Power inspection system and method

Abstract

The present invention relates to the field of intelligent operation and maintenance of power systems and data processing technology, specifically to a power inspection system and method; the method comprises: collecting equipment status data, constructing a third-order state transition tensor field to model the time-series evolution characteristics of the spatial state of the equipment, and identifying short-term high-risk equipment based on a weighted anomaly score and a dynamic threshold strategy; recording historical state peaks through a state-time matrix fusion exponential decay model, calculating the forgetting risk factor in combination with the current level of inactivity, and proactively reviewing equipment that was once severely abnormal but recently inactive; constructing a dynamic evolution graph structure, optimizing inspection paths using a genetic-like perturbation mechanism, and scheduling multiple teams in parallel through task clustering and queue matching; establishing an equipment anomaly linkage propagation model to predict cascade risks, and automatically matching the optimal disposal plan in combination with a policy template library. The present invention forms a closed-loop control from anomaly identification, path optimization, to policy execution, improving the accuracy and timeliness of power operation and maintenance.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent operation and maintenance and data processing of power systems, and in particular to a power inspection system and method. Background Art

[0002] The power inspection sector faces core challenges such as insufficient dynamic status monitoring, lack of a long-term hidden danger tracing mechanism, limited resource scheduling efficiency, and difficulty in quantifying risk transmission across equipment.

[0003] The Chinese invention patent with announcement number CN118734231B discloses a large-scale model-based intelligent inspection system for power plants, which realizes data transmission and processing through communication connections between the server side and multiple inspection devices and the inspected equipment. The invention includes multiple units such as sensor fault detection, drone inspection optimization, data fusion analysis, and data acquisition fault repair. It aims to solve the inspection distortion problem caused by sensor failure or data transmission interference, while improving the inspection stability and accuracy of drones under harsh conditions. It uses the random forest algorithm to build a sensor data model, dynamically adjusts the drone inspection path, and uses deep learning to perform data fusion analysis to generate accurate inspection conclusions and disposal plans. This not only improves the accuracy and efficiency of inspections, but also further enhances the intelligence and adaptability of the system.

[0004] Current methods lack the ability to fine-tune modeling of the spatiotemporal dynamic evolution of equipment status, making it difficult to capture short-term deterioration trends in real time. Historical abnormal events are easily overlooked during quiet periods due to the lack of an active memory mechanism, posing a risk of recurrence. Fixed-path scheduling strategies are unable to collaboratively optimize the real-time status, spatial topology, and resource constraints of equipment, resulting in coverage blind spots and cost waste. At the same time, there is a lack of a predictive model for the transmission paths of abnormalities between devices, which restricts the timeliness and systematic nature of risk management. There is an urgent need for a power inspection solution that integrates dynamic state quantification, long-term memory enhancement, intelligent scheduling deduction, and closed-loop strategy generation. Summary of the Invention

[0005] The purpose of the present invention is to address the problems existing in the background technology and to propose a power inspection system and method.

[0006] The technical solution of the present invention is a power inspection method, which includes the following specific implementation steps:

[0007] S1. Collect real-time status data of power equipment and construct a third-order state transition tensor of the equipment in spatial position, state category, and time dimension based on the state transition tensor field and state time gradient. Calculate the confidence change rate of each state at different positions and times, and then identify short-term high-risk equipment.

[0008] S2. Build a long-term memory structure for the device's historical status. Combined with the exponential decay mechanism and state confidence reinforcement, dynamically update the historical memory value of each state, calculate the forgetting risk factor based on the historical peak value and the current quiet level, and generate a comprehensive inspection priority score.

[0009] S3. Combining the spatial topology structure with scheduling resource constraints, we construct an inspection task graph model, adopt a genetic perturbation mechanism to optimize and evolve the inspection path, and implement path parallel scheduling through task clustering and multi-queue scheduling strategies.

[0010] S4. Based on the topological structure between devices and the current anomaly score, a local linkage transfer function and risk propagation diagram are constructed to simulate the potential linkage path of abnormal events. Combined with the policy template library, risk patterns are identified and response strategies are matched, and the inspection path is dynamically adjusted.

[0011] Preferably, the identification process for short-term high-risk devices is as follows:

[0012] Deploy multimodal sensors to collect power equipment status data in real time and perform signal filtering, denoising, time synchronization, and normalization processing;

[0013] The power inspection area is divided into discrete spatial units, the equipment status is discretized into a set of categories based on key parameters, and multimodal real-time status data is used to construct a time series to capture dynamic changes in status.

[0014] A third-order tensor is constructed based on a set of spatial units, state categories, and time frames. Its elements are calculated through frequency statistics to calculate the conditional probability that the device is in a certain state at a specific location and time, that is, the state confidence.

[0015] Calculate the time difference of the state confidence of adjacent time points to form the state change gradient and quantify the rate of change of the device state;

[0016] Based on the equipment status confidence and deterioration trend gradient, the abnormal potential score of each spatial unit at the current moment is calculated by weight. Only the part where the status increases and deteriorates is counted, and a weight coefficient is introduced to distinguish the degree of influence of different statuses on the priority, and the abnormal potential score is output;

[0017] A global dynamic threshold is set based on the mean and standard deviation of the abnormal potential scores of all spatial units in the current frame, and spatial units with scores exceeding the threshold are screened and added to the high-priority task pool.

