, pre-stored in the offline distillation phase, are invoked. Physical residual loss is calculated based on the theoretical state predicted by the training hyperbolic latent features and the physical control equations. The network parameters of the feature encoder and classifier are updated based on the weighted sum of the cross-entropy loss and the physical residual loss. It should be noted that during the model training phase, the coefficients of the distilled physical control equations remain unchanged and do not participate in the gradient update of the network parameters. The network only adjusts the parameters of the feature encoder through backpropagation of the physical residual loss, so that the features output by the encoder gradually approach the solution space that conforms to physical laws. Through this phased approach, the discovery of physical laws and the training of the feature extraction network are decoupled, ensuring the accuracy of the physical equations while reducing the complexity of joint optimization.

[0034] Between the offline distillation stage and the model training stage, a physically consistent noise enhancement strategy can preferably be employed to improve the robustness of the model. This strategy is executed during the model training stage, specifically as follows: In each training iteration, instead of using only the original clean training samples, a perturbation sample set is simultaneously constructed. This perturbation sample is generated by simulating the photoelectric characteristics of the distributed acoustic sensing system and the coupling degradation mechanism between the optical fiber and the ground. Specifically, the generation process of the perturbation sample is as follows: First, an action function is constructed using preset physical parameters. This function includes coupling attenuation parameters and elastic hysteresis parameters to simulate the amplitude attenuation and distortion effects of the signal when it is transmitted at the coupling interface between the optical fiber and the ground. Then, an additive interference term is superimposed. This interference term includes photoelectric drift parameters, phase parameters, and system noise parameters to simulate the drift caused by factors such as changes in ambient temperature and fluctuations in the light source frequency during photoelectric detection, as well as the inherent noise of the system. The perturbation sample generated in this way covers a variety of physical degradation factors, including coupling attenuation effect, elastic hysteresis effect, photoelectric drift effect, and inherent noise of the system. The original samples and perturbed samples are mixed at a preset ratio and then input into the model for training. This forces the feature encoder to learn physical features that remain stable under different degradation conditions, thereby effectively suppressing noise artifacts that are unrelated to actual physical laws and enhancing the model's generalization ability and stability in real industrial noise environments. In a preferred embodiment, the mixing ratio is set to approximately 40% clean samples and approximately 60% perturbed samples.

[0035] In some embodiments, the feature encoder is any one of a Transformer architecture encoder, an InceptionTime architecture encoder, or a hyperbolic Transformer architecture encoder; the hyperbolic geometric space is a Poincaré sphere model or a Lorentz model; the spatiotemporal features are projected onto the hyperbolic geometric space through an exponential mapping of the hyperbolic geometric model.

[0036] Specifically, a hyperbolic Transformer architecture encoder is a preferred implementation. It utilizes the Transformer's attention mechanism to capture the long-range dependencies of vibration signals and embeds the extracted features into hyperbolic space through an exponential mapping of a hyperbolic geometric model. Hyperbolic geometric space, with its negative curvature, can express complex hierarchical structures with less distortion in low-dimensional conditions compared to Euclidean space. The projection process is as follows: the feature encoder first outputs a feature vector in Euclidean space, and then uses an exponential mapping to map this vector from the tangent space to the hyperbolic manifold, thereby obtaining the hyperbolic latent features. The mathematical meaning of this exponential mapping is: the feature vector is first divided by its own norm to obtain a unit direction vector, and then multiplied by the hyperbolic tangent function value of that norm, ultimately obtaining the mapped point in hyperbolic space.

[0037] In some embodiments, the geometric distance is any one of geodesic distance, Klein model distance, or Riemannian metric distance.

[0038] Specifically, geodesic distance refers to the shortest path length along a manifold between two points in hyperbolic space, accurately reflecting the true similarity between two feature points on a negative curvature manifold. In hyperbolic geometric space, the calculation of geodesic distance differs from that of straight-line distance in Euclidean space; it is measured using the inverse function of hyperbolic cosines. Specifically, for any two points in the Poincaré sphere model, the geodesic distance is determined by the Euclidean distance between the two points and their respective margins to the sphere's boundary: the distance increases with the Euclidean distance between the two points, and also increases rapidly as any point approaches the sphere's boundary. Therefore, it has higher sensitivity to differences between points near the boundary, providing a favorable geometric basis for distinguishing different types of vibration characteristics.

[0039] In some embodiments, the physical law discovery module is any one of a physical-symbolic consistency constraint symbolic regression module, a physical information neural network, or a sparse recognition nonlinear dynamics algorithm, and the physical control equation is obtained by the physical law discovery module from the training data.

