Fault detection method and system for power distribution cabinet

Abstract

The invention provides a fault detection method and system for a power distribution cabinet, and the method comprises the steps: combining the data of an internal sensor of the power distribution cabinet with spatial distribution to generate a dynamic topological graph, and taking the dynamic topological graph as a global state expression; extracting spatio-temporal features related to the fault based on the dynamic topological graph, compressing complexity and forming a feature sequence; constructing a multi-branch time model framework based on the feature sequence to predict the fault of the power distribution cabinet, and performing risk assessment; and judging whether to trigger an alarm or not according to a prediction result, and updating the dynamic topological graph through a maintenance result to form a closed loop. Multi-dimensional data of a sensor network in the power distribution cabinet are combined with time sequence modeling through dynamic topology analysis, and real-time fault detection and trend prediction are achieved.

Description

Technical Field

[0001] The present invention belongs to the field of fault detection, and in particular relates to a fault detection method and system for a power distribution cabinet. Background Art

[0002] With the continuous development of modern power systems, the operation status of distribution cabinets, as an important part of power equipment, directly affects the stability and safety of the power grid. However, due to load fluctuations, environmental influences (such as humidity and temperature changes) and aging of components during long-term operation, faults such as overheating, short circuits, and poor contact are prone to occur inside the distribution cabinet. These faults usually manifest as hidden changes in the early stages and are difficult to detect in time through traditional detection methods, which can easily lead to the accumulation of faults and cause serious consequences.

[0003] The current mainstream fault detection methods have the following technical bottlenecks:

[0004] Lack of real-time performance: Single-point sensor monitoring technology relies on the monitoring of a single specific data (such as temperature or current), which requires manual analysis to determine the fault. It is not only inefficient, but also difficult to meet the demand for rapid response in the operation of the distribution cabinet.

[0005] Lack of predictive capability: Existing methods mainly focus on static fault detection and fail to capture the temporal evolution of faults. Due to the lack of modeling of the temporal characteristics of multi-sensor data, existing systems are unable to provide reliable early fault warnings.

[0006] Insufficient global correlation: There are complex interactions between the components inside the distribution cabinet. For example, overheating may affect bus ducts, cables, circuit breakers and other components at the same time. However, current detection methods fail to establish dynamic correlation analysis between these components, resulting in low accuracy in fault location and cause analysis.

[0007] In response to these problems, how to develop a system that can detect faults in real time and predict their future trends, especially a technology that can combine the dynamic correlation between various components, has become a difficult problem that needs to be solved urgently in the field of distribution cabinet fault detection. Summary of the invention

[0008] The purpose of the present invention is to propose a fault detection method and system for distribution cabinets. By modeling the spatiotemporal series, dynamic topology analysis is combined with time series modeling, which not only solves the real-time problem of the traditional method, but also realizes early fault prediction and global correlation analysis, which has important application value and technical advantages.

[0009] In order to achieve the above object, a first aspect of the present invention provides a fault detection method for a power distribution cabinet, the method comprising:

[0010] S1. Generate a dynamic topology diagram by combining the internal sensor data of the power distribution cabinet with its spatial distribution as a global state expression; wherein the internal sensor data of the power distribution cabinet includes temperature T i 、Current I i and voltage V i , i is the sensor number, i=1,2,…,N;

[0011] S2, extracting fault-related spatiotemporal features based on the dynamic topology graph, compressing the complexity and forming a feature sequence;

[0012] S3. Construct a multi-branch time model framework based on the feature sequence to predict the failure of the distribution cabinet and perform risk assessment;

[0013] S4, judging whether to trigger an alarm based on the prediction results, and updating the dynamic topology map through the maintenance results to form a closed loop;

[0014] The S1 specifically includes:

[0015] The sensor nodes are grouped into a set V = {1, 2, ..., N}, and the initial topology graph G = 9V, E is defined ) , where E is the connection edge between sensor nodes, and the weight w of each edge (i, j)∈E ij Initialized to:

[0016]

[0017] Where d(i,j) is the physical distance between sensors i and j; the initial topological graph is used to capture the spatial relationship of nodes;

[0018] Dynamically adjust the edge weight w according to the real-time collected sensor data ij , so that it can reflect the dynamic coupling relationship between components:

[0019] w ij,t =α·w ij +β·|ΔP i,t -ΔP j,t | -1

[0020] Among them, w ij,t is the dynamic weight of edge (i, j) at time t; w ij is the static edge weight at initialization; ΔP i,t is the parameter change of sensor i at time t, α and β are edge weight adjustment factors, satisfying α+β=1; |ΔP i,t -ΔP j,t | is the difference in state changes between sensors i and j;

[0021] Sparse processing is used to reduce noise and computational complexity, and only the k edges with the largest weight are retained, and k≤N, N is the total number of sensors, then the sparse graph is G t =(V,E t ),in Each edge (i,j)∈E t The weight w ij,t After updating, it is used to reflect the real-time dynamic association between sensors;

[0022] The final generated dynamic topology graph G t =(V,E t )include:

[0023] The node set V represents the sensor location;

[0024] Edge set E t Describe the dynamic relationship between sensors, edge weight w ij,t Reflects the coupling characteristics of component status.