[0018] Preferably, the forgetting risk factor construction process is as follows:

[0019] The historical average confidence of each device state is recorded through a two-dimensional matrix of state and time period, and a long-term memory map is constructed to track the state change trajectory;

[0020] Introducing an exponential decay model and a current state confidence mechanism to update the historical impression strength: when a state appears, it is updated based on the decay factor and the current confidence, otherwise only the historical value is decayed;

[0021] Quantify the forgetting risk factor based on the difference between the historical highest state confidence and the current window average:

[0022] ;

[0023] in, Represents the spatial position r i In state s j The risk factor for forgetting under represents the state sensitivity adjustment coefficient; Indicates that in all historical time periods, state s j On the device i The highest confidence level ever seen; Indicates device r i In state s j is the average state confidence within the historical time period t; L is the total number of time frames; T is the length of the short-term observation window.

[0024] Preferably, the comprehensive inspection priority score generation process is as follows:

[0025] ;

[0026] in, Indicates device r i Comprehensive inspection priority score; represents the weight adjustment coefficient; represents the abnormal potential score, i.e., position r i At time t k The potential inspection priority score is based on the state upward trend and the current state confidence; M represents the total number of discretized device states.

[0027] Preferably, the evolution process of optimizing the inspection path using a quasi-genetic perturbation mechanism is as follows:

[0028] Constructing attributed graph structure G: ;

[0029] Among them, G represents the inspection task scheduling graph, that is, the topology and task relationship structure between devices; V represents the node set of the graph, , any node r i represents an electric inspection device; E represents the edge set of the graph, that is, there is a passable inspection path between two devices; W represents the node weight, which means that the device r i The comprehensive inspection priority score P(r i); A represents edge weight;

[0030] Define the optimization objective function : ;

[0031] in, Indicates a specific inspection path; Represents the path optimization objective function value, that is, the path Comprehensive benefit score; Indicates inspection access device r i The required dispatch cost; Represents node r i The historical scheduling frequency penalty value; 、 and represents the weight coefficient;

[0032] A genetic perturbation mechanism is introduced to perform online path evolution and self-correction to address path failure caused by dynamic changes in device status over time. This includes three types of evolution operations:

[0033] Local perturbation, i.e. local replacement: for the node r with low score in the current path k , in its neighborhood N(r k ) to search for a higher-optimal node r m , try to replace to form a new path ,like , then accept the replacement;

[0034] High priority insertion, that is, task insertion: when a node priority P(r j ) suddenly increases, traverse the current path to find the insertion position l, minimize the insertion cost, and insert it into the existing path;

[0035] Pruning and compression, i.e. task removal: If a node's current priority decreases and its cost is high, and it affects the overall efficiency of the task, it will be removed from the path and temporarily stored for backup.

[0036] Preferably, the implementation process of achieving path parallel scheduling through task cluster division and multi-queue scheduling strategy is as follows:

[0037] A1. Construct cluster sample vector: ;

[0038] in, Indicates device r i The three-dimensional space coordinates of represents the scaling factor of the status score; Indicates device r i Current comprehensive inspection priority score;

[0039] And the density clustering method DBSCAN is used to divide all devices into K task clusters;

[0040] A2. Define resource modeling conditions: ;

[0041] in, represents the resource capacity vector of the jth inspection queue; represents the kth task cluster; Represents a task cluster The minimum scheduling resource requirement required, i.e. the sum of the scheduling costs of the nodes in the cluster;

[0042] A3. After completing cluster division and resource modeling, the task clusters are matched and assigned to scheduling queues with corresponding resource capabilities. Specifically:

[0043] Initial priority matching: for each queue , calculate its relationship with the task cluster Task score:

[0044] ;

[0045] in, Indicates that the jth queue executes task cluster C k Scheduling score; Represents node r i The cost of access; Represents the scheduling cost adjustment factor, which controls the impact of the cost on the task cluster score;

[0046] Conflict coordination: If multiple queue targets overlap task clusters, conflict coordination is performed based on the queue's current load and remaining available resources.

[0047] Dynamic headroom replenishment: If a task in a cluster has high priority but no queue matches it, the redundancy replenishment mechanism is activated. Some high-priority nodes are extracted from the task cluster to form temporary tasks and distributed to queues with spare capacity for replenishment.

[0048] A4. Set the boundary node readjustment threshold , if a node is covered by multiple clusters, its If the task is a high-value task, it will be assigned to the queue with the shortest feasible path in the inspection queue; another queue will cache the node status and set it as a skip node.

[0049] A5. After the task clusters are matched with the queues, each queue will independently execute its path planning and task scheduling.

[0050] Preferably, the local linkage transfer function is:

[0051] ;

[0052] in, Represents node r i Abnormal pair node r j The impact strength of abnormal probability; Represents node r i Current anomaly score, also known as anomaly potential score ; Indicates the topological structure association; 、 represents the weight system; Represents the linkage probability function, outputting the probability in the interval [0,1];

[0053] The local anomaly influence matrix T is formed to capture the initial propagation conditions of the implicit anomaly diffusion path.