[0040] Specifically, the physics-symmetric consistency-constrained symbolic regression module is a preferred implementation. This module assumes that the system's dynamic behavior follows the general form of a second-order oscillatory manifold, meaning that the system's acceleration is determined by both the current state and velocity, without needing to pre-define a specific equation structure. During the distillation process, the module maintains a symbol pool predefined with basic mathematical operators, function templates, and equivalent transformation rules to limit the search space. The module's optimization objective comprises two parts: the first measures the fitting accuracy of the candidate equations to the training data, and the second measures the structural deviation of the candidate equations from the pre-defined templates. Through this dual constraint, the module can distill concise and physically meaningful governing equations while ensuring fitting accuracy, thereby describing complex nonlinear dynamic processes such as soil coupling and strain hysteresis.

[0041] In some embodiments, determining the final event category based on the classification confidence level and the physical residual includes: The category corresponding to the maximum classification confidence score is determined as the first candidate category; The category corresponding to the minimum physical residual is determined as the second candidate category; When the first candidate category matches the second candidate category, the first candidate category or the second candidate category is determined as the final event category; When the first candidate category and the second candidate category are inconsistent, the second candidate category is determined as the final event category.

[0042] Specifically, the decision-making logic described above reflects the design principle of prioritizing physics. When data-driven statistical results conflict with physical law verification results, the system chooses the category with the smallest physical residual, which best conforms to physical laws. This mechanism effectively overcomes potential misjudgments that may occur in data-driven models under noise interference.

[0043] In some embodiments, when the first candidate category is inconsistent with the second candidate category, the classification confidence of the first candidate category and the physical residual of the second candidate category are also output synchronously as evidence for physical-attribution consistency analysis.

[0044] Specifically, by simultaneously outputting conflicting statistical and physical evidence, the system provides maintenance personnel with complete attribution information, enhancing the transparency of human-machine collaborative decision-making.

[0045] In some embodiments, before determining the final event category based on the classification confidence level and the physical residual, the method further includes: Determine whether the physical residuals corresponding to each category are all greater than a preset threshold. If so, determine that the current vibration signal to be identified is unknown interference or noise, and do not output the event category. Otherwise, the final event category is determined based on the classification confidence level and the physical residual.

[0046] If the physical residuals of the signal to be identified are very large relative to the physical control equations of all known categories, it means that the signal does not conform to any known physical laws. In this case, it can be identified as an unknown interference event or pure noise. This mechanism gives the model the ability to identify unknown anomalies, effectively reducing the false alarm rate of misclassifying new types of vibrations as known events.

[0047] In some embodiments, the physical residual loss is calculated as follows: First, the hyperbolic latent feature points corresponding to the training samples are obtained in the hyperbolic geometric space. At the same time, the theoretical next state point is predicted from the current state using the physical control equation corresponding to the true class of the sample. Then, the geodesic distance between the two points is calculated and the distance is used as the physical residual loss.

[0048] Specifically, this physical residual loss is not applied to all samples; it is only activated when the model's classification prediction matches the sample's true label. In other words, physical constraints only participate in gradient backpropagation if the model correctly identifies the sample's class. For ease of implementation, an indicator variable can be set, which takes a value of 1 when the predicted class matches the true class and 0 otherwise. This indicator variable is then multiplied by the geodesic distance to obtain the class-specific physical residual loss. Through this mechanism, the model is guided by physical laws only when the classification is correct, avoiding the introduction of unnecessary interference when the classification is incorrect.

[0049] The total training loss is obtained by weighted summation of cross-entropy loss and physical residual loss. Cross-entropy loss reflects the model's performance on statistical classification tasks, while physical residual loss reflects whether the features extracted by the model conform to physical laws. The weights between the two are controlled by a preset hyperparameter. Through this joint optimization approach, the model pursues classification accuracy during training while ensuring that the extracted feature trajectories do not deviate from the solution space allowed by physical laws, thereby achieving physical constraints on the model training process.

