Fault detection method and device based on graph structure difference and electronic equipment

Abstract

The application discloses a fault detection method and device based on graph structure difference and electronic equipment, and relates to the technical field of intelligent monitoring. The method comprises the following steps: acquiring an observation data vector of a physical system to be monitored at time t, a linear mapping matrix and an upper state matrix vector, the observation data vector being determined by collecting sensing data of the physical system to be monitored in real time, and the linear mapping matrix and the upper state matrix vector being constructed based on a basic structure model, operating parameters and predictive input of the physical system to be monitored; constructing an optimization problem objective function at time t based on the observation data vector, the linear mapping matrix and the upper state matrix vector at time t; obtaining a sparse difference vector at time t by solving the optimization problem objective function at time t; comparing the sparse difference vector at time t with a threshold value to obtain a fault detection result, and sending a control instruction to a system controller. The method can improve the accuracy of fault diagnosis.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring technology, specifically relating to a fault detection method, device, and electronic equipment based on graph structure differences. Background Technology

[0002] Taking power grid planning systems or urban transportation systems as examples, in actual operation, factors such as equipment aging, environmental interference, and sudden failures often cause the actual state of the system to deviate from its expected planning model. A typical difference is reflected in the graph structure. This graph difference not only has sparsity, local correlation, and temporal evolution, but may also manifest as anomalies in the group structure at the node or edge level.

[0003] However, existing detection methods based on single indicators or global statistics are difficult to accurately characterize and separate differences in graph structure, resulting in high rates of missed and false alarms in power grid planning systems or urban transportation systems, making it difficult to handle system faults in a timely manner. Summary of the Invention

[0004] The technical problem to be solved by the present invention is that existing detection methods based on single indicators or global statistics are difficult to accurately characterize and separate graph structure differences, resulting in high rates of missed and false alarms. In order to solve the above problems, the present invention provides a fault detection method, device and electronic equipment based on graph structure differences.

[0005] The content of this invention includes: In a first aspect, embodiments of the present invention provide a fault detection method based on graph structure differences, comprising: The observation data vector, linear mapping matrix, and upper-level state matrix vector of the physical system to be monitored at time t are obtained. The physical system to be monitored is a power grid planning system or an urban transportation system. The observation data vector is determined by real-time acquisition of sensor data of the physical system to be monitored. The linear mapping matrix and the upper-level state matrix vector are constructed based on the basic structural model, operating parameters, and predictive inputs of the physical system to be monitored, and are used to determine the expected output of the physical system to be monitored. Based on the observation data vector, linear mapping matrix, and upper-level state matrix vector at time t, construct the objective function for the optimization problem at time t; By solving the objective function of the optimization problem at time t, a sparse difference vector at time t is obtained. The sparse difference vector at time t is used to characterize the deviation between the actual operating state and the expected state of the physical system to be monitored. The sparse difference vector at time t is compared with a threshold to obtain the fault detection result. Based on the fault detection result, a control command is sent to the system controller of the physical system to be monitored to trigger the corresponding system adjustment operation. Wherein, the objective function of the optimization problem at time t is Represented as: ; in, Let be the linear mapping matrix at time t. Let be the upper-level state matrix vector at time t. Let be the observation data vector at time t. Let be the sparse difference vector at time t to be solved. for transpose, for Dimensions for The data in the i-th row, Let be the sparse difference vector at time t-1. Let be the graphical Laplace matrix at time t. Used for characterization Norm, Used to characterize the L1 norm, Used to characterize the L2 norm, Used to characterize the trace of a matrix. The first regularization parameter is used. This is the second regularization parameter. This is the third regularization parameter. t is the fourth regularization parameter, where t is a positive integer.

[0006] Optionally, after obtaining the sparse difference vector at time t by solving the objective function of the optimization problem at time t, the method further includes: The gradient contribution of the objective regularization parameter to the objective function of the optimization problem at time t is calculated. The objective regularization parameter includes a first regularization parameter, a second regularization parameter, a third regularization parameter, and a fourth regularization parameter. Corresponding gradient contribution for: ; Based on the reconstructed residual and / or the gradient contribution, the target regularization parameter is adjusted to obtain the adjusted target regularization parameter, which is used to construct the objective function of the optimization problem at time t+1. Wherein, the reconstructed residual for: .

[0007] Optionally, the adjusted target regularization parameter is: ; in, The adjusted target regularization parameters, Let be the target regularization parameter. The gradient vector is composed of the gradient contributions of the first regularization parameter, the second regularization parameter, the third regularization parameter, and the fourth regularization parameter. To reconstruct the residual in the corresponding The result of the change in direction To reconstruct the standard deviation of the residuals, As the first learning rate, As the second learning rate, Used to characterize exponential operations.

[0008] Optionally, before comparing the sparse difference vector at time t with a threshold to obtain the fault detection result, the method further includes: Obtain historical datasets , Let i be the sparse difference vector at time i, and n be the total number of data in the historical dataset. Both i and n are positive integers, and i is less than or equal to n. An average excess graph is plotted based on the historical dataset, and a candidate threshold interval is determined based on the average excess graph. For each candidate threshold in the candidate threshold interval, a generalized Pareto distribution model is fitted, and the target tail threshold and the corresponding generalized Pareto distribution parameters are selected according to the goodness-of-fit index. The threshold is determined based on the target tail threshold, the corresponding generalized Pareto distribution parameters, and the preset detection performance indicators.

[0009] Optionally, before constructing the objective function for the optimization problem at time t based on the observation data vector, linear mapping matrix, and upper-level state matrix vector, the method further includes: Based on the observation data vector at time t and the upper-level state matrix vector at time t, the correlation matrix is ​​learned using sparse inverse covariance estimation or a method based on mutual information. Construct the graphical Laplace matrix for time t based on the correlation matrix.

[0010] Optionally, the step of obtaining the sparse difference vector at time t by solving the objective function of the optimization problem at time t includes: Determine whether the current scenario can only be processed by a single machine; In the case where only single-machine computation is possible, the objective function of the optimization problem is solved using the alternating direction multiplier method to estimate the sparse difference vector at time t; otherwise, the fast iterative shrinking threshold algorithm is used to solve the objective function of the optimization problem to estimate the sparse difference vector at time t.

[0011] Optionally, when the physical system to be monitored is the power grid planning system, the linear mapping matrix and the upper-level state matrix vector are constructed based on the historical load forecast, network topology connection relationship and power grid equipment parameters of the power grid planning system, the observation data vector is determined based on the collected power grid operation data, and the non-zero elements in the sparse difference vector at time t are used to characterize the deviation between the actual power grid operation state at time t and the expected operation state of the power grid planning system. When the physical system to be monitored is the urban transportation system, the linear mapping matrix and the upper-level state matrix vector are constructed based on the road network structure, traffic assignment model and travel demand prediction of the urban transportation system. The observation data vector is determined based on the data collected by the full road network detector. The non-zero elements in the sparse difference vector at time t are used to characterize the deviation between the actual traffic state and the expected traffic state of the urban transportation system.