[0054] Preferably, the risk propagation map construction process is as follows:

[0055] Define the risk propagation graph G risk : ;

[0056] Where V represents the node set of the graph; E represents the edge set of the graph; R represents the risk score set of the node;

[0057] Construct deduction propagation rules, use deduction round t to control the propagation depth, and set the risk diffusion damping coefficient To avoid the infinite spread of abnormalities:

[0058] ;

[0059] in, Represents the node r after the tth propagation round i risk score; Represents node r i Initial risk situation; N(r j ) represents node r j adjacent nodes of ; Indicates the node r at the next moment obtained by deduction i risk score;

[0060] The termination condition is: reaching stability or the upper limit of deduction rounds, and obtaining the final risk map .

[0061] Preferably, the inspection path dynamic adjustment process is as follows:

[0062] Build a policy template library Model: ;

[0063] Among them, Model represents the policy template library, which contains the anomaly patterns P that have been identified in the past.k and response strategy A k ;P k A represents the kth known abnormal pattern in the template; k Indicates that P k The set of recommended response strategies matched; K is the total number of abnormal patterns identified in the past;

[0064] Calculate matching similarity:

[0065] ;

[0066] If sim> , then output strategy A k , otherwise it is marked as a new mode to be constructed as a temporary response;

[0067] in, Represents the set of abnormal features shown by the currently observed abnormal propagation graph; Represents the similarity score between the current anomaly propagation graph and the historical template; Indicates the threshold parameters of each template; Indicates the threshold value of each template set;

[0068] Define the scheduling profit change function: ;

[0069] in, Indicates the change in the overall scheduling benefit of the inspection system after applying the strategy; Indicates the updated and optimized inspection path after applying the strategy; Indicates the currently generated inspection path sequence;

[0070] like , then accept the strategy, update the path and scheduling queue;

[0071] The top n risk nodes of the new pattern are extracted to construct an emergency path cluster, and an insertion sequence is generated based on the greedy algorithm and cost constraints, and resources are allocated preferentially.

[0072] The technical solution of the present invention is a power inspection system, which is used to perform the above-mentioned power inspection method, including:

[0073] State perception and data fusion module, used to collect multimodal data of power equipment and construct state tensor field and anomaly scoring matrix;

[0074] Long-term memory construction and forgetting modeling module, used to record historical status changes of the device and calculate the forgetting risk factor;

[0075] Multi-constrained path scheduling and evolution module, used to build task scheduling graphs and generate optimal inspection paths through perturbation mechanisms;

[0076] The anomaly deduction and strategy adaptation module is used to simulate the propagation chain of abnormal events and match response strategies to achieve dynamic optimization of scheduling plans.

[0077] Compared with the prior art, the above technical solution of the present invention has the following beneficial technical effects:

[0078] This paper constructs a power inspection system and method that integrates behavior deviation, state evolution, and historical forgetting mechanisms to form a dynamically updated scoring system, effectively improving the ability to identify potential risks:

[0079] By constructing a third-order state transition tensor field to accurately quantify the spatiotemporal evolution characteristics of equipment status, combined with dynamic gradient analysis to capture short-term abnormal deterioration trends in real time, the accuracy and timeliness of high-risk equipment identification are significantly improved;

[0080] The introduction of a dynamically updated long-term memory structure combines historical state peaks with the current level of inactivity to calculate the forgetting risk factor. This proactively locates devices that were once severely abnormal but have recently become inactive, effectively resolving the issue of missing long-term hidden danger traceability in traditional inspections.

[0081] A dynamic task scheduling graph is constructed based on resource constraints and topological relationships. A genetic perturbation mechanism is used to optimize inspection paths. Combined with task clustering and multi-queue parallel scheduling, this significantly reduces path execution costs and improves resource utilization.

[0082] An abnormal linkage propagation model is established to predict cascading fault chains, and the optimal response plan is generated in real time by matching the strategy template library, forming a "monitoring-dispatch-disposal" closed-loop control, comprehensively enhancing the risk prevention and control capabilities and operation and maintenance efficiency of the power system. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 This is a flow chart of a power inspection method proposed by the present invention. DETAILED DESCRIPTION

[0084] Example 1, as Figure 1 As shown, the present invention proposes a power inspection method, which includes the following specific implementation steps:

[0085] S1. Based on high-dimensional multimodal data, the time-series evolution characteristics of the equipment spatial state are characterized by state transition tensor fields and dynamic gradient analysis. Combined with weighted anomaly scoring and dynamic threshold strategies, the short-term abnormal risks of power equipment are accurately identified. The specific implementation process is as follows:

[0086] S11. In the power inspection area, deploy multimodal sensors to collect real-time status data of equipment, including but not limited to: temperature sensors, infrared thermal imaging, partial discharge sensors, vibration sensors, and acoustic sensors;

[0087] Preprocessing of the collected real-time status data includes signal filtering, denoising, time synchronization and normalization to ensure the accurate temporal and spatial correspondence of different modal data;

[0088] S12. Divide the inspection area where the power equipment is located into multiple spatial units, each unit representing a certain physical area, to facilitate spatial mapping of sensor data. Discretize the equipment status according to key parameters (such as temperature level, voltage anomaly, vibration amplitude, etc.) to form a set of status categories. Multimodal data forms a time series at discrete time points to capture the change of equipment status over time. Specifically:

[0089] Define the power equipment inspection area as a set of discrete spatial units , and define the device state space ;

[0090] The collected multimodal data forms a discrete time series in the time dimension, and the time point set is ;