[0050] During the model inference phase, the physical control equations have already been distilled offline by the physical law discovery module and are pre-stored in the model in parameterized form, requiring no further modification during inference. During inference, for the input vibration signal to be identified, the system first extracts spatiotemporal features using a feature encoder and projects them onto hyperbolic geometric space. Then, it calls the distilled physical control equations corresponding to each category, predicting the theoretical state for each category from the current feature state. Next, it calculates the geodesic distance between the current feature state and the theoretical state for each category as the physical residual. Simultaneously, it inputs the hyperbolic latent features into a classifier to obtain the classification confidence for each category. Finally, it combines the classification confidence and physical residuals to determine the event category and simultaneously outputs the physical control equations for that category, along with its dynamic parameters, such as natural frequency and damping ratio. For specific types of vibration events, the output physical parameters have clear physical meaning and engineering reference value. For example, for engine-related vibration events, the distillation equations typically take the form of stable periodic terms, and their corresponding physical parameters contain significant fundamental frequency components (such as a 50Hz power frequency component). For drilling operation-related vibration events, the distillation equations contain higher-order terms characterizing the nonlinear constitutive relationship of the soil, and their damping ratio parameters exhibit a regular change with drilling depth. These physical parameters are output along with the physical control equations, serving as the physical basis for safety monitoring, and are provided for secondary verification by maintenance personnel in conjunction with on-site conditions.

[0051] Another embodiment of this application provides a pipeline safety monitoring device, comprising: The acquisition module is used to acquire the vibration signal to be identified output by the acquisition and demodulation instrument; The feature extraction and embedding module is used to input the vibration signal to be identified into the trained physical information sparse structure learning model, use the feature encoder in the physical information sparse structure learning model to extract the spatiotemporal features of the vibration signal to be identified, and project the spatiotemporal features onto the hyperbolic geometric space to obtain hyperbolic latent features. The physical residual calculation module is used to call the physical control equations corresponding to each category of events that are pre-stored in the physical information sparse structure learning model and obtained offline by the physical law discovery module. Based on the hyperbolic latent features and the physical control equations corresponding to each category, it predicts the theoretical state corresponding to each category and calculates the geometric distance between the hyperbolic latent features and each theoretical state in the hyperbolic geometric space. The geometric distance is used as the physical residual corresponding to each category. The classification module is used to input the hyperbolic latent features into the classifier of the physical information sparse structure learning model to obtain the classification confidence of each category; The decision output module is used to determine the final event category based on the classification confidence level and the physical residual, and simultaneously output the physical control equation and its physical parameters corresponding to the final event category as the physical basis for safety monitoring.

[0052] The modules described above are connected sequentially. Data flows through the acquisition module and the feature extraction and embedding module in sequence, then flows in parallel into the physical residual calculation module and the classification module, and finally converges into the decision output module. The specific implementation of each module's function can be found in the corresponding steps of the method embodiment, and will not be repeated here. This device can be deployed in edge computing devices to achieve low-latency, high-reliability real-time monitoring.

[0053] Another embodiment of this application also provides a pipeline safety monitoring system, including: The sensing fiber is laid along the pipeline to sense vibration signals along the pipeline. The acquisition and demodulation instrument is connected to the sensing optical fiber and is used to transmit probe light pulses to the sensing optical fiber and receive Rayleigh backscattered light signals, and demodulate and output digitized vibration signals to be identified. An edge computing module is connected to the acquisition and demodulation device, and the edge computing module includes the pipeline safety monitoring device as described in the above embodiments. The monitoring terminal is communicatively connected to the edge computing module and is used to display the final event category and the corresponding physical control equation and physical parameters.

[0054] The sensing fiber optic cable is physically connected to the acquisition and demodulation unit via an optical interface. The acquisition and demodulation unit transmits the digitized vibration signal to the edge computing module via a data interface such as Ethernet or an industrial data bus. The edge computing module runs a trained physical information sparse structure learning model, executes the above method, and sends the identification results and physical evidence to a remote monitoring terminal for display and alarm via wired or wireless network. After receiving the alarm information, the monitoring terminal can simultaneously display the physical control equations and parameters on which the alarm is based for verification by maintenance personnel.

[0055] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical applications, or technological improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A pipeline safety monitoring method, characterized in that, include: Acquire the vibration signal to be identified output by the acquisition and demodulation instrument; The vibration signal to be identified is input into a trained physical information sparse structure learning model. The spatiotemporal features of the vibration signal to be identified are extracted using the feature encoder in the physical information sparse structure learning model. The spatiotemporal features are then projected onto hyperbolic geometric space to obtain hyperbolic latent features. The physical control equations corresponding to each category of events are pre-stored in the physical information sparse structure learning model and obtained offline by the physical law discovery module. Based on the hyperbolic latent features and the physical control equations corresponding to each category, the theoretical states corresponding to each category are predicted respectively. The geometric distance between the hyperbolic latent features and each theoretical state in the hyperbolic geometric space is calculated, and the geometric distance is used as the physical residual corresponding to each category. The hyperbolic latent features are input into the classifier of the physical information sparse structure learning model to obtain the classification confidence of each category; The final event category is determined based on the classification confidence level and the physical residual, and the physical control equation and its physical parameters corresponding to the final event category are output synchronously as the physical basis for safety monitoring.