[0025] Preferably, the parameter change ΔP of the sensor i at time t i,t ΔP i,t =[ΔT i,t ,ΔI i,t ,ΔV i,t ], where ΔT i,t is the temperature change in degrees Celsius; ΔI i,t is the current change in amperes; ΔV i,t is the voltage change in volts.

[0026] Preferably, the S2 specifically includes:

[0027] According to the generated dynamic topology graph G t =(V,E t ), extract the spatiotemporal feature embedding vector h of each sensor node i,t ,capture the dynamic relationship between nodes, while ensuring that the generated embedding vector can express the potential information of the fault;

[0028] The spatiotemporal features of each sensor node are embedded into the vector h i,t Aggregate into the global topological feature vector H at time t t , indicating the operating status of the entire power distribution cabinet at the current moment;

[0029] The generated global eigenvector H t Perform regularization to obtain the global eigenvector H′ after regularization t , ensuring that the feature sequences at different time points have smooth temporal continuity;

[0030] The regularized global eigenvector H′ of all time points t Combine into feature sequence

[0031] Preferably, the dynamic topology graph G generated according to t =(V,E t ), extract the spatiotemporal feature embedding vector h of each sensor node i,t , using dynamic feature embedding, expressed as:

[0032]

[0033] Among them, h i,t is the embedding vector of sensor i at time t, which is used to characterize the fault state of the node; σ(·) is a nonlinear activation function, which is used to limit the value range of the embedding and enhance the nonlinear expression ability of the model; is the neighbor set of node i; w ij,t is the edge weight in the dynamic topology graph, indicating the association strength between sensors i and j; f(x j ,x i ) is the node state difference function; λ· It is a dynamic change regularization term, which is used to constrain the sensitivity of the embedding vector to drastic changes in weights and is suitable for capturing fast-changing dynamic faults.

[0034] The spatiotemporal features of each sensor node are embedded into the vector h i,t Aggregate into the global topological feature vector H at time t t , expressed as:

[0035]

[0036] Among them, H t is the global feature vector at time t, which represents the overall operating status of the distribution cabinet; γ i The node importance coefficient is assigned a weight according to the physical location or device type of the sensor;

[0037] The regularization process is expressed as:

[0038]

[0039] Among them, H′ t is the regularized global eigenvector, eliminating unnecessary fluctuations in the feature sequence; is the gradient of the feature vector at time t, which is used to capture the drastic changes of features in time; η is the smoothing factor, which controls the strength of feature regularization;

[0040] The characteristic sequence Expressed as

[0041] Preferably, the node state difference function is defined as:

[0042] f(x j ,x i )=[|T j -T i |,|I j -I i |,|V j -V i |]

[0043] Among them, T j ,T i is the temperature value of nodes j and i; I j ,I i is the current value; V j ,V i is the voltage value.

[0044] Preferably, the S3 specifically includes:

[0045] According to the global feature {H′ after regularization t}Calculate dynamic residual features {R t}, combined into the final combined feature X t =[H′ t ,R t ], where X t is the input feature vector at time t, which contains basic features and dynamic residuals, and fully describes the state and changes;

[0046] The final combined feature is X t =[H′ t ,R t ] to form an input sequence Then, a multi-branch time model framework based on distribution cabinet fault detection is constructed for processing, and a feature prediction sequence is output;

[0047] Multi-task failure risk assessment is performed based on feature prediction sequences, including prediction of potential failure locations, failure types, and occurrence times.

[0048] Preferably, the multi-branch time model framework is constructed as follows:

[0049] Constructing a long-term trend branch and a short-term fluctuation branch; wherein the long-term trend branch is used to capture the slowly changing pattern of the operating state of the power distribution cabinet; and the short-term fluctuation branch is used to capture the pattern of rapid changes and sudden failures;

[0050] The advancement relationship of the multi-branch time model framework is constructed as follows:

[0051]

[0052] in, Predict the feature sequence for the next k time steps; F trend (X t ) is the long-term trend branch output, used to describe the slow fault trend; F short (X t ) is the short-term fluctuation branch output, focusing on fast fault detection;

[0053] The multi-task failure risk assessment specifically includes:

[0054] Define the risk scoring function and perform risk analysis for each time step:

[0055]

[0056] Among them, Risk t+k is the risk prediction value at time t+k, including the probability distribution of risk level; W risk is the weight matrix for risk assessment; is a regularization term, which is used to measure the deviation of future prediction values ​​from the current state and strengthen the detection of abnormal changes.

[0057] Preferably, the long-term trend branch extracts the slowly changing pattern of the input sequence based on a temporal convolutional network, expressed as:

[0058] F trend (X t) =ReLU(W trend *X t )

[0059] Among them, W trend is the temporal convolution kernel weight, which is used to extract the temporal pattern of a fixed window; ReLU is a nonlinear activation function, which is used to enhance the sparsity of features;

[0060] The short-term fluctuation branch adopts a dynamic residual network structure to strengthen the modeling of residual features, which is expressed as:

[0061] F short (X t )=λ·R t +g(W short ·R t )

[0062] Among them, λ is the dynamic amplification factor, which is used to enhance the sensitivity to residual errors; g(W short ·R t ) is a nonlinear activation used to capture complex fluctuation patterns.