[0091] Among them, r i represents the i-th spatial unit, i.e., a certain power equipment or the physical location area where it is located, i=1,2,…,N, where N is the total number of spatial units; s j represents the jth type of discrete device state, representing a certain state interval of a certain physical attribute (such as "overheating" or "high discharge"), j = 1, 2, ..., M, where M is the total number of discrete device states set; t k represents the kth time frame, i.e., the time point in a unified sampling period, k=1,2,…,L, where L is the total number of time frames; R represents the complete set of spatial units, i.e., the set of all monitored inspection areas; S represents the complete set of state categories, i.e., the set of all discriminable state categories; T represents the set of time frames, i.e., the range of time series considered in this modeling;

[0092] S13. Based on the full set of spatial units R, the full set of state categories S, and the set of time frames T, define the third-order state transition tensor , whose elements ;

[0093] in, Represents the state confidence, that is, at time point t k , spatial position r i At, the power equipment is in state s j The conditional probability of Represents the state s counted in the original data j Appears at position r i , time t k frequency;

[0094] S14. Calculate the time difference of tensor elements between consecutive time points to form a state change gradient and quantify the speed of change of the device state over time. Specifically:

[0095] ;

[0096] in, Represents the state time gradient, that is, the device state s j At spatial position r i The state confidence of k to t k+1 rate of change; represents the interval between consecutive time frames;

[0097] It should be noted that when When , it indicates that the status is deteriorating, such as the thermal image temperature continues to rise; when When , it indicates that the state tends to recover or be repaired;

[0098] S15. Calculate the abnormal potential of each spatial unit in the current time frame based on the combined information of the tensor field and the gradient field, specifically: ;

[0099] in, represents the abnormal potential score, i.e., position r i At time t k The potential inspection priority score is based on the upward trend of the status and the confidence level of the current status; It means that only the rising (deteriorating) part of the status is considered, and the trend of status improvement is ignored; Represents the state weight coefficient, that is, state s j Impact on inspection priority (e.g., high-voltage breakdown > abnormal vibration > slight temperature rise);

[0100] For example, when the infrared temperature of a device increases, the PD signal strength increases, and the corresponding X i,j,k and G i,j,k At the same time, the position If a device is in a stable state, the score will not increase significantly.

[0101] S16. Based on the abnormal potential score, dynamically adjusted thresholds are set to screen out high-risk spatial units as short-term priority inspection targets. Specifically:

[0102] Defining global dynamic thresholds : ;

[0103] Filter to meet: The spatial unit r i, join the high priority task pool;

[0104] in, represents the mean potential score of all spatial units in the kth frame, which is used to construct the basis of dynamic threshold; Represents the standard deviation of the score of the kth frame, which is used to describe the degree of dispersion of the score distribution; represents the abnormal score sensitivity coefficient;

[0105] The high-priority task pool is a collection of spatial units identified as "high risk" or "high priority".

[0106] S2. Build a dynamically updated long-term memory structure for power events to store, retrieve, and enhance the evolution of equipment inspection status over time. Combined with the current status potential score, this structure identifies forgotten but potentially recurring risk factors, ensuring inspection continuity and the ability to trace hidden dangers. The specific implementation process is as follows:

[0107] S21. Continuously update the state occurrence density through the state × time period two-dimensional matrix to form a personalized memory map of space-state, specifically:

[0108] Define the memory matrix: ;

[0109] in, Indicates device r i In state s j The average state confidence within the historical time period t, t=1,2,…,H; H represents the time window length of the long-term memory structure; represents the state-time memory matrix, device r i The long-term state memory structure records the change trajectories of M states in H historical time periods;

[0110] S22. Introduce an exponential decay model and a current state confidence reinforcement mechanism to dynamically update the historical impression strength of each state. Specifically:

[0111] Define the long-term memory update rule as: ;

[0112] in, Represents the memory decay factor, that is, the proportion of historical state memory retained at the current moment (the smaller the better, the faster the forgetting, the larger the historical impact). ; Represents the state confidence, that is, at time point t and spatial position r i At, the power equipment is in state s j The conditional probability of Indicates the historical memory value in the previous time period;

[0113] S23. Construct a forgetting risk factor function that integrates historical peak memory and current silence level, define the forgetting risk of each state in a certain spatial unit as:

[0114] ;

[0115] in, Represents the spatial position r i In state s j the forgetting risk factor under (the larger the value, the higher the degree of severity but recent dormancy); Represents the state sensitivity adjustment coefficient, which is used to adjust the weight of the impact of forgetting risk on the final score (configurable or learned); Indicates that in all historical time periods, state s j On the device i The highest confidence level ever seen; T represents the length of the short-term observation window;

[0116] S24. Define inspection priority factors and comprehensively evaluate the inspection scheduling priority of each space unit:

[0117] ;

[0118] in, Indicates device r i Comprehensive inspection priority score (the higher the value, the higher the priority for inspection); represents the weight adjustment coefficient, i.e., the weight ratio of the fusion short-term potential score and the long-term forgetting risk factor;

[0119] S25. Build an active reactivation mechanism that triggers supplementary inspections or intelligent diagnosis tasks when areas with high forgetting risk but low short-term potential are discovered. The activation conditions are as follows:

[0120] ;

[0121] in, Represents the preset forgetting risk threshold, when R(r i ,s j )>θ, the state is considered to have a high risk of recurrence; Represents the lower threshold of the current potential score. When S(r i ,s j )<ε, it is considered that there is no obvious abnormality at present.