[0012] Secondly, embodiments of the present invention provide a fault detection device based on graph structure differences, comprising: The first acquisition module is used to acquire the observation data vector, linear mapping matrix, and upper-level state matrix vector of the physical system to be monitored at time t. The physical system to be monitored is a power grid planning system or an urban transportation system. The observation data vector is determined by real-time acquisition of sensor data of the physical system to be monitored. The linear mapping matrix and the upper-level state matrix vector are constructed based on the basic structural model, operating parameters, and predictive inputs of the physical system to be monitored, and are used to determine the expected output of the physical system to be monitored. The construction module is used to construct the objective function of the optimization problem at time t based on the observation data vector, linear mapping matrix, and upper-level state matrix vector at time t. The solution module is used to solve the objective function of the optimization problem at time t to obtain the sparse difference vector at time t. The sparse difference vector at time t is used to characterize the deviation between the actual operating state and the expected state of the physical system to be monitored. The processing module is used to compare the sparse difference vector at time t with a threshold to obtain the fault detection result, and send a control command to the system controller of the physical system to be monitored based on the fault detection result to trigger the corresponding system adjustment operation. Wherein, the objective function of the optimization problem at time t is Represented as: ; in, Let be the linear mapping matrix at time t. Let be the upper-level state matrix vector at time t. Let be the observation data vector at time t. Let be the sparse difference vector at time t to be solved. for transpose, for Dimensions for The data in the i-th row, Let be the sparse difference vector at time t-1. Let be the graphical Laplace matrix at time t. Used for characterization Norm, Used to characterize the L1 norm, Used to characterize the L2 norm, Used to characterize the trace of a matrix. The first regularization parameter is used. This is the second regularization parameter. This is the third regularization parameter. t is the fourth regularization parameter, where t is a positive integer.

[0013] Thirdly, embodiments of the present invention provide an electronic device, including: a memory, a processor, and a program stored in the memory and executable on the processor; the processor is configured to read the program in the memory to implement the steps in the fault detection method based on graph structure differences as described in the first aspect.

[0014] Fourthly, embodiments of the present invention provide a readable storage medium for storing a program, which, when executed by a processor, implements the steps in the fault detection method based on graph structure differences as described in the first aspect.

[0015] The beneficial effects of this invention are as follows: In the embodiments of this invention, an optimization problem objective function integrating sparsity, group sparsity, graph smoothness, and temporal smoothness is constructed. Sparsity constraints cause the model to focus on a few anomalous nodes, effectively suppressing noise interference; group sparsity encourages synchronous activation of anomalous nodes in adjacent time steps or functionally related nodes, which helps to identify multi-source faults with spatial clustering or cooperative behavior; graph smoothness utilizes the system topology to force anomalous signals to present local consistency on the graph, avoiding isolated false alarms and enhancing the accuracy of locating structurally related faults; temporal smoothness, by constraining the evolution continuity of anomalous signals in the time dimension, effectively distinguishes between transient disturbances and persistent deviations, improving the ability to identify long-term anomalous patterns. The above four regularization terms work synergistically in the optimization problem objective function to form a joint prior guidance for multi-scale and multi-type anomalous signals in complex systems. This makes the difference vector obtained from the optimization solution not only meet the requirements of fitting the observation data but also have clear physical interpretability and structural rationality. Thus, without the need for a large number of labeled samples, high-precision detection, accurate location, and effective separation of diverse fault modes such as sparse transient faults, persistent deviations, and regional anomalies can be achieved. By using the above methods to monitor the graph differences of power grid planning systems or urban transportation systems, faults in these systems can be detected quickly and accurately, and timely responses can be made to address them, thereby improving the operational reliability of the power grid planning systems or urban transportation systems. Attached Figure Description

[0016] Figure 1 This is one of the flowcharts for a fault detection method based on graph structure differences provided in an embodiment of the present invention; Figure 2 The second flowchart is a fault detection method based on graph structure differences provided in an embodiment of the present invention. Figure 3 A schematic diagram of a fault detection device based on graph structure differences provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0017] In the embodiments of this application, the term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. The character " / " generally indicates that the preceding and following related objects have an "or" relationship. In the embodiments of this application, the term "multiple" refers to two or more, and other quantifiers are similar. The terms "first," "second," etc., in the specification of this application are used to distinguish similar objects and are not used to describe a specific order or sequence. It should be understood that such terms can be interchanged where appropriate so that embodiments of this application can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first" and "second" are usually of the same class, without limiting the number of objects. For example, the first object can be one or multiple.

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0019] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0020] This application provides a fault detection method, apparatus, and electronic device based on graph structure differences, which aims to significantly improve the early detection capability and diagnostic robustness of potential faults, and enhance the reliability of power grid planning systems or urban transportation systems under real disturbance environments.

[0021] Please see Figure 1 , Figure 1 This is a flowchart illustrating the fault detection method based on graph structure differences provided in an embodiment of the present invention. The method specifically includes the following steps: Step 101: Obtain the observation data vector, linear mapping matrix, and upper-level state matrix vector of the physical system to be monitored at time t. The physical system to be monitored is a power grid planning system or an urban transportation system. The observation data vector is determined by real-time acquisition of sensor data of the physical system to be monitored. The linear mapping matrix and the upper-level state matrix vector are constructed based on the basic structural model, operating parameters, and predictive inputs of the physical system to be monitored, and are used to determine the expected output of the physical system to be monitored.

[0022] Step 102: Based on the observation data vector, linear mapping matrix, and upper-level state matrix vector at time t, construct the objective function for the optimization problem at time t.

[0023] Step 103: By solving the objective function of the optimization problem at time t, a sparse difference vector at time t is obtained. The sparse difference vector at time t is used to characterize the deviation between the actual operating state and the expected state of the physical system to be monitored.

[0024] Step 104: Compare the sparse difference vector at time t with a threshold to obtain the fault detection result, and send a control command to the system controller of the physical system to be monitored based on the fault detection result to trigger the corresponding system adjustment operation.

[0025] The method provided in this invention is applied to existing industrial and urban intelligent systems to address the problem of difficulty in accurately detecting and separating graph structure differences in power grid planning systems or urban transportation systems due to factors such as equipment aging, environmental interference, and sudden failures. In power grid planning systems or urban transportation systems, the system's state and output are typically modeled to obtain a planning model. Graph structure differences refer to the deviations between the actual system state diagram and the expected graph structure in the planning model caused by abnormal factors during system operation. When graph structure differences exist (such as topology changes, parameter anomalies, etc.), sparse and anomalous differences will appear in the observed data.

[0026] The observation data vector is determined by real-time acquisition of sensor data from the physical system under monitoring. After time alignment and preprocessing, the observation data constitutes the observation vector at time t. For example, in a power grid planning system, the observation data consists of measurements such as voltage, current, and active / reactive power collected by synchronous phasor measurement units, remote terminal units, or smart meters deployed at substations, feeders, or buses. In an urban transportation system, the observation data includes traffic flow, average speed, occupancy rate, or travel time collected by devices such as geomagnetic coils and video detectors.