[0063] Preferably, the S4 specifically includes:

[0064] Design a comprehensive alarm index A based on the risk prediction value t , used to measure the overall failure risk level of the distribution cabinet in the future period; wherein the comprehensive alarm index A t , expressed as:

[0065]

[0066] Among them, W alarm is the risk weight vector, which assigns weights to risk levels; Dynamic change term, used to amplify outliers of future feature changes; Indicates the difference between the future feature prediction and the current state; λ is the amplification factor, which enhances the sensitivity to significant changes;

[0067] If A t When the set threshold τ is exceeded, an alarm signal is triggered:

[0068]

[0069] Wherein, Alarm is the alarm signal, 1 indicates triggering an alarm, and 0 indicates no warning; τ is the alarm threshold, which is dynamically adjusted according to the environment and operation strategy; for the moment when the alarm is triggered, an alarm report is automatically generated; wherein, the content includes fault type prediction, fault location prediction and risk time range;

[0070] When an alarm is triggered, the system's dynamic topology map and the parameters of the multi-branch time model framework are updated in real time to ensure that they adapt to the latest status changes.

[0071] In a second aspect of the present invention, a fault detection system for a power distribution cabinet is provided, the system comprising:

[0072] A dynamic topology map construction module is used to generate a dynamic topology map by combining the internal sensor data of the distribution cabinet with its spatial distribution as a global state expression; wherein the internal sensor data of the distribution cabinet includes temperature T i 、Current I i and voltage V i , i is the sensor number, i=1,2,…,N;

[0073] The spatiotemporal feature construction module is used to extract fault-related spatiotemporal features based on the dynamic topology graph, compress the complexity and form a feature sequence;

[0074] The risk prediction module is used to build a multi-branch time model framework based on the feature sequence to predict the failure of the distribution cabinet and perform risk assessment;

[0075] The detection, warning and optimization module is used to determine whether to trigger an alarm based on the prediction results, and to update the dynamic topology map through maintenance results to form a closed loop;

[0076] The dynamic topology map construction module specifically includes:

[0077] The sensor nodes are organized into a set V = {1, 2, ..., N}, and the initial topology graph G = 9V, E) is defined, where E is the connection edge between the sensor nodes, and the weight w of each edge (i, j) ∈ E ij Initialized to:

[0078]

[0079] Where d(i,j) is the physical distance between sensors i and j; the initial topological graph is used to capture the spatial relationship of nodes;

[0080] According to the real-time sensor data collected, the edge weight w is dynamically adjusted ij , so that it can reflect the dynamic coupling relationship between components:

[0081]

[0082] Among them, w ij,t is the dynamic weight of edge (i, j) at time t; w ij is the static edge weight at initialization; ΔP i,t is the parameter change of sensor i at time t, α and β are edge weight adjustment factors, satisfying α+β=1; |ΔP i,t -ΔP j,t | is the difference in state changes between sensors i and j;

[0083] Sparse processing is used to reduce noise and computational complexity, and only the k edges with the largest weight are retained, and k≤N, N is the total number of sensors, then the sparse graph is G t =(V,E t ),in Each edge (i,j)∈E t The weight w ij,t After updating, it is used to reflect the real-time dynamic association between sensors;

[0084] The final generated dynamic topology graph G t =(V,E t )include:

[0085] The node set V represents the sensor location;

[0086] Edge set E t Describe the dynamic relationship between sensors, edge weight w ij,t Reflects the coupling characteristics of component status.

[0087] The beneficial technical effects of the present invention are at least as follows:

[0088] The present invention captures the spatial correlation between components by constructing a dynamic topology map of the sensor network inside the power distribution cabinet. During operation, when a component is abnormal (such as temperature rise or increased electromagnetic interference), dynamic topology modeling can analyze the propagation path and impact range of the abnormality on the entire system in real time, providing support for rapid fault location.

[0089] The present invention uses particle swarm optimization (PSO) algorithm combined with real-time monitoring data to perform multi-objective optimization on the operating parameters of the screening equipment (such as vibration frequency, mesh size, and inclination angle). By introducing a comprehensive evaluation function of screening efficiency and energy consumption, the dynamic adjustment of the equipment operating parameters is achieved, solving the problem that the existing equipment cannot adapt to changes in mineral flow characteristics.

[0090] Based on dynamic topology, the present invention uses a temporal convolutional network (TCN) to model the time evolution characteristics of multidimensional sensor data and extract potential abnormal trend signals, thereby predicting the evolution direction of faults and the time when they may occur.

[0091] The present invention combines the multi-dimensional data of the sensor network inside the power distribution cabinet through dynamic topology analysis and time series modeling to achieve real-time detection and trend prediction of faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0093] Figure 1 The present invention is a flow chart of a fault detection method for a power distribution cabinet.

[0094] Figure 2 The present invention is a framework diagram of a fault detection system for a power distribution cabinet. DETAILED DESCRIPTION

[0095] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0096] like Figure 1 As shown, an embodiment of the present invention provides a fault detection method for a power distribution cabinet, the method comprising the following steps:

[0097] S1. Generate a dynamic topology diagram by combining the internal sensor data of the power distribution cabinet with its spatial distribution as a global state expression; wherein the internal sensor data of the power distribution cabinet includes temperature T i 、Current I i and voltage V i , i is the sensor number, i=1,2,…,N.