[0122] S3, spatial priority scoring based on the output of step S2 , combined with the actual resource scheduling constraints (including but not limited to personnel distribution, inspection time window, path distance, and historical load), a dynamically evolving task scheduling graph is constructed, and the optimal scheduling sequence is generated through the inspection path optimization algorithm. The specific implementation process is as follows:

[0123] S31. Construct an attributed graph structure G, integrating spatial topology, inspection task urgency, and access cost factors into a unified graph structure to support task path construction and multi-objective scheduling optimization:

[0124] ;

[0125] Among them, G represents the inspection task scheduling graph, that is, the topology and task relationship structure between devices; V represents the node set of the graph, , any node r i represents an electric inspection device; E represents the edge set of the graph, that is, there is a passable inspection path between two devices; W represents the node weight, which means that the device r i The comprehensive inspection priority score P(r i ); A represents edge weight;

[0126] S32. Define the optimization objective function , comprehensively considering the coverage of high-priority tasks, path execution cost and historical load balancing to improve overall scheduling efficiency:

[0127] ;

[0128] in, Represents a specific inspection path, consisting of several r i Composition, indicating the inspection visit sequence; Represents the path optimization objective function value, that is, the path The higher the comprehensive benefit score, the better; Indicates inspection access device r i The required scheduling cost, including but not limited to the cost of manpower and energy resources; Represents node r i The historical scheduling frequency penalty value is used to suppress repeated patrol behavior; 、 and Represents the weight coefficient, which is used to balance the influence of different scheduling objectives in the overall optimization;

[0129] S33 introduces a quasi-genetic perturbation mechanism to perform online evolution and self-correction on the path, solving the problem of path failure caused by dynamic changes in device status over time. It includes three types of evolution operations:

[0130] Local perturbation (local replacement): perturb the node r with low score in the current pathk , in its neighborhood N(r k ) to search for a higher-optimal node r m , try to replace to form a new path ,like , then accept the replacement;

[0131] High priority insertion (task insertion): When a node priority P(r j ) suddenly increases (such as a high temperature alarm), the current path is traversed to find the insertion position l, and after minimizing the insertion cost, it is inserted into the existing path;

[0132] Pruning and compression (task removal): If a node's current priority decreases and its cost is high, and it affects the overall efficiency of the task, it can be removed from the path and temporarily stored for backup to reduce resource usage;

[0133] S34. A scheduling queue generation method based on task clusters automatically divides the task space by region and state concentration, and assigns it to different inspection teams for parallel execution. Specifically:

[0134] A1. Construct cluster sample vector: ;

[0135] in, Indicates device r i The three-dimensional space coordinates of represents the scaling factor of the state score, which is used to balance the effects of spatial distance and state urgency; Indicates device r i Current comprehensive inspection priority score;

[0136] And use the density clustering method (DBSCAN is used in this embodiment) to divide all devices into K task clusters;

[0137] A2. Define resource modeling conditions: ;

[0138] in, The resource capacity vector representing the jth inspection queue, including but not limited to load, path type, and available time; represents the kth task cluster; Represents a task cluster The minimum scheduling resource requirements required (i.e., the sum of the scheduling costs of nodes in the cluster);

[0139] A3. After completing cluster division and resource modeling, a reasonable mapping strategy needs to be established to match task clusters to scheduling queues with corresponding resource capabilities. Specifically:

[0140] Initial priority matching: for each queue , calculate its relationship with the task cluster Task score:

[0141] ;

[0142] in, Indicates that the jth queue executes task cluster C k Scheduling score; Represents node r i The cost of access; Represents the scheduling cost adjustment factor, which controls the impact of the cost on the task cluster score;

[0143] Conflict coordination: If multiple queues have overlapping task clusters (e.g., two queues can schedule a certain cluster), conflict coordination is performed based on the queue's current load and remaining available resources.

[0144] Dynamic headroom replenishment: If a task in a cluster has high priority but no queue matches it, the "redundancy replenishment mechanism" is activated. Some high-priority nodes are extracted from the task cluster to form temporary tasks and distributed to queues with spare capacity for replenishment.

[0145] A4. In actual scheduling, there is often overlap of boundary nodes between task clusters (for example, substation equipment is located at the intersection of multiple task clusters). Coordinated scheduling is required to prevent resource duplication and path conflicts. To this end: set a boundary node readjustment threshold , if a node is covered by multiple clusters, its And it is a high-value task, then:

[0146] Actively assigned to the queue with the shortest feasible path among the inspection queues;

[0147] Another queue caches the node status and sets it as a skip node;

[0148] A5. After the task clusters are matched with the queues, each queue will independently perform its path planning and task scheduling;

[0149] S35. Use the feedback results of each inspection to update the scheduling diagram, equipment scores and frequency penalty items in real time;

[0150] For example, if a device r i If the status of the device is confirmed to have deteriorated during inspection, the scores of the adjacent devices will be increased simultaneously; their own scores will be strengthened in memory in future cycles (entering the key status cache); if the device has no abnormalities for N consecutive rounds, its historical frequency penalty item D(r i ), release resources; new equipment or fault points (such as newly deployed smart meters) are synchronously added to the scheduling diagram structure G;

[0151] For example, if node r 25 In the long term, high temperature alarms have occurred many times, but they have been missed in recent rounds of inspections. The score is P(r25 )=0.94, cost C=0.12, frequency penalty D=0.08, then its scheduling score is higher in path optimization; if the new device r 37 It is found that there is an overvoltage problem through feedback, and the feedback result F(r 37 ) = abnormal, its score ΔP(r 37 )=+0.35, it will be inserted into the current scheduling graph and increase the priority; if the node r 14 After 5 consecutive rounds of normal operation, the system will change its D(r 14 ) is halved, reducing resource waste.