[0027] The linear mapping matrix and upper-level state vector are constructed based on the basic structural model, operating parameters, and predictive inputs of the physical system to be monitored. For example, in a power grid planning system, the basic structural model includes equipment parameters such as power grid topology connections (node-branch correlation matrix), line impedance, and transformer turns ratio; predictive inputs include short-term load forecasts and renewable energy output forecasts generated based on historical data and meteorological information; the upper-level state vector represents the expected system state of the planning model at time t, such as the voltage amplitude and phase angle of each node, and branch power flow; the linear mapping matrix is ​​approximated by the power grid Jacobian matrix or DC power flow sensitivity matrix, reflecting the physical mapping relationship from theoretical state to observable quantities. In an urban transportation system, the basic structural model includes static attributes such as road network topology (node-road segment connections), number of lanes, speed limits, and signal timing schemes; predictive inputs include travel demand forecasts generated based on land use, population flow, and multi-source big data; the upper-level state vector represents the expected flow, density, or speed of each road segment at time t under a traffic allocation model (such as user equilibrium or random user equilibrium); the linear mapping matrix is ​​determined by the detector deployment location and road network topology.

[0028] The linear mapping matrix and the upper-level state vector are used together to determine the expected output of the physical system to be monitored at time t. This expected output represents the ideal measurement value that should be observed under the current planning model assumptions, and serves as the benchmark for subsequent anomaly detection, deviation identification and model correction.

[0029] The observation equation is a core concept in system modeling, state estimation, signal processing, and control theory. It describes the mathematical relationship between observable external data and the internal, unmeasurable states of the system.

[0030] In this embodiment, the observation equation is expressed as follows: ; in, Represents the vector of observed data. Represents a known linear mapping matrix. The upper-level state matrix vector of the system is calculated. It can be used to characterize the expected output of a system under normal conditions. This represents the sparse difference vector to be estimated. This represents the noise vector.

[0031] Due to the presence of noise vectors, sparse dissimilarity vectors are difficult to compute directly. In some embodiments, in order to estimate from the observation equations... Add regularization terms based on prior knowledge Regularization term Represents hope Based on the physical or mathematical properties it possesses, the optimization equation is obtained. for: ; in, Used for characterization Norm, Norm refers to the sum of the squares of all elements of a matrix; it is the fitted term. represent of Norm, thus guaranteeing Consistency with the observed data vector, This is the regularization parameter, used to control the balance between regularization and fitting error accuracy.

[0032] Physical system under monitoring It may exhibit the following characteristics, such as at different times. They may have similarities, or sometimes need to be The sparsity of the signal itself or its adjacency matrix, and the fact that different edge nodes may share some common characteristics, necessitate encouraging group sparsity. Therefore, in this embodiment, regularization terms such as sparsity, group sparsity, graph smoothness, and time smoothness are introduced to construct an optimization problem at time t. It can be represented as follows: ; in, Let be the linear mapping matrix at time t. Let be the upper-level state matrix vector at time t. Let be the observation data vector at time t. Let be the sparse difference vector at time t to be solved. for transpose, for Dimensions for The i-th row, To estimate the sparse difference vector at time t-1, Let be the graphical Laplace matrix at time t. Used for characterization Norm, Used to characterize the L1 norm, and to characterize the absolute sum and magnitude of vector elements, often used to induce sparsity. Used to characterize the L2 norm, and to characterize the length / magnitude of a vector, it is often used to control smoothness and overall amplitude, and to prevent overfitting. Used to characterize the trace of a matrix.

[0033] Specifically, sparsity refers to the characteristic that only a few nodes or edges in a differential signal are affected, while the majority of elements are zero or close to zero. Represents sparsity constraints, i.e. Only a few nodes or feature dimensions are affected, resulting in overall sparsity, meaning many elements are 0. This is the non-negative first regularization parameter corresponding to the sparsity constraint, used to emphasize sparsity. The sparsity constraint prompts the model to focus on a few outlier nodes, effectively suppressing noise interference, and is suitable for detecting transient equipment failures or local mutations.

[0034] Group sparsity refers to the occurrence of differential signals in "groups," such as multiple feature dimensions of the same node being abnormal simultaneously, or certain related edges deviating collectively, exhibiting a structured anomaly pattern. This represents the sparsity constraint of the group, where this term corresponds to the deviation of the i-th node across all feature dimensions. In this case, anomalies typically occur at the "node" level, rather than just a single feature. The second regularization parameter is non-negative and is used to emphasize group-structured sparsity. Group sparsity encourages anomalous synchronous activation on adjacent time steps or functionally related nodes, which helps identify multi-source faults with spatial clustering or cooperative behavior, such as regional load abrupt changes and cascading faults.

[0035] Graph smoothness, based on graph Laplace constraints, encourages adjacent nodes or edges in a graph to have similar difference values, allowing anomalies to spread or cluster smoothly in space, rather than appearing in isolation. This represents a graph smoothness constraint, which encourages adjacent nodes in the graph to have similarities. The value indicates that anomalies propagate or cluster smoothly on the graph, rather than appearing randomly and in isolation. The third regularization parameter is non-negative and is used to emphasize the local manifold structure of data points in the graph. Graph smoothing utilizes the system topology, such as the connections between power grids, communication networks, and transportation networks, to force abnormal signals to exhibit local consistency on the graph, avoid isolated false alarms, and enhance the accuracy of locating structure-related faults.

[0036] Temporal smoothness refers to the gradual change in the difference signal between adjacent time steps, without any abrupt jumps in the difference value. Representing a time smoothing constraint, this term encourages adjacent time steps to... The changes are gradual, meaning that the anomalies / deviations do not change drastically, but evolve smoothly over time. The fourth regularization parameter is non-negative and is used to emphasize the effect of the time smoothing term. Time smoothing effectively distinguishes between transient disturbances and persistent deviations, such as sensor drift and slow degradation faults, by constraining the continuity of the evolution of anomalous signals in the time dimension, thereby improving the ability to identify long-term anomalous patterns.

[0037] It should be understood that, , , and The larger the value, the greater its weight in the physical system being monitored, and the more important different constraints are in different application scenarios. These parameters need to be selected according to the specific scenario. In some embodiments, , , and The value is predetermined and is set and adjusted through manual optimization or cross-validation.

[0038] In other embodiments, to enhance the model's adaptability, the weights of each regularization term are dynamically adjusted by monitoring gradient changes and reconstruction residuals during the optimization process online, thereby achieving flexible responses to different anomaly patterns. Optionally, in some embodiments, after step 103, the method further includes: The gradient contribution of the objective regularization parameter to the objective function of the optimization problem at time t is calculated. The objective regularization parameter includes a first regularization parameter, a second regularization parameter, a third regularization parameter, and a fourth regularization parameter. Corresponding gradient contribution for: ; Based on the reconstructed residual and / or the gradient contribution, the target regularization parameter is adjusted to obtain the adjusted target regularization parameter, which is used to construct the objective function of the optimization problem at time t+1. Wherein, the reconstructed residual for: .