[0098] The specific implementation method of this step includes:

[0099] The sensor network collects key physical parameters inside the power distribution cabinet, including temperature T i 、Current I i and voltage V i , where i is the sensor number, i = 1, 2, ..., N. These sensor nodes form a set V = {1, 2, ..., N}, and define an initial topology graph G = (V, E), where E is the connection edge between sensor nodes.

[0100] Among them, the weight w of each edge (i, j)∈E ij Initialized to:

[0101]

[0102] Where d(i,j) is the physical distance between sensors i and j (in meters). The initial topology graph is used to capture the spatial relationship of nodes.

[0103] Furthermore, the edge weight w is dynamically adjusted according to the real-time collected sensor data. ij , so that it can reflect the dynamic coupling relationship between components:

[0104] w ij,t =α·w ij +β·|ΔP i,t -ΔP j,t | -1

[0105] Among them, w ij,t is the dynamic weight of edge (i, j) at time t. ij is the static edge weight at initialization (no time change). ΔP i,t =[ΔT i,t ,ΔI i,t ,ΔV i,t ] is the parameter change of sensor i at time t, where ΔT i,t is the temperature change in degrees Celsius; ΔI i,t is the current change in amperes; ΔV i,t is the voltage change in volts. α and β are edge weight adjustment factors, satisfying α+β=1. |ΔP i,t-ΔP j,t | is the difference in state change between sensors i and j (vector norm).

[0106] Furthermore, in order to reduce noise and computational complexity, sparse processing is used to retain only the k edges with the largest weights (k≤N). The sparse graph is G t =(V,E t ),in Each edge (i,j)∈E t The weight w ij,t After updating, it is used to reflect the real-time dynamic association between sensors. This thinning step selects the edges with the largest weight by sorting and retains important topological relationships.

[0107] Furthermore, the generated dynamic topology graph G t =(V,E t )include:

[0108] The node set V represents the sensor location;

[0109] Edge set E t Describe the dynamic relationship between sensors, edge weight w ij,t Reflects the coupling characteristics of component status.

[0110] The dynamic topology graph G obtained by the above steps t It will serve as the input of step 2 to extract topological features.

[0111] S2. Extract fault-related spatiotemporal features based on the dynamic topology graph, compress the complexity and form a feature sequence.

[0112] The specific implementation method of this step includes:

[0113] According to the dynamic topology graph G generated in step 1 t =(V,E t ), extract the spatiotemporal feature embedding vector h of each sensor node i,t The purpose of this step is to use the topological structure and the state changes of the sensor to capture the dynamic relationship between nodes, while ensuring that the generated embedding vector can express the potential information of the fault. The dynamic feature embedding formula is as follows:

[0114]

[0115] Among them, h i,t is the embedding vector of sensor i at time t, which is used to characterize the fault state of the node. σ(·) is a nonlinear activation function, which is used to limit the value range of the embedding and enhance the nonlinear expression ability of the model. is the neighbor set of node i. ij,tis the edge weight in the dynamic topology graph of step 1, indicating the association strength between sensors i and j. j ,x i ) is the node state difference function, defined as:

[0116] f(x j ,x i )=[|T j -T i |,|I j -I i |,|V j -V i |]

[0117] Among them, T j ,T i is the temperature value of nodes j and i; I j ,I i is the current value; V j ,V i is the voltage value. is a dynamic change regularization term, which is used to constrain the sensitivity of the embedding vector to drastic changes in weights and is suitable for capturing fast-changing dynamic faults.

[0118] Furthermore, the formula dynamically changes the regularization term λ·|w ij,t -w ij,t-1 |Captures the changes in weights between sensors, which is particularly suitable for sudden fault scenarios in distribution cabinets (such as short circuits or local overheating). Embedding vector h i,t is the dynamic weight w ij,t and node state change f(x j ,x i ) directly reflects the local and global characteristics of the operating status of the distribution cabinet.

[0119] Furthermore, the embedding vector h of each node is i,t Aggregate into the global topological feature vector H at time t t , indicating the operating status of the entire distribution cabinet at that moment.

[0120]

[0121] Among them, H t is the global feature vector at time t, which represents the overall operating status of the distribution cabinet. i It is the node importance coefficient, which is weighted according to the physical location or device type of the sensor (for example, the bus duct sensor has a greater weight than the corner sensor).

[0122] Furthermore, the polymerization process is carried out by γ iThe introduction of H reflects the difference in importance of different components in the distribution cabinet, for example, focusing on equipment with higher temperature rise (such as circuit breakers). t is for all embedding vectors h i,t The weighted average of , while retaining the global information of the dynamic topology.

[0123] Furthermore, the generated global feature vector H t Regularization is performed to ensure that the feature sequences at different time points have smooth temporal continuity, thus facilitating subsequent time series modeling:

[0124]

[0125] Among them, H′ t is the regularized global eigenvector, which eliminates unnecessary fluctuations in the feature sequence. is the gradient of the feature vector at time t, which is used to capture the drastic changes of features in time. η is the smoothing factor, which controls the strength of feature regularization.

[0126] Furthermore, the time gradient regularization term The introduction of solves the problem of characteristic fluctuation of sensor data in the distribution cabinet caused by noise and improves the stability of the characteristic sequence.

[0127] Furthermore, the regularized global feature vector H′ at all time points t Combine into feature sequence As input for time series modeling:

[0128]

[0129] The output feature sequence completely characterizes the dynamic evolution process of the operating status of the distribution cabinet, can capture time correlation and potential trends, and provide a basis for the next step of time modeling.