[0152] S4. Build a mechanism that features linkage analysis, pattern deduction, and automatic matching of response strategies. Focusing on abnormal events, establish a behavioral pattern evolution chain across devices, time, and multimodal data. Combined with the actual strategy database, generate the optimal response plan in real time and provide feedback to the dispatch system. The specific implementation process is as follows:

[0153] S41. Model the impact relationship between abnormal devices to identify potential linkage risk paths that may be activated due to a node anomaly. Specifically:

[0154] Introduce the local linkage transfer function:

[0155] ;

[0156] in, Represents node r i Abnormal pair node r j The impact strength of abnormal probability; Represents node r i Current anomaly score, also known as anomaly potential score ; Represents the topological structure association, that is, the connection weight between two nodes in an electrical diagram or wiring diagram; 、 represents the weight coefficient, which weighs the coupling effect between the anomaly intensity and the topological correlation; Represents the linkage probability function, outputting the probability in the interval [0,1];

[0157] Based on this, the local anomaly impact matrix T is output to capture the initial propagation conditions of the implicit anomaly diffusion path;

[0158] S42. Construct risk propagation map G risk ,The local linkage relationship constructed in the previous step is embedded in the global scheduling graph structure to form a dynamic risk graph with risk propagation capability, simulate the cascading failure chain that may be caused by equipment anomalies, and predict the potential diffusion direction of high-risk areas. Specifically:

[0159] ;

[0160] Construct deduction propagation rules, use deduction round t to control the propagation depth, and set the risk diffusion damping coefficient To avoid the infinite spread of abnormalities:

[0161] ;

[0162] Among them, R represents the risk score set of the node, R(r i ) represents node r i risk score; Represents the node r after the tth propagation round i risk score; Represents node r i Initial risk situation; N(r j ) represents node r j adjacent nodes of ; Indicates the node r at the next moment obtained by deduction i risk score;

[0163] The termination condition is: reaching stability or the upper limit of deduction rounds, and obtaining the final risk map , used for the next step of strategy matching;

[0164] S43. Build a policy template library Model to identify the abnormal pattern characteristics presented in the propagation diagram, match them with the response strategy library preset in the system, and quickly select the empirically effective response plan, specifically:

[0165] ;

[0166] Among them, Model represents the policy template library, which contains the anomaly patterns P that have been identified in the past. k and response strategy A k ;P k A represents the kth known abnormal pattern in the template; k Indicates that P k The set of recommended response strategies matched; K is the total number of abnormal patterns identified in the past;

[0167] Calculate matching similarity: ;

[0168] If sim> (Set the threshold parameters of each template), then the output strategy A k , otherwise it is marked as a new mode to be constructed as a temporary response;

[0169] in, Represents the abnormal feature set shown by the currently observed abnormal propagation graph, risk graph R ∗ Extracted from Represents the similarity score between the current anomaly propagation graph and the historical template; Indicates the threshold parameter of each template, which is used to control the tightness of the match and is adjusted based on historical template statistics;

[0170] S44. Define a scheduling benefit change function. Based on the identified abnormal linkage structure and recommended strategy, automatically adjust the original inspection plan, insert high-priority paths, or reorder tasks to ensure that risk control is prioritized. Specifically:

[0171] Define the scheduling profit change function: ;

[0172] in, Indicates the change in the overall scheduling benefit of the inspection system after applying the strategy; Indicates the updated and optimized inspection path after applying the strategy; Indicates the currently generated inspection path sequence;

[0173] like , then accept the strategy, update the path and scheduling queue;

[0174] If a new pattern is identified, an emergency path cluster C is constructed based on the current high-risk nodes. risk , perform rapid response task scheduling;

[0175] The path construction method is: based on Sorting and extracting the top n risk nodes; using greedy + cost constraints to generate an insertion sequence; dynamically adjusting the inspection resource queue to prioritize allocation to policy nodes;

[0176] S45. Integrate the execution results with the actual feedback loop to enhance the long-term adaptability of the strategy template and propagation model, and realize the knowledge evolution and behavior update of the power inspection system;

[0177] For example, if strategy A k After execution, the anomaly rate decreases significantly: enhance the matching degree of the policy template; update its historical success rate; enhance the state transition coefficient between related nodes; if the policy execution is invalid: reduce the credibility of the policy; mark the event as a potential new pattern P new , forming new policy base entries in future executions.

[0178] In the second embodiment, the present invention proposes an electric power inspection system, which is used to execute the electric power inspection method proposed in the first embodiment, including: a state perception and data fusion module, a long-term memory construction and forgetting modeling module, a multi-constraint path scheduling and evolution module, and an abnormality deduction and strategy adaptation module.

[0179] The state perception and data fusion module is used to collect multi-source heterogeneous data on the operating status of power equipment, including but not limited to temperature, voltage, current, vibration, historical alarm records, and environmental parameters. It then constructs a unified heterogeneous state quantity matrix by fusing temporal stability scoring, behavioral deviation detection, and spatial distribution coupling features, and then outputs anomaly scoring results for inspection priority sorting.