[0039] As a specific implementation, during the optimization iteration process, the gradient contribution of each regularization parameter to the objective function of the optimization problem at time t is calculated. : ; in, This is an optimization problem at time t. The first regularization parameter The corresponding gradient contribution, The second regularization parameter The corresponding gradient contribution, The third regularization parameter The corresponding gradient contribution, The fourth regularization parameter The corresponding gradient contribution. In some embodiments, if the gradient of any regularization parameter remains large, it indicates that the current data is sensitive to the constraint, and the weight should be increased appropriately; otherwise, it should be decreased.

[0040] Simultaneously, the reconstruction residuals are calculated. The reconstruction residuals are the differences between the observed data and the model's predicted values, reflecting the model's fit and the strength of anomalies. They are crucial for adjusting regularization parameters and determining the system's state. Then, based on the reconstruction residuals... Based on the distribution characteristics such as mean, variance, and kurtosis, and the actual situation, the regularization parameters are adjusted in real time. Specifically, the residuals are reconstructed. for: .

[0041] In some embodiments, if the reconstructed residuals exhibit a heavy-tailed distribution with a large number of outliers, the sparse regularization term and the graph regularization term are enhanced, i.e., the first regularization parameter is increased. and the third regularization parameter If the reconstructed residuals exhibit a systematically non-zero mean, then the time smoothing term should be enhanced, i.e., the fourth regularization parameter should be increased. If the reconstructed residuals exhibit abnormal group structure, such as certain graph nodes or edges, then the group sparsity regularization term should be enhanced, i.e., the second regularization parameter should be increased. .

[0042] As a specific implementation, at each time t, the regularization parameter is updated according to the following rules, and the adjusted target regularization parameter is: ; in, Both β and t are pre-set learning rates. For ease of distinction, β is used to define the learning rate. This is referred to as the first learning rate, and β as the second learning rate. In some embodiments, adjusting the learning rate setting controls the rate of change of system parameters, thus preventing sudden and rapid system switching that could lead to insufficient performance stability.

[0043] In this embodiment, The gradient vector is composed of the gradient contributions of each regularization parameter. For all gradient vectors The 2-norm, The result of the change in the reconstructed residual in the direction of the corresponding regularization term can be represented by the Frobenius norm pair of the reconstructed residual. Calculation of partial derivatives, Let be the standard deviation of all reconstructed residuals.

[0044] In this embodiment, It can be understood as , It can also be understood as ,Right now The regularization parameters used to construct the objective function of the optimization problem at time t+1. The adjusted regularization parameters... , , and The data is stored for use in establishing the optimization equations for the next time step. Through this method, the weights of each regularization term are dynamically adjusted based on the gradient changes and residual distribution during the optimization process, enabling the model to flexibly respond to changes in system state and improve its adaptability to different abnormal modes.

[0045] At each time point, after obtaining the sparse difference vector, its variation amplitude needs to be continuously monitored to identify abnormal events. If the test statistic is less than the threshold, the amplitude of the evolutionary difference is considered to be within a reasonable range. Otherwise, it is considered to be a transient fluctuation caused by external disturbances such as equipment failure or sudden events, and the relevant information is fed back to the decision-making end (i.e., the system controller) of the monitored physical system in real time to trigger corresponding system adjustment operations, thereby adjusting the system response in a timely manner. The specific system adjustment operations are determined by the system controller based on the fault monitoring results, which will not be elaborated here.

[0046] In this embodiment, the dynamic adaptive regularization parameter adjustment mechanism can flexibly respond to the dynamic changes in the system's operating state, thus significantly improving the early detection capability and diagnostic robustness of potential faults, and enhancing the reliability of power grid planning systems or urban transportation systems under real disturbance environments.

[0047] Extreme value theory is a statistical theory used to analyze the distribution of extreme values. In this invention, it is used to fit the tail features of differential data, adaptively determine the anomaly threshold, and achieve reliable monitoring of rare abnormal events. In this embodiment, the threshold is determined based on extreme value theory.

[0048] Optionally, in some embodiments, before step 104, the method further includes: Obtain historical datasets , Let i be the sparse difference vector at time i, and n be the total number of data in the historical dataset. Both i and n are positive integers, and i is less than or equal to n. An average excess graph is plotted based on the historical dataset, and a candidate threshold interval is determined based on the average excess graph. For each candidate threshold in the candidate threshold interval, a generalized Pareto distribution model is fitted, and the target tail threshold and the corresponding generalized Pareto distribution parameters are selected according to the goodness-of-fit index. The threshold is determined based on the target tail threshold, the corresponding generalized Pareto distribution parameters, and the preset detection performance indicators.

[0049] As a specific implementation, firstly, historical data is preprocessed and stabilized to obtain a historical dataset. For example, missing values ​​and obvious outliers (such as maxima / minimum outliers caused by communication errors) are removed from the original historical data, ensuring that all retained data are nominal data. Finally, the historical dataset is obtained. .

[0050] In some embodiments, if the data in the historical dataset is non-stationary, a difference transformation is used to obtain the differencing sequence, i.e., for the ... Data at each time point, considering Subsequent Monitoring will be conducted. For simplicity, this embodiment assumes that subsequent data is stable and uses historical datasets. Proceed with the subsequent operations.

[0051] The excess function is calculated based on the stored historical dataset using the following formula. : ; in, The tail threshold, Medium greater than the tail threshold The data can be considered as the tail of the distribution, which can be described by the Generalized Pareto Distribution (GPD). Let be an indicator function, representing if ,but =1, otherwise It is 0.

[0052] Based on the above analysis, an average excess graph is constructed, where the horizontal axis is... The value ranges from 0 to The vertical axis is the corresponding Based on the average excess chart, the following were identified: The starting point where a linear trend begins Furthermore, it requires that the number of samples exceeding the threshold exceeds a preset value (e.g., 50), thereby constructing a candidate threshold range. .

[0053] For each candidate threshold in the candidate threshold interval ,extract Medium greater than the candidate threshold The sample is used to calculate the difference between this sample and u, thus constructing an excess sample set. ,Right now Using an excess sample set By fitting the parameters of the GPD, the shape parameters are obtained. and scale parameters The distribution function of GPD is: ; in, For use in describing Shape parameters of the distribution shape is the scaling parameter. In some embodiments, the maximum likelihood estimation method is used. and A fitting process is then performed. Therefore, for each candidate threshold, a generalized Pareto distribution function is obtained through maximum likelihood estimation.

[0054] In some embodiments, the average of the sum of the absolute distances between the Empirical Cumulative Distribution Function (ECDF) values ​​and the generalized Pareto distribution at each sample point is used as a measure of difference, and the parameter corresponding to the curve with the smallest difference is selected as the optimal parameter. Specifically, the GPD model and the excess sample set are calculated. The fitting difference between ECDF values ​​and theoretical GPD values ​​is defined as the average of the absolute deviations of ECDF values ​​from the theoretical GPD values ​​at all excess sample points.

[0055] Iterate through all candidate thresholds and select the target tail threshold corresponding to the GPD model with the smallest fitting difference. and its corresponding target shape parameters and target scale parameters As the optimal GPD model, based on this optimal GPD model, and combining the target performance index and expected performance requirements, a suitable threshold T is finally determined through the expected quantile. The calculation method for T is as follows: ; in, It exceeds the target tail threshold The sample proportion can be obtained from historical data that is greater than The number of samples is estimated by dividing the total number of samples, and P is a preset detection performance index.