[0130] This step generates From the dynamic topology graph G t The extracted feature sequence directly takes over the output of step 1. Through innovative designs such as dynamic regularization terms and feature smoothing, the feature sequence is ensured to have good temporal and spatial correlation and stability, which is particularly suitable for fault detection and prediction in complex and changeable environments of distribution cabinets.

[0131] S3. Based on the feature sequence, a multi-branch time model framework is constructed to predict the failure of the distribution cabinet and perform risk assessment.

[0132] The specific implementation method of this step includes:

[0133] According to the regularized feature sequence output in step 2 Construct input data suitable for time series modeling. Considering the complexity of the distribution cabinet operating environment and possible abnormal trends, it is necessary to construct more representative and robust input features.

[0134] The input features are divided into two parts:

[0135] Basic global features {H′ t}, indicating the overall operating status of the power distribution cabinet at each time point;

[0136] Dynamic residual feature {R t}, by comparing the state changes at adjacent time points to capture abnormal trends:

[0137] R t =H′ t -H′ t-1

[0138] Among them, R t It is the residual feature vector at time t, which indicates the change between the current state and the state at the previous moment, and is suitable for capturing the mutation mode caused by the fault. t ,H′ t-1 are the regularized global feature vectors at time t and t-1, respectively.

[0139] The final combined features are:

[0140] X t =[H′ t ,R t ]

[0141] Among them, X t It is the input feature vector at time t, which contains basic features and dynamic residuals, and fully describes the state and changes.

[0142] Furthermore, the input sequence The problem is solved by a multi-branch temporal modeling framework designed for fault detection in power distribution cabinets. The framework consists of two parts:

[0143] Long-term trend branch: captures slowly changing patterns in the operating status of the distribution cabinet, suitable for gradually evolving faults (such as overtemperature or equipment aging).

[0144] Short-term fluctuation branch: focuses on fast changes and sudden faults (such as short circuits or arcing phenomena).

[0145] The core recursive relationship of the combination model is as follows:

[0146]

[0147] in, Predict the sequence of features for the next k time steps. trend (Xt ) is the long-term trend branch output, used to describe the slow fault trend. short (X t ) is the short-term fluctuation branch output, focusing on fast fault detection.

[0148] Further, the detailed calculation is:

[0149] Long-term trend branches:

[0150] The long-term trend branch is based on the Temporal Convolutional Network (TCN), which extracts the slowly changing patterns of the input sequence:

[0151] F trend (X t )=ReLU(W trend *X t )

[0152] Among them, W trend is the temporal convolution kernel weight, which is used to extract the temporal pattern of a fixed window. ReLU is a nonlinear activation function, which is used to enhance the sparsity of features.

[0153] Short-term volatility branch:

[0154] The short-term volatility branch adopts the Dynamic Residual Network (DRN) structure to strengthen the modeling of residual features:

[0155] F short (X t )=λ·R t +g(W short ·R t )

[0156] Among them, λ is the dynamic amplification factor, which is used to enhance the sensitivity to residual errors; g(W short ·R t ) is a nonlinear activation used to capture complex fluctuation patterns.

[0157] Furthermore, by predicting the future feature sequence Perform multi-task failure risk assessment, including prediction of failure risk level, type, and location: Define a risk scoring function and perform risk analysis for each time step:

[0158]

[0159] Among them, Risk t+k is the risk prediction value at time t+k, including the probability distribution of risk level. risk is the weight matrix for risk assessment. is a regularization term, which is used to measure the deviation of the future predicted value from the current state and strengthen the detection of abnormal changes.

[0160] According to the risk scoring function, the following information is output:

[0161] Potential fault location: Determine the specific sensor or device by combining node characteristics and high risk scores;

[0162] Fault type: classified into overheating, short circuit, arc, etc. based on the change pattern of the predicted characteristics;

[0163] Occurrence time: Outputs the predicted time range through high risk scoring time points.

[0164] Furthermore, the prediction results are output, which include the following two parts:

[0165] Feature prediction sequence for the future time period t+1 to t+k Used for system status trend analysis;

[0166] Multi-task risk assessment results Risk t+1:t+k , providing a basis for the alarm and feedback optimization in step 4.

[0167] S4. Determine whether to trigger an alarm based on the prediction results, and update the dynamic topology map through the maintenance results to form a closed loop.

[0168] The specific implementation method of this step includes:

[0169] Based on the future risk assessment results output in step 3, Risk t+1:t+k , design comprehensive alarm index A t , which is used to measure the overall failure risk level of the power distribution cabinet in the future period. In order to meet the needs of complex environments, special items are introduced to enhance the sensitivity to potential abnormalities. The alarm indicator formula is:

[0170]

[0171] Among them, A t is the comprehensive alarm index at time t. t+k is the risk prediction result output from step 3, which represents the risk value at the next k moments. alarm is the risk weight vector, which assigns weights to risk levels. It is a dynamic change term, which is used to amplify the abnormal values ​​of future feature changes. It represents the difference between the future feature prediction and the current state. λ is the amplification factor, which increases the sensitivity to significant changes.