[0180] The long-term memory construction and forgetting modeling module is used to capture hidden abnormal behavior patterns in the equipment operation history. It simulates the experiential forgetting process in manual operation and maintenance by constructing a decreasing memory function. It also dynamically quantifies the forgetting risk factor to reflect potential structural risks in the long term.

[0181] The multi-constrained path scheduling and evolution module is used to integrate scoring results, historical forgetting intensity, and geographical or electrical layout constraints to build a path evolution scheduling function based on multiple factors such as abnormality priority, inspection cost, and memory decay. The module outputs the optimal inspection path through dynamically updated task queues and resource allocation strategies.

[0182] The anomaly deduction and strategy adaptation module is used to identify possible abnormal causal relationships between devices, build a dynamic abnormal state propagation model for the diffusion of linkage risks, match the optimal response plan based on the historical strategy library, and make real-time adjustments and plug-in scheduling to the original path, thereby realizing the prediction, response and self-update of strategy knowledge of abnormal events.

[0183] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited thereto. Various changes can be made within the scope of knowledge possessed by those skilled in the art without departing from the spirit of the present invention.

Claims

1. A power inspection method, characterized in that: The specific implementation steps include the following: S1. Collect real-time status data of power equipment and construct a third-order state transition tensor of the equipment in spatial position, state category, and time dimension based on the state transition tensor field and state time gradient. Calculate the confidence change rate of each state at different positions and times, and then identify short-term high-risk equipment. The process for identifying short-term high-risk devices is as follows: Deploy multimodal sensors to collect power equipment status data in real time and perform signal filtering, denoising, time synchronization, and normalization processing; The power inspection area is divided into discrete spatial units, the equipment status is discretized into a set of categories based on key parameters, and multimodal real-time status data is used to construct a time series to capture dynamic changes in status. A third-order tensor is constructed based on a set of spatial units, state categories, and time frames. Its elements are calculated through frequency statistics to calculate the conditional probability that the device is in a certain state at a specific location and time, that is, the state confidence. Calculate the time difference of the state confidence of adjacent time points to form the state change gradient and quantify the rate of change of the device state; Based on the equipment status confidence and deterioration trend gradient, the abnormal potential score of each spatial unit at the current moment is calculated by weight. Only the part where the status increases and deteriorates is counted, and a weight coefficient is introduced to distinguish the degree of influence of different statuses on the priority, and the abnormal potential score is output; A global dynamic threshold is set based on the mean and standard deviation of the abnormal potential scores of all spatial units in the current frame, and spatial units with scores exceeding the threshold are screened and added to the high-priority task pool; S2. Build a long-term memory structure for the device's historical status. Combined with the exponential decay mechanism and state confidence reinforcement, dynamically update the historical memory value of each state, calculate the forgetting risk factor based on the historical peak value and the current quiet level, and generate a comprehensive inspection priority score. S3. Combining the spatial topology structure with scheduling resource constraints, we construct an inspection task graph model, adopt a genetic perturbation mechanism to optimize and evolve the inspection path, and implement path parallel scheduling through task clustering and multi-queue scheduling strategies. S4. Based on the topological structure between devices and the current anomaly score, a local linkage transfer function and risk propagation diagram are constructed to simulate the potential linkage path of abnormal events. Combined with the policy template library, risk patterns are identified and response strategies are matched, and the inspection path is dynamically adjusted.

2. A power inspection method according to claim 1, characterized in that: The process of constructing the forgetting risk factor is as follows: The historical average confidence of each device state is recorded through a two-dimensional matrix of state and time period, and a long-term memory map is constructed to track the state change trajectory; Introducing an exponential decay model and a current state confidence mechanism to update the historical impression strength: when a state appears, it is updated based on the decay factor and the current confidence, otherwise only the historical value is decayed; Quantify the forgetting risk factor based on the difference between the historical highest state confidence and the current window average: ； in, Represents the spatial position r i In state s j The risk factor for forgetting under represents the state sensitivity adjustment coefficient; Indicates that in all historical time periods, state s j On the device i The highest confidence level ever seen; Indicates device r i In state s j is the average state confidence within the historical time period t; L is the total number of time frames; T is the length of the short-term observation window.

3. The power inspection method according to claim 2, characterized in that: The comprehensive inspection priority score generation process is as follows: ； in, Indicates device r i Comprehensive inspection priority score; represents the weight adjustment coefficient; represents the abnormal potential score, i.e., position r i At time t k The potential inspection priority score is based on the state upward trend and the current state confidence; M represents the total number of discretized device states.

4. The power inspection method according to claim 3, characterized in that: The evolution process of optimizing the inspection path using a genetic perturbation mechanism is as follows: Constructing attributed graph structure G: ; Among them, G represents the inspection task scheduling graph, that is, the topology and task relationship structure between devices; V represents the node set of the graph, , any node r i represents an electric inspection device; E represents the edge set of the graph, that is, there is a passable inspection path between two devices; W represents the node weight, which means that the device r i The comprehensive inspection priority score P(r i ); A represents edge weight; Define the optimization objective function : ; in, Indicates a specific inspection path; Represents the path optimization objective function value, that is, the path Comprehensive benefit score; Indicates inspection access device r i The required dispatch cost; Represents node r i The historical scheduling frequency penalty value; 、 and represents the weight coefficient; A genetic perturbation mechanism is introduced to perform online path evolution and self-correction to address path failure caused by dynamic changes in device status over time. This includes three types of evolution operations: Local perturbation, i.e. local replacement: for the node r with low score in the current path k , in its neighborhood N(r k ) to search for a higher-optimal node r m , replace to form a new path; like , then accept the replacement; among them, Indicates the new path; High priority insertion, that is, task insertion: when a node priority P(r j ) suddenly increases, traverse the current path to find the insertion position l, minimize the insertion cost, and insert it into the existing path; Pruning and compression, i.e. task removal: If a node's current priority decreases and its cost is high, and it affects the overall efficiency of the task, it will be removed from the path and temporarily stored for backup.