[0056] Will Compare with the final threshold T: If If the value is greater than T, a warning about the existence of the difference is triggered and corresponding actions are taken; otherwise, the sample at the next time step is processed.

[0057] In practical implementation, the distribution changes of sparse difference vectors are continuously monitored, and threshold re-evaluation and model updates are performed periodically based on long-term collected data. Historical difference distribution is automatically updated according to seasonal and periodic changes to avoid algorithm failure due to environmental changes / scene migration.

[0058] In some embodiments, the graph Laplacian matrix is ​​predefined and static. In other embodiments, considering that the underlying relational structure of the system may drift slowly, making it difficult for static results to reflect the dynamic changes in the relations between nodes during system operation, the graph Laplacian matrix is ​​updated at each time step.

[0059] Optionally, in some embodiments, before step 102, the method further includes: Based on the observation data vector at time t and the upper-level state matrix vector at time t, the correlation matrix is ​​learned using sparse inverse covariance estimation or a method based on mutual information. Construct the graphical Laplace matrix for time t based on the correlation matrix.

[0060] In this embodiment, at time t, the real-time dependencies between nodes are learned using the current observation residuals (or original observations) to construct a sparse, time-varying graph, and from this, a graph Laplacian matrix is ​​generated for subsequent smoothness regularization or anomaly propagation modeling.

[0061] Specifically, at time t, based on the upper-level state matrix vector The system calculates the real-time correlation between nodes. Specifically, it uses sparse inverse covariance estimation or mutual information-based methods to learn a sparse correlation matrix online. Correlation matrix Each element of this matrix is ​​used to characterize the conditional dependencies between system state variables. State estimation The Middle The component and the first The strength of conditional dependency or the size of mutual information between components.

[0062] Based on the learned correlation matrix Construct the corresponding time-varying graph Laplacian matrix : ; in, for The degree matrix.

[0063] In some embodiments, it is necessary to select an appropriate algorithm based on the target of the planning subsystem. A centralized scenario refers to a scenario where data can be aggregated on a single machine, and computation is performed through a unified data center. A distributed scenario refers to a scenario where each worker node relies only on local data, solves local model parameters in parallel, and then aggregates and updates the data on a global server.

[0064] Optionally, in some embodiments, step 103 includes: Based on the performance requirements of the actual scenario, determine whether the current scenario can only be operated on a single machine; In the case where only single-machine computation is possible, the objective function of the optimization problem is solved using the alternating direction multiplier method to estimate the sparse difference vector at time t; otherwise, the fast iterative shrinking threshold algorithm is used to solve the objective function of the optimization problem to estimate the sparse difference vector at time t.

[0065] In this embodiment, the determination is made based on whether single-machine operation is the only possible scenario. If single-machine operation is the only possible scenario, the current scenario is considered centralized; otherwise, it is considered distributed. Appropriate scenario algorithms are selected based on the objectives of the power grid planning system or the urban transportation system. In environments that encourage centralized and rapid processing, the Fast Iterative Shrinkage-Thresholding Algorithm (FISTA) is used; in environments that encourage distributed collaboration and data locality, the Alternating Direction Method of Multipliers (ADMM) is used.

[0066] In some embodiments, the objective function of the optimization problem is solved using the distributed ADMM algorithm. Although the ADMM algorithm is relatively slow in terms of convergence speed, its advantage lies in its ability to directly utilize the already dispersed subsystem data to perform parameter estimation in a distributed manner.

[0067] The ADMM algorithm decomposes the original complex optimization problem into several relatively simple subproblems, and approximates the global optimum by alternately optimizing these subproblems and updating the multipliers. As a specific example, the steps for solving the objective function of the optimization problem using the ADMM algorithm are as follows: First, based on the ADMM algorithm definition, auxiliary parameters Z for the ADMM-driven optimization problem are established: ; st ; Based on the principles of the ADMM algorithm, it is easy to prove that this problem is equivalent to minimizing... function, where, for: ; in, and These are all auxiliary variables introduced to implement the ADMM method. For vectors The i-th variable, Auxiliary parameters are introduced to implement the ADMM method.

[0068] Will , Initialize as an all-zero vector or all-zero matrix according to the required structure, where the upper-right subscript represents the iteration number of the ADMM algorithm. Simultaneously, based on the estimation from the previous time step (i.e., time t-1)... Construct time smoothing constraints for the current moment (i.e., time t).

[0069] In the In the next iteration, repeat the following operation until... Convergence occurs, and the convergence result is the estimated value at the current moment. Specifically, during the iteration process, updates are performed as follows: : ; in, For the first The estimated value obtained in the second iteration For the first In the next loop , It is the identity matrix. These are the auxiliary parameters mentioned above.

[0070] Specifically, during the iteration process, it is further updated in the following manner. : ; in, Representative variable The nearest neighbor operator specifically includes the following steps: First, calculate the intermediate variables. The results are as follows: ; in, For a symbolic function, for If greater than 0, then Conversely, it is 0. This is used to achieve soft threshold shrinkage. Then, the following calculations are performed: ; in, yes The Group observation, This represents the L2 norm of all vectors in the group. This step shrinks the group to achieve group sparsity. For each group i, the above process is repeated until all results for each group have been processed, and the results are combined to form a vector. .

[0071] Specifically, during the iteration process, updates are performed as follows: : ; Specifically, during the iteration process, updates are performed as follows: : ; in, It is a constant greater than 1. It is a constant less than 1, used to accelerate ADMM convergence.

[0072] In some embodiments, the objective function is solved using the centralized FISTA algorithm. The advantage of the FISTA algorithm lies in its applicability to centralized scenarios and its ability to provide higher iterative computation speed. The FISTA algorithm accelerates convergence by introducing a Nesterov momentum extrapolation step, building upon the standard proximal gradient method. The detailed steps for solving the objective function of the optimization problem using the FISTA algorithm are as follows: Will , Initialize the vector or matrix to all zeros according to the required structure, and store the estimate from the previous time step. , used for estimation at this moment.

[0073] Repeat the following steps until... Convergence, obtaining Specifically, during the iteration process, the gradient is updated as follows: : ; Where k is the number of iterations, It is the Lipschitz constant. It is an identity matrix.

[0074] Specifically, during the iteration process, updates are performed as follows: : ; Among them, the nearest neighbor operator The specific calculation method can be found in the description of the ADMM algorithm, and will not be elaborated here.

[0075] Specifically, during the iteration process, the momentum is updated as follows: ; Specifically, during the iteration process, updates are performed as follows: : .