[0172] If A t When the set threshold τ is exceeded, an alarm signal is triggered:

[0173]

[0174] Among them, Alarm is the alarm signal, 1 means triggering an alarm, and 0 means no warning. τ is the alarm threshold, which is dynamically adjusted according to the environment and operation strategy.

[0175] It is understandable that by introducing dynamic changes Strengthen the ability to respond to drastic changes in future conditions. alarm The risk weight design enables the system to pay more attention to high-risk level failures.

[0176] Furthermore, when an alarm is triggered, the system automatically generates an alarm report, which includes:

[0177] Failure type prediction: By analyzing Risk t+k The class probability is used to determine the most likely fault type (such as overheating, short circuit, or arc).

[0178] Fault location prediction: Based on For nodes with higher medium risk scores, locate possible faulty sensors and associated components.

[0179] Risk time range: Provides the future high risk time range [t+k min ,t+k max ] to facilitate maintenance decisions.

[0180] Report format example: Alarm trigger time: t Fault type: Short circuit Fault location: Node 5 (bus duct) Predicted risk time range: t+3 ​​to t+6

[0181] Furthermore, after the alarm is triggered, the dynamic topology map G of the system is updated in real time. t and model parameters to ensure that they adapt to the latest state changes. The specific update steps are as follows:

[0182] Sensor data update: Add new sensor data after collecting alarms {T i,t ,I i,t ,V i,t}, recalculate the edge weight w ij,t , generate the updated topology graph G′ t .

[0183] Risk weight optimization: Risk assessment weight W risk Adjustments have been made to add feedback learning items for recent alarms:

[0184] W′ risk =W risk +η·Δ feedback

[0185] Among them, Δfeedback : The feedback increment calculated based on the deviation between the actual alarm and the predicted result. η: Learning rate, which controls the amplitude of weight adjustment.

[0186] It is understandable that the real-time feedback mechanism improves the system's ability to adapt to environmental changes. feedback , so that the model can be gradually optimized after multiple rounds of running, reducing false positives and false negatives.

[0187] Furthermore, the updated topology graph is finally outputted, and the updated topology graph G′ t Pass it back to step 1 as the basic input for the next round of fault detection, forming a complete closed-loop process.

[0188] Furthermore, the alarm mechanism combines characteristic changes and risk levels to achieve rapid response. Dynamic feedback optimization improves the long-term adaptability of the system. The closed-loop process ensures that the system can adapt to the operating status of the distribution cabinet in real time, achieving the real-time and predictive requirements of the patented goal.

[0189] like Figure 2 As shown, an embodiment of the present invention further provides a fault detection system for a power distribution cabinet, the system comprising:

[0190] The dynamic topology map construction module 101 is used to generate a dynamic topology map by combining the internal sensor data of the power distribution cabinet with its spatial distribution as a global state expression; wherein the internal sensor data of the power distribution cabinet includes temperature T i 、Current I i and voltage V i , i is the sensor number, i=1,2,…,N;

[0191] A spatiotemporal feature construction module 102, for extracting fault-related spatiotemporal features based on a dynamic topology graph, compressing the complexity and forming a feature sequence;

[0192] The risk prediction module 103 is used to construct a multi-branch time model framework based on the feature sequence to predict the failure of the power distribution cabinet and perform risk assessment;

[0193] The detection warning and optimization module 104 is used to determine whether to trigger an alarm based on the prediction results, and to update the dynamic topology map through the maintenance results to form a closed loop;

[0194] The dynamic topology map construction module specifically includes:

[0195] The sensor nodes are organized into a set V = {1, 2, ..., N}, and the initial topology graph G = (V, E) is defined, where E is the connection edge between the sensor nodes, and the weight w of each edge (i, j) ∈ E is ij Initialized to:

[0196]

[0197] Where d(i,j) is the physical distance between sensors i and j; the initial topological graph is used to capture the spatial relationship of nodes;

[0198] According to the real-time sensor data collected, the edge weight w is dynamically adjusted ij , so that it can reflect the dynamic coupling relationship between components:

[0199] w ij,t =α·w ij +β·|ΔP i,t -ΔP j,t | -1

[0200] Among them, w ij,t is the dynamic weight of edge (i, j) at time t; w ij is the static edge weight at initialization; ΔP i,t is the parameter change of sensor i at time t, α and β are edge weight adjustment factors, satisfying α+β=1; |ΔP i,t -ΔP j,t | is the difference in state changes between sensors i and j;

[0201] Sparse processing is used to reduce noise and computational complexity, and only the k edges with the largest weight are retained, and k≤N, N is the total number of sensors, then the sparse graph is G t =(V,E t ),in Each edge (i,j)∈E t The weight w ij,t After updating, it is used to reflect the real-time dynamic association between sensors;

[0202] The final generated dynamic topology graph G t =(V,E t )include:

[0203] The node set V represents the sensor location;

[0204] Edge set E t Describe the dynamic relationship between sensors, edge weight w ij,t Reflects the coupling characteristics of component status.

[0205] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0206] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a division of logical functions. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, and the indirect coupling or communication connection of the device or unit can be electrical, mechanical or other forms.

[0207] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can be essentially or partly embodied in the form of a software product that contributes to the prior art. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server or network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program codes.

[0208] Although embodiments of the present invention have been shown and described, those skilled in the art will appreciate that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the claims and their equivalents.