5. The power inspection method according to claim 4, characterized in that: The implementation process of path parallel scheduling through task cluster division and multi-queue scheduling strategy is as follows: A1. Construct cluster sample vector: ; in, Indicates device r i The three-dimensional space coordinates of represents the scaling factor of the status score; Indicates device r i Current comprehensive inspection priority score; And the density clustering method DBSCAN is used to divide all devices into K task clusters; A2. Define resource modeling conditions: ; in, represents the resource capacity vector of the jth inspection queue; represents the kth task cluster; Represents a task cluster The minimum scheduling resource requirement required, i.e. the sum of the scheduling costs of the nodes in the cluster; A3. After completing cluster division and resource modeling, the task clusters are matched and assigned to scheduling queues with corresponding resource capabilities. Specifically: Initial priority matching: for each queue , calculate its relationship with the task cluster Task score: ； in, Indicates that the jth queue executes task cluster C k Scheduling score; Represents node r i The cost of access; Represents the scheduling cost adjustment factor, which controls the impact of the cost on the task cluster score; Conflict coordination: If multiple queue targets overlap task clusters, conflict coordination is performed based on the queue's current load and remaining available resources. Dynamic headroom replenishment: If a task in a cluster has high priority but no queue matches it, the redundancy replenishment mechanism is activated. Some high-priority nodes are extracted from the task cluster to form temporary tasks and distributed to queues with spare capacity for replenishment. A4. Set the boundary node readjustment threshold , if a node is covered by multiple clusters, its If the task is a high-value task, it will be assigned to the queue with the shortest feasible path in the inspection queue. Another queue will cache the node status and set it as a skip node. A5. After the task clusters are matched with the queues, each queue will independently execute its path planning and task scheduling.

6. The power inspection method according to claim 5, characterized in that: The local linkage transfer function is: ； in, Represents node r i Abnormal pair node r j The impact strength of abnormal probability; Represents node r i Current anomaly score, also known as anomaly potential score ; Indicates the topological structure association; 、 represents the weight system; Represents the linkage probability function, outputting the probability in the interval [0,1]; The local anomaly influence matrix T is formed to capture the initial propagation conditions of the implicit anomaly diffusion path.

7. The power inspection method according to claim 6, characterized in that: The process of constructing a risk propagation map is as follows: Define the risk propagation graph G risk : ; Where V represents the node set of the graph; E represents the edge set of the graph; R represents the risk score set of the node; Construct deduction propagation rules, use deduction round t to control the propagation depth, and set the risk diffusion damping coefficient To avoid the infinite spread of abnormalities: ； in, Represents the node r after the tth propagation round i risk score; Represents node r i Initial risk situation; N(r j ) represents node r j adjacent nodes of ; Represents the node r at the next moment obtained by deduction i risk score; The termination condition is: reaching stability or the upper limit of deduction rounds, and obtaining the final risk map .

8. The power inspection method according to claim 7, characterized in that: The inspection path dynamic adjustment process is as follows: Build a policy template library Model: ; Among them, Model represents the policy template library, which contains the anomaly patterns P that have been identified in the past. k and response strategy A k ;P k A represents the kth known abnormal pattern in the template; k Indicates that P k The set of recommended response strategies matched; K is the total number of abnormal patterns identified in the past; Calculate matching similarity: ; If sim> , then output strategy A k , otherwise it is marked as a new mode to be constructed as a temporary response; in, Represents the set of abnormal features shown by the currently observed abnormal propagation graph; Represents the similarity score between the current anomaly propagation graph and the historical template; Indicates the threshold parameters of each template; Indicates the threshold value of each template set; Define the scheduling profit change function: ; in, Indicates the change in the overall scheduling benefit of the inspection system after applying the strategy; Indicates the updated and optimized inspection path after applying the strategy; Indicates the currently generated inspection path sequence; like , then accept the strategy, update the path and scheduling queue; The top n risk nodes of the new pattern are extracted to construct an emergency path cluster, and an insertion sequence is generated based on the greedy algorithm and cost constraints, and resources are allocated preferentially.

9. A power inspection system, used to execute the power inspection method according to any one of claims 1 to 8, characterized in that: include: State perception and data fusion module, used to collect multimodal data of power equipment and construct state tensor field and anomaly scoring matrix; Long-term memory construction and forgetting modeling module, used to record historical status changes of the device and calculate the forgetting risk factor; Multi-constrained path scheduling and evolution module, used to build task scheduling graphs and generate optimal inspection paths through perturbation mechanisms; The anomaly deduction and strategy adaptation module is used to simulate the propagation chain of abnormal events and match response strategies to achieve dynamic optimization of scheduling plans.

Images