[0076] Optionally, in some embodiments, when the physical system to be monitored is the power grid planning system, the linear mapping matrix and the upper-level state matrix vector are constructed based on the historical load forecasts, network topology connections and power grid equipment parameters of the power grid planning system, the observation data vector is determined based on the collected power grid operation data, and the non-zero elements in the sparse difference vector at time t are used to characterize the deviation between the actual power grid operation state at time t and the expected operation state of the power grid planning system. In the power grid planning system, the upper-level state matrix vector X and the linear mapping matrix A are determined based on historical load forecasts, network topology connections, and equipment parameters. Specifically, historical loads can be the active / reactive loads of each node, network topology connections include the topology results between lines, transformers, and buses, and equipment parameters include line impedance, transformer turns ratio, and generator output limits. This system is deployed in regional dispatch centers, forming a distributed monitoring network. In actual power grid operation, each node collects real-time measurement data through sensors and synchronous phasor measurement units (PMUs), such as node voltage amplitude and phase angle, branch current, active / reactive power, switch status, and circuit breaker location, thus obtaining an observation data vector.

[0077] As a specific implementation, the observation data vector consists of real-time measurements collected by synchronous phasor measurement units, remote terminal units, or smart meters deployed at substations, feeders, or buses. The linear mapping matrix is ​​approximated by the power grid's Jacobian matrix or DC power flow sensitivity matrix. The upper-level state matrix vector specifically includes the voltage amplitude and phase angle of each node, branch power flow, etc. These are ideal states derived from load forecasting and network topology calculations. The Graph Laplace matrix is ​​constructed based on the physical topology of the power grid. Each non-zero element directly corresponds to a specific physical measurement point in the power grid (such as the active power of a transmission line or the voltage of a bus). The magnitude and direction of this element represent the degree of deviation between the actual measured value and the model prediction value at that point.

[0078] The method provided in this invention can identify local deviations between the actual topology, equipment parameters, or load distribution in a region and the planning model. Short-term abnormal deviations may originate from instantaneous equipment failures or sudden load surges, while long-term abnormal deviations may stem from continuously increasing distributed energy penetration or parameter drift caused by line aging. In this embodiment, the ADMM algorithm can be used for solving the problem. Through distributed iterative coordination, without processing all the original data, each sub-part individually estimates the global sparse difference vector E and feeds the results back to the upper-level system controller to trigger corresponding protection or isolation operations, thereby achieving dynamic verification and rolling correction of the power grid.

[0079] Optionally, in some embodiments, when the physical system to be monitored is the urban transportation system, the linear mapping matrix and the upper-level state matrix vector are constructed based on the road network structure, traffic assignment model and travel demand prediction in the urban transportation system, the observation data vector is determined based on the data collected by the full road network detector, and the non-zero elements in the sparse difference vector at time t are used to characterize the deviation between the actual traffic state and the expected traffic state of the urban transportation system.

[0080] In urban transportation systems, based on road network structure, traffic assignment models, and travel demand forecasts, a higher-level state matrix vector X and a linear mapping matrix A are constructed under the planning state. The road network structure describes the topological connections and physical attributes of the urban road system. The traffic assignment model allocates the predicted travel demand to specific road networks, obtaining the traffic flow distribution for each road segment. Travel demand forecasts represent the predicted travel volume between different areas within a future time period. This system is deployed at the traffic control center, accessing the entire road network detectors in real time, and then obtaining the observation data vector Y based on the collected data.

[0081] As a specific implementation, the observation data vector consists of real-time traffic data collected by devices such as geomagnetic coils and video detectors deployed on the road network, specifically including segment flow, average speed, occupancy rate, or travel time. The linear mapping matrix is ​​determined by the detector deployment locations and the road network topology. The upper-level state matrix vector represents the expected state of each road segment at time t under a traffic assignment model (such as a user equilibrium model), specifically including the expected flow, density, or speed of each road segment. The graph Laplace matrix is ​​constructed based on the topological structure of the urban road network. Each non-zero element directly corresponds to a specific road segment or intersection in the road network, and the value of the element represents the actual traffic flow status at that location (such as flow rate and speed).

[0082] In this embodiment, the FISTA algorithm can be used to quickly solve for the sparse differences across the entire road network, yielding a sparse difference vector E. Non-zero elements indicate the road segments and their intensity that deviate from the planned traffic conditions. This system is suitable for small and medium-sized cities or regional traffic management centers. Leveraging centralized computing resources, it can monitor the graph differences between the planning model and the operational status. Through graph difference monitoring, it determines fault detection results and sends control commands to the system controller based on these results, triggering corresponding scheduling operations on the urban traffic network.

[0083] Please see Figure 2The following is an example of a specific embodiment. In this embodiment, the graph structure differences of the power grid planning system or urban transportation system are continuously monitored. At each time point, the sparse difference vector at that time point is calculated using the method provided in this embodiment of the invention to achieve fault monitoring.

[0084] The following explanation uses any time t as an example. First, we obtain the observation data vector of the power grid planning system or urban transportation system at time t. Linear mapping matrix and the upper-level state matrix vector The objective function of the optimization problem at the current moment can be expressed as: .

[0085] Based on the system scenario requirements, the appropriate solution algorithm is selected. By determining whether single-machine computation is feasible, the FISTA algorithm is used to solve the objective function of the optimization problem in an environment that encourages centralized, fast processing. In an environment that encourages distributed collaboration and data locality, the ADMM algorithm is used. Finally, the iterative output of the algorithm is determined as the sparse difference vector at time t. .

[0086] An adaptive regularization parameter selection method based on gradient statistics and residual analysis dynamically adjusts the weights of each regularization term by monitoring gradient changes and reconstruction residuals online during the optimization process, thus obtaining the adjusted regularization parameters. , , and ,Will , , and Stored for use in establishing the objective function for the optimization problem in the next time step.

[0087] At each time t, based on the observed data vector and the upper-level state matrix vector The real-time correlation between nodes is calculated to obtain a sparse correlation matrix. Based on the learned correlation matrix Construct the corresponding time-varying graph Laplacian matrix .

[0088] Obtain the sparse difference vector at time t. Subsequently, the magnitude of its changes is continuously monitored to identify abnormal events. Specifically, the threshold used for judgment is determined based on extreme value theory. First, based on historical data, data preprocessing and stabilization are performed to obtain a historical dataset. Then, based on the stored historical dataset, the excess function is calculated, and candidate threshold intervals are determined by analyzing the average excess map. Finally, the threshold T is determined by fitting a generalized Pareto distribution.

[0089] If the sparse difference vector at time t If the magnitude of the evolutionary difference is less than or equal to the threshold T, it is considered to be within a reasonable range. Otherwise, it is considered to be a transient fluctuation caused by external disturbances such as equipment failure or sudden events, and a control command is sent to the system controller based on the fault detection results to trigger corresponding system adjustment operations. For example, for a power grid planning system, this involves adjusting the topology of the power grid or the operating status of equipment; for an urban transportation network, it involves adjusting the traffic status of various roads.

[0090] In this embodiment, a multi-constraint optimization model integrating sparsity, group sparsity, graph smoothness, and temporal smoothness is first constructed. Then, based on the system scenario requirements, either a centralized FISTA or distributed ADMM algorithm is selected for efficient solution. An adaptive regularization parameter adjustment mechanism based on gradient statistics and residual analysis, along with a dynamic graph structure learning stage, is introduced to achieve online updating of the graph Laplacian matrix. Finally, the differential evolution process is monitored and adaptively thresholded based on extremum theory. Through these methods, sparsity anomalies can be effectively detected and separated, improving the system's robustness and early warning capabilities under dynamic perturbation environments.