Claims

1. A fault detection method for a power distribution cabinet, characterized in that: The method comprises: S1. Generate a dynamic topology diagram by combining the internal sensor data of the power distribution cabinet with its spatial distribution as a global state expression; wherein the internal sensor data of the power distribution cabinet includes temperature T i 、Current I i and voltage V i , i is the sensor number, i=1,2,…,N; S2, extracting fault-related spatiotemporal features based on the dynamic topology graph, compressing the complexity and forming a feature sequence; S3. Construct a multi-branch time model framework based on the feature sequence to predict the failure of the distribution cabinet and perform risk assessment; S4, judging whether to trigger an alarm based on the prediction results, and updating the dynamic topology map through the maintenance results to form a closed loop; The S1 specifically includes: The sensor nodes are organized into a set V = {1, 2, ..., N}, and the initial topology graph G = (V, E) is defined, where E is the connection edge between the sensor nodes, and the weight w of each edge (i, j) ∈ E is ij Initialized to: Where d(i,j) is the physical distance between sensors i and j; the initial topological graph is used to capture the spatial relationship of nodes; Dynamically adjust the edge weight w according to the real-time collected sensor data ij , so that it can reflect the dynamic coupling relationship between components: w ij,t =α·w ij +β·|ΔP i,t -ΔP j,t | -1 Among them, w ij,t is the dynamic weight of edge (i, j) at time t; w ij is the static edge weight at initialization; ΔP i,t is the parameter change of sensor i at time t, α and β are edge weight adjustment factors, satisfying α+β=1; |ΔP i,t -ΔP j,t | is the difference in state changes between sensors i and j; Sparse processing is used to reduce noise and computational complexity, and only the k edges with the largest weight are retained, and k≤N, N is the total number of sensors, then the sparse graph is G t =(V,E t ),in Each edge (i,j)∈E t The weight w ij,t After updating, it is used to reflect the real-time dynamic association between sensors; The final generated dynamic topology graph G t =(V,E t )include: The node set V represents the sensor location; Edge set E t Describe the dynamic relationship between sensors, edge weight w ij,t Reflects the coupling characteristics of component status.

2. A fault detection method for a power distribution cabinet according to claim 1, characterized in that: The parameter change ΔP of the sensor i at time t i,t ΔP i,t =[ΔT i,t ,ΔI i,t ,ΔV i,t ], where ΔT i,t is the temperature change in degrees Celsius; ΔI i,t is the current change in amperes; ΔV i,t is the voltage change in volts.

3. A fault detection method for a power distribution cabinet according to claim 1, characterized in that: The S2 specifically includes: According to the generated dynamic topology graph G t =(V,E t ), extract the spatiotemporal feature embedding vector h of each sensor node i,t ,capture the dynamic relationship between nodes, while ensuring that the generated embedding vector can express the potential information of the fault; The spatiotemporal features of each sensor node are embedded into the vector h i,t Aggregate into the global topological feature vector H at time t t , indicating the operating status of the entire power distribution cabinet at the current moment; The generated global eigenvector H t Perform regularization to obtain the global eigenvector H′ after regularization t , ensuring that the feature sequences at different time points have smooth temporal continuity; The regularized global eigenvector H at all time points t Combine into feature sequence 4. A fault detection method for a power distribution cabinet according to claim 3, characterized in that: The dynamic topology graph G generated according to t =(V,E t ), extract the spatiotemporal feature embedding vector h of each sensor node i,t , using dynamic feature embedding, expressed as: Among them, h i,t is the embedding vector of sensor i at time t, which is used to characterize the fault state of the node; σ(·) is a nonlinear activation function, which is used to limit the value range of the embedding and enhance the nonlinear expression ability of the model; is the neighbor set of node i; w ij,t is the edge weight in the dynamic topology graph, indicating the association strength between sensors i and j; f(x j ,x i ) is the node state difference function; λ· It is a dynamic change regularization term, which is used to constrain the sensitivity of the embedding vector to drastic changes in weights and is suitable for capturing fast-changing dynamic faults. The spatiotemporal features of each sensor node are embedded into the vector h i,t Aggregate into the global topological feature vector H at time t t , expressed as: Among them, H t is the global feature vector at time t, which represents the overall operating status of the distribution cabinet; γ i The node importance coefficient is assigned a weight according to the physical location or device type of the sensor; The regularization process is expressed as: Among them, H t is the regularized global eigenvector, eliminating unnecessary fluctuations in the feature sequence; is the gradient of the feature vector at time t, which is used to capture the drastic changes of features in time; η is the smoothing factor, which controls the strength of feature regularization; The characteristic sequence Expressed as 5. A fault detection method for a power distribution cabinet according to claim 4, characterized in that: The node state difference function is defined as: f(x j ,x i )=[|T j -T i |,|I j -I i |,|V j -V i |] Among them, T j ,T i is the temperature value of nodes j and i; I j ,I i is the current value; V j ,V i is the voltage value.

6. A fault detection method for a power distribution cabinet according to claim 4, characterized in that: The S3 specifically includes: According to the global feature {H′ after regularization t }Calculate dynamic residual features {R t }, combined into the final combined feature X t =[H′ t ,R t ], where X t is the input feature vector at time t, which contains basic features and dynamic residuals, and fully describes the state and changes; The final combined feature is X t =[H′ t ,R t ] to form an input sequence Then, a multi-branch time model framework based on distribution cabinet fault detection is constructed for processing, and a feature prediction sequence is output; Multi-task failure risk assessment is performed based on feature prediction sequences, including prediction of potential failure locations, failure types, and occurrence times.