2. The method according to claim 1, characterized in that, The physical information sparse structure learning model is obtained through training in the following manner: Obtain a training sample set, which includes distributed acoustic sensing vibration signals and their corresponding real event category labels; The training samples are input into the physical information sparse structure learning model to be trained, the spatiotemporal features of the training samples are extracted using the feature encoder, and the spatiotemporal features are projected onto the hyperbolic geometric space to obtain the training hyperbolic latent features. The trained hyperbolic latent features are input into the classifier to obtain the classification prediction result, and the cross-entropy loss is calculated based on the classification prediction result and the real event category label. The physical control equations corresponding to the real event category, which are pre-stored in the physical information sparse structure learning model and obtained offline by the physical law discovery module, are called. The physical residual loss is calculated based on the training hyperbolic latent features and the theoretical state predicted by the physical control equations. The parameters of the physical information sparse structure learning model are updated based on the weighted sum of the cross-entropy loss and the physical residual loss.

3. The method according to claim 1, characterized in that, The feature encoder is any one of the Transformer architecture encoder, InceptionTime architecture encoder, or hyperbolic Transformer architecture encoder; the hyperbolic geometric space is a Poincaré sphere model or a Lorentz model; the spatiotemporal features are projected onto the hyperbolic geometric space through the exponential mapping of the hyperbolic geometric model.

4. The method according to claim 1, characterized in that, The geometric distance is any one of geodesic distance, Klein model distance, or Riemann metric distance.

5. The method according to claim 1, characterized in that, The physical law discovery module is any one of the following: physical-symbolic consistency constraint symbolic regression module, physical information neural network, or sparse recognition nonlinear dynamics algorithm. The physical control equation is obtained by the physical law discovery module from the training data.

6. The method according to claim 1, characterized in that, The step of determining the final event category based on the classification confidence level and the physical residual includes: The category corresponding to the maximum classification confidence score is determined as the first candidate category; The category corresponding to the minimum physical residual is determined as the second candidate category; When the first candidate category matches the second candidate category, the first candidate category or the second candidate category is determined as the final event category; When the first candidate category and the second candidate category are inconsistent, the second candidate category is determined as the final event category.

7. The method according to claim 6, characterized in that, When the first candidate category is inconsistent with the second candidate category, the classification confidence of the first candidate category and the physical residual of the second candidate category will be output simultaneously as evidence for physical-attribution consistency analysis.

8. The method according to claim 1, characterized in that, Before determining the final event category based on the classification confidence level and the physical residual, the method further includes: Determine whether the physical residuals corresponding to each category are all greater than a preset threshold. If so, determine that the current vibration signal to be identified is unknown interference or noise, and do not output the event category. Otherwise, the final event category is determined based on the classification confidence level and the physical residual.

9. A pipeline safety monitoring device, characterized in that, include: The acquisition module is used to acquire the vibration signal to be identified output by the acquisition and demodulation instrument; The feature extraction and embedding module is used to input the vibration signal to be identified into the trained physical information sparse structure learning model, use the feature encoder in the physical information sparse structure learning model to extract the spatiotemporal features of the vibration signal to be identified, and project the spatiotemporal features onto the hyperbolic geometric space to obtain hyperbolic latent features. The physical residual calculation module is used to call the physical control equations corresponding to each category of events that are pre-stored in the physical information sparse structure learning model and obtained offline by the physical law discovery module. Based on the hyperbolic latent features and the physical control equations corresponding to each category, it predicts the theoretical state corresponding to each category and calculates the geometric distance between the hyperbolic latent features and each theoretical state in the hyperbolic geometric space. The geometric distance is used as the physical residual corresponding to each category. The classification module is used to input the hyperbolic latent features into the classifier of the physical information sparse structure learning model to obtain the classification confidence of each category; The decision output module is used to determine the final event category based on the classification confidence level and the physical residual, and simultaneously output the physical control equation and its physical parameters corresponding to the final event category as the physical basis for safety monitoring.

10. A pipeline safety monitoring system, characterized in that, include: The sensing fiber is laid along the pipeline to sense vibration signals along the pipeline. The acquisition and demodulation instrument is connected to the sensing optical fiber and is used to transmit probe light pulses to the sensing optical fiber and receive Rayleigh backscattered light signals, and demodulate and output digitized vibration signals to be identified. An edge computing module is connected to the acquisition and demodulation device, and the edge computing module includes the pipeline safety monitoring device as described in claim 9; The monitoring terminal is communicatively connected to the edge computing module and is used to display the final event category and the corresponding physical control equation and physical parameters.

Images