[0091] like Figure 3 As shown, this embodiment of the invention also provides a fault detection device 300 based on graph structure differences, comprising: The first acquisition module 301 is used to acquire the observation data vector, linear mapping matrix, and upper-level state matrix vector of the physical system to be monitored at time t. The physical system to be monitored is a power grid planning system or an urban transportation system. The observation data vector is determined by real-time acquisition of sensor data of the physical system to be monitored. The linear mapping matrix and the upper-level state matrix vector are constructed based on the basic structure model, operating parameters, and predictive inputs of the physical system to be monitored, and are used to determine the expected output of the physical system to be monitored. The construction module 302 is used to construct the objective function of the optimization problem at time t based on the observation data vector, linear mapping matrix and upper-level state matrix vector at time t. The solution module 303 is used to solve the objective function of the optimization problem at time t to obtain the sparse difference vector at time t. The sparse difference vector at time t is used to characterize the deviation between the actual operating state and the expected state of the physical system to be monitored. The processing module 304 is used to compare the sparse difference vector at time t with a threshold to obtain a fault detection result, and send a control command to the system controller of the physical system to be monitored based on the fault detection result to trigger the corresponding system adjustment operation. Wherein, the objective function of the optimization problem at time t is Represented as: ; in, Let be the linear mapping matrix at time t. Let be the upper-level state matrix vector at time t. Let be the observation data vector at time t. Let be the sparse difference vector at time t to be solved. for transpose, for Dimensions for The data in the i-th row, Let be the sparse difference vector at time t-1. Let be the graphical Laplace matrix at time t. Used for characterization Norm, Used to characterize the L1 norm, Used to characterize the L2 norm, Used to characterize the trace of a matrix. The first regularization parameter is used. This is the second regularization parameter. This is the third regularization parameter. t is the fourth regularization parameter, where t is a positive integer.

[0092] Optionally, the fault detection device 300 based on graph structure differences further includes: The calculation module is used to calculate the gradient contribution of the target regularization parameter to the objective function of the optimization problem at time t. The target regularization parameter includes a first regularization parameter, a second regularization parameter, a third regularization parameter, and a fourth regularization parameter. Corresponding gradient contribution for: ; An adjustment module is used to adjust the target regularization parameter based on the reconstruction residual and / or the gradient contribution to obtain the adjusted target regularization parameter, which is used to construct the objective function of the optimization problem at time t+1. Wherein, the reconstructed residual for: .

[0093] Optionally, the adjusted target regularization parameter is: ; in, The adjusted target regularization parameters, Let be the target regularization parameter. The gradient vector is composed of the gradient contributions of the first regularization parameter, the second regularization parameter, the third regularization parameter, and the fourth regularization parameter. To reconstruct the residual in the corresponding The result of the change in direction To reconstruct the standard deviation of the residuals, As the first learning rate, As the second learning rate, Used to characterize exponential operations.

[0094] Optionally, the fault detection device 300 based on graph structure differences further includes: The second acquisition module is used to acquire historical datasets. , Let i be the sparse difference vector at time i, and n be the total number of data in the historical dataset. Both i and n are positive integers, and i is less than or equal to n. The first processing module is used to draw an average excess map based on the historical dataset and determine the candidate threshold interval in the average excess map. The second processing module is used to fit a generalized Pareto distribution model to each candidate threshold in the candidate threshold interval, and select the target tail threshold and the corresponding generalized Pareto distribution parameters according to the goodness-of-fit index. The determination module is used to determine the threshold based on the target tail threshold, the corresponding generalized Pareto distribution parameter, and a preset detection performance index.

[0095] Optionally, the fault detection device 300 based on graph structure differences further includes: The learning module is used to learn the correlation matrix based on the observation data vector at time t and the upper-level state matrix vector at time t, using sparse inverse covariance estimation or a mutual information-based method. A construction module is used to construct the graphical Laplacian matrix at time t based on the correlation matrix.

[0096] Optionally, the solution module 303 includes: The judgment unit is used to determine whether the current scene can only be operated by a single machine. The solution unit is used to solve the objective function of the optimization problem using the alternating direction multiplier method when only single-machine operation is possible, and to estimate the sparse difference vector at time t. Otherwise, it uses the fast iterative shrinking threshold algorithm to solve the objective function of the optimization problem and to estimate the sparse difference vector at time t.

[0097] Optionally, when the physical system to be monitored is the power grid planning system, the linear mapping matrix and the upper-level state matrix vector are constructed based on the historical load forecast, network topology connection relationship and power grid equipment parameters of the power grid planning system, the observation data vector is determined based on the collected power grid operation data, and the non-zero elements in the sparse difference vector at time t are used to characterize the deviation between the actual power grid operation state at time t and the expected operation state of the power grid planning system. When the physical system to be monitored is the urban transportation system, the linear mapping matrix and the upper-level state matrix vector are constructed based on the road network structure, traffic assignment model and travel demand prediction of the urban transportation system. The observation data vector is determined based on the data collected by the full road network detector. The non-zero elements in the sparse difference vector at time t are used to characterize the deviation between the actual traffic state and the expected traffic state of the urban transportation system.

[0098] The fault detection device 300 based on graph structure differences provided in this application embodiment can execute the above method embodiment, and its implementation principle and technical effect are similar, so it will not be described again here.

[0099] It should be noted that the division of units in the embodiments of this application is illustrative and only represents one logical functional division. In actual implementation, other division methods may be used. Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units.

[0100] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a processor-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0101] like Figure 4 As shown, this application provides an electronic device 400, including: a memory 402, a processor 401, and a program stored in the memory 402 and executable on the processor 401; the processor 401 is used to read the program in the memory 402 to implement the steps in the fault detection method based on graph structure differences as described above.

[0102] This application also provides a readable storage medium storing a program. When the program is executed by a processor, it implements the various processes of the above-described fault detection method embodiment based on graph structure differences and achieves the same technical effect. To avoid repetition, it will not be described again here. The readable storage medium can be any available medium or data storage device that the processor can access, including but not limited to magnetic storage (such as floppy disks, hard disks, magnetic tapes, magneto-optical disks (MOs), etc.), optical storage (such as compact disks (CDs), digital video discs (DVDs), Blu-ray discs (BDs), high-definition universal discs (HVDs), etc.), and semiconductor storage (such as read-only memory (ROMs), erasable programmable read-only memory (EPROMs), electrically erasable programmable read-only memory (EEPROMs), non-volatile memory (NAND flash), solid-state drives (SSDs)).

[0103] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0104] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, disk, optical disk) and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0105] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other modifications under the guidance of this application without departing from its spirit, and all of these modifications are within the scope of protection of this application.