7. A fault detection method for a power distribution cabinet according to claim 6, characterized in that: The multi-branch time model framework is constructed as follows: Constructing a long-term trend branch and a short-term fluctuation branch; wherein the long-term trend branch is used to capture the slowly changing pattern of the operating state of the power distribution cabinet; and the short-term fluctuation branch is used to capture the pattern of rapid changes and sudden failures; The advancement relationship of the multi-branch time model framework is constructed as follows: in, Predict the feature sequence for the next k time steps; F trend (X t ) is the long-term trend branch output, used to describe the slow fault trend; F short (X t ) is the short-term fluctuation branch output, focusing on fast fault detection; The multi-task failure risk assessment specifically includes: Define the risk scoring function and perform risk analysis for each time step: Among them, Risk t+k is the risk prediction value at time t+k, including the probability distribution of risk level; W risk is the weight matrix for risk assessment; is a regularization term, which is used to measure the deviation of the future predicted value from the current state and strengthen the detection of abnormal changes.

8. A fault detection method for a power distribution cabinet according to claim 7, characterized in that: The long-term trend branch extracts the slowly changing pattern of the input sequence based on a temporal convolutional network, expressed as: F trend (X t )=ReLU(W trend *X t ) Among them, W trend is the temporal convolution kernel weight, which is used to extract the temporal pattern of a fixed window; ReLU is a nonlinear activation function, which is used to enhance the sparsity of features; The short-term fluctuation branch adopts a dynamic residual network structure to strengthen the modeling of residual features, which is expressed as: F short (X t )=λ·R t +g(W short ·R t ) Among them, λ is the dynamic amplification factor, which is used to enhance the sensitivity to residual errors; g(W short ·R t ) is a nonlinear activation used to capture complex fluctuation patterns.

9. A fault detection method for a power distribution cabinet according to claim 7, characterized in that: The S4 specifically includes: Design a comprehensive alarm index A based on the risk prediction value t , used to measure the overall failure risk level of the distribution cabinet in the future period; wherein the comprehensive alarm index A t , expressed as: Among them, W alarm is the risk weight vector, which assigns weights to risk levels; Dynamic change term, used to amplify outliers of future feature changes; Indicates the difference between the future feature prediction and the current state; λ is the amplification factor, which enhances the sensitivity to significant changes; If A t When the set threshold τ is exceeded, an alarm signal is triggered: Wherein, Alarm is the alarm signal, 1 indicates triggering an alarm, and 0 indicates no warning; τ is the alarm threshold, which is dynamically adjusted according to the environment and operation strategy; for the moment when the alarm is triggered, an alarm report is automatically generated; wherein, the content includes fault type prediction, fault location prediction and risk time range; When an alarm is triggered, the system's dynamic topology map and the parameters of the multi-branch time model framework are updated in real time to ensure that they adapt to the latest status changes.

10. A fault detection system for a power distribution cabinet, characterized in that: The system comprises: A dynamic topology map construction module is used to generate a dynamic topology map by combining the internal sensor data of the distribution cabinet with its spatial distribution as a global state expression; wherein the internal sensor data of the distribution cabinet includes temperature T i 、Current I i and voltage V i , i is the sensor number, i=1,2,…,N; The spatiotemporal feature construction module is used to extract fault-related spatiotemporal features based on the dynamic topology graph, compress the complexity and form a feature sequence; The risk prediction module is used to build a multi-branch time model framework based on the feature sequence to predict the failure of the distribution cabinet and perform risk assessment; The detection, warning and optimization module is used to determine whether to trigger an alarm based on the prediction results, and to update the dynamic topology map through maintenance results to form a closed loop; The dynamic topology map construction module specifically includes: The sensor nodes are organized into a set V = {1, 2, ..., N}, and the initial topology graph G = (V, E) is defined, where E is the connection edge between the sensor nodes, and the weight w of each edge (i, j) ∈ E is ij Initialized to: Where d(i,j) is the physical distance between sensors i and j; the initial topological graph is used to capture the spatial relationship of nodes; Dynamically adjust the edge weight w according to the real-time collected sensor data ij , so that it can reflect the dynamic coupling relationship between components: w ij,t =α·w ij +β|ΔP i,t -ΔP j,t | -1 Among them, w ij,t is the dynamic weight of edge (i, j) at time t; w ij is the static edge weight at initialization; ΔP i,t is the parameter change of sensor i at time t, α and β are edge weight adjustment factors, satisfying α+β=1; |ΔP i,t -ΔP j,t | is the difference in state changes between sensors i and j; Sparse processing is used to reduce noise and computational complexity, and only the k edges with the largest weight are retained, and k≤N, N is the total number of sensors, then the sparse graph is G t =(V,E t ),in Each edge (i,j)∈E t The weight w ij,t After updating, it is used to reflect the real-time dynamic association between sensors; The final generated dynamic topology graph G t =(V,E t )include: The node set V represents the sensor location; Edge set E t Describe the dynamic relationship between sensors, edge weight w ij,t Reflects the coupling characteristics of component status.

Images