Claims

1. A fault detection method based on graph structure differences, characterized in that, include: The observation data vector, linear mapping matrix, and upper-level state matrix vector of the physical system to be monitored at time t are obtained. The physical system to be monitored is a power grid planning system or an urban transportation system. The observation data vector is determined by real-time acquisition of sensor data of the physical system to be monitored. The linear mapping matrix and the upper-level state matrix vector are constructed based on the basic structural model, operating parameters, and predictive inputs of the physical system to be monitored, and are used to determine the expected output of the physical system to be monitored. Based on the observation data vector, linear mapping matrix, and upper-level state matrix vector at time t, construct the objective function for the optimization problem at time t; By solving the objective function of the optimization problem at time t, a sparse difference vector at time t is obtained. The sparse difference vector at time t is used to characterize the deviation between the actual operating state and the expected state of the physical system to be monitored. The sparse difference vector at time t is compared with a threshold to obtain the fault detection result. Based on the fault detection result, a control command is sent to the system controller of the physical system to be monitored to trigger the corresponding system adjustment operation. Wherein, the objective function of the optimization problem at time t is Represented as: ； in, Let be the linear mapping matrix at time t. Let be the upper-level state matrix vector at time t. Let be the observation data vector at time t. Let be the sparse difference vector at time t to be solved. for transpose, for Dimensions for The data in the i-th row, Let be the sparse difference vector at time t-1. Let be the graphical Laplace matrix at time t. Used for characterization Norm, Used to characterize the L1 norm, Used to characterize the L2 norm, Used to characterize the trace of a matrix. The first regularization parameter is used. This is the second regularization parameter. This is the third regularization parameter. t is the fourth regularization parameter, where t is a positive integer.

2. The fault detection method based on graph structure differences according to claim 1, characterized in that, After solving the objective function of the optimization problem at time t to obtain the sparse difference vector at time t, the method further includes: The gradient contribution of the objective regularization parameter to the objective function of the optimization problem at time t is calculated. The objective regularization parameter includes a first regularization parameter, a second regularization parameter, a third regularization parameter, and a fourth regularization parameter. Corresponding gradient contribution for: ； Based on the reconstructed residual and / or the gradient contribution, the target regularization parameter is adjusted to obtain the adjusted target regularization parameter, which is used to construct the objective function of the optimization problem at time t+1. Wherein, the reconstructed residual for: 。 3. The fault detection method based on graph structure differences according to claim 2, characterized in that, The adjusted target regularization parameter is: ； in, The adjusted target regularization parameters, Let be the target regularization parameter. The gradient vector is composed of the gradient contributions of the first regularization parameter, the second regularization parameter, the third regularization parameter, and the fourth regularization parameter. To reconstruct the residual in the corresponding The result of the change in direction To reconstruct the standard deviation of the residuals, As the first learning rate, As the second learning rate, Used to characterize exponential operations.

4. The fault detection method based on graph structure differences according to claim 1, characterized in that, Before comparing the sparse difference vector at time t with a threshold to obtain the fault detection result, the method further includes: Obtain historical datasets , Let i be the sparse difference vector at time i, and n be the total number of data in the historical dataset. Both i and n are positive integers, and i is less than or equal to n. An average excess graph is plotted based on the historical dataset, and a candidate threshold interval is determined based on the average excess graph. For each candidate threshold in the candidate threshold interval, a generalized Pareto distribution model is fitted, and the target tail threshold and the corresponding generalized Pareto distribution parameters are selected according to the goodness-of-fit index. The threshold is determined based on the target tail threshold, the corresponding generalized Pareto distribution parameters, and the preset detection performance indicators.

5. The fault detection method based on graph structure differences according to claim 1, characterized in that, Before constructing the objective function for the optimization problem at time t based on the observation data vector, linear mapping matrix, and upper-level state matrix vector at time t, the method further includes: Based on the observation data vector at time t and the upper-level state matrix vector at time t, the correlation matrix is ​​learned using sparse inverse covariance estimation or a method based on mutual information. Construct the graphical Laplace matrix for time t based on the correlation matrix.

6. The fault detection method based on graph structure differences according to claim 1, characterized in that, The process of solving the objective function of the optimization problem at time t to obtain the sparse difference vector at time t includes: Determine whether the current scenario can only be processed by a single machine; In the case where only single-machine computation is possible, the objective function of the optimization problem is solved using the alternating direction multiplier method to estimate the sparse difference vector at time t; otherwise, the fast iterative shrinking threshold algorithm is used to solve the objective function of the optimization problem to estimate the sparse difference vector at time t.

7. The fault detection method based on graph structure differences according to claim 1, characterized in that, When the physical system to be monitored is the power grid planning system, the linear mapping matrix and the upper-level state matrix vector are constructed based on the historical load forecast, network topology connection relationship and power grid equipment parameters of the power grid planning system. The observation data vector is determined based on the collected power grid operation data. The non-zero elements in the sparse difference vector at time t are used to characterize the deviation between the actual power grid operation state at time t and the expected operation state of the power grid planning system. When the physical system to be monitored is the urban transportation system, the linear mapping matrix and the upper-level state matrix vector are constructed based on the road network structure, traffic assignment model and travel demand prediction of the urban transportation system. The observation data vector is determined based on the data collected by the full road network detector. The non-zero elements in the sparse difference vector at time t are used to characterize the deviation between the actual traffic state and the expected traffic state of the urban transportation system.

8. A fault detection device based on graph structure differences, characterized in that, include: The first acquisition module is used to acquire the observation data vector, linear mapping matrix, and upper-level state matrix vector of the physical system to be monitored at time t. The physical system to be monitored is a power grid planning system or an urban transportation system. The observation data vector is determined by real-time acquisition of sensor data of the physical system to be monitored. The linear mapping matrix and the upper-level state matrix vector are constructed based on the basic structural model, operating parameters, and predictive inputs of the physical system to be monitored, and are used to determine the expected output of the physical system to be monitored. The construction module is used to construct the objective function of the optimization problem at time t based on the observation data vector, linear mapping matrix, and upper-level state matrix vector at time t. The solution module is used to solve the objective function of the optimization problem at time t to obtain the sparse difference vector at time t. The sparse difference vector at time t is used to characterize the deviation between the actual operating state and the expected state of the physical system to be monitored. The processing module is used to compare the sparse difference vector at time t with a threshold to obtain the fault detection result, and send a control command to the system controller of the physical system to be monitored based on the fault detection result to trigger the corresponding system adjustment operation. Wherein, the objective function of the optimization problem at time t is Represented as: ； in, Let be the linear mapping matrix at time t. Let be the upper-level state matrix vector at time t. Let be the observation data vector at time t. Let be the sparse difference vector at time t to be solved. for transpose, for Dimensions for The data in the i-th row, Let be the sparse difference vector at time t-1. Let be the graphical Laplace matrix at time t. Used for characterization Norm, Used to characterize the L1 norm, Used to characterize the L2 norm, Used to characterize the trace of a matrix. The first regularization parameter is used. This is the second regularization parameter. This is the third regularization parameter. t is the fourth regularization parameter, where t is a positive integer.

9. An electronic device, comprising: A memory, a processor, and a program stored in the memory and executable on the processor; characterized in that the processor is configured to read the program from the memory to implement the steps in the fault detection method based on graph structure differences as described in any one of claims 1 to 7.

10. A readable storage medium for storing a program, characterized in that, When the program is executed by the processor, it implements the steps in the fault detection method based on graph structure differences as described in any one of claims 1 to 7.

Images