Centralized control resource intelligent inspection method and system based on multi-source data fusion

Abstract

The invention discloses a centralized control resource intelligent inspection method based on multi-source data fusion. The method comprises the following steps: collecting full-dimensional data in real time, endowing all collected data with a uniform label of unique ID and coordinates in a dynamic semantic map, and constructing a multi-source data collection platform; performing preprocessing and space-time alignment on multi-source data based on a multi-source data acquisition platform, extracting modal features of different types of data through a feature extraction network, generating a final fusion feature vector, and constructing an intelligent diagnosis model; and dividing equipment levels by using equipment operation state data output by the intelligent diagnosis model, determining a dynamic inspection period through an intelligent algorithm based on the equipment levels, planning an optimal inspection path, and performing autonomous inspection and data acquisition by a corresponding agent according to the optimal inspection path. According to the invention, an intelligent solution can be provided for daily operation and maintenance of the centralized control center, the operation and maintenance cost is reduced, the operation and maintenance efficiency and quality are improved, and the intelligent development of the power industry is promoted.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance of power systems, and in particular to an intelligent inspection method and system for centralized control resources based on multi-source data fusion. Background Technology

[0002] With the large-scale and intelligent development of hydropower, centralized control and dispatching systems are playing an increasingly important role in the operation and management of hydropower station clusters. However, current centralized control system resource management faces challenges such as resource dispersion, low management efficiency, and insufficient intelligence. The current state of intelligent resource management in centralized control systems is mainly manifested in the following aspects: First, manual inspection is inefficient; traditional inspections rely on manual labor, are prone to missed inspections and misjudgments, and cannot respond to faults in real time. Second, fault early warning is delayed; historically frequent fault points lack dynamic monitoring, making root cause analysis difficult. Third, resource utilization is uneven; server, storage, and other equipment resources are allocated irrationally, resulting in high operation and maintenance costs. Finally, multi-source data is fragmented; intelligent agent inspection, video surveillance, and sensor data are not effectively integrated, making collaborative analysis difficult.

[0003] This project's research into key technologies for intelligent inspection and maintenance of centralized control resources will help overcome existing technological bottlenecks, improve the intelligence level of centralized control systems, and achieve efficient resource management and optimized allocation. Furthermore, exploring typical application scenarios such as intelligent inspection and intelligent maintenance can provide intelligent solutions for the daily operation and maintenance of centralized control centers, reducing maintenance costs and improving efficiency and quality, which is of great significance for promoting the intelligent development of the power industry. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for intelligent inspection of centralized control resources based on multi-source data fusion. This method and system can improve the intelligence level of the centralized control system and realize efficient management and optimized allocation of resources.

[0005] The objective of this invention can be achieved through the following technical solutions: A method for intelligent inspection of centralized control resources based on multi-source data fusion includes the following steps: Historical, multi-dimensional monitoring data of substation equipment and environment are assigned unique IDs and coordinates in a dynamic semantic map; Historical full-dimensional monitoring data with unique IDs and coordinate unified labels is preprocessed to obtain a fused data package of historical full-dimensional monitoring data. The modal features of the fused data package are extracted by a feature extraction network. The modal features of the fused data package are input into a graph attention network to generate a final fused feature vector. The operating status data of the substation equipment is output based on the final fused feature vector. The substation equipment operation status data is combined with the equipment criticality to divide the equipment into levels. Based on the divided equipment levels, the dynamic inspection cycle is determined by a dynamic inspection cycle algorithm that integrates equipment health and predicted fault probability, and the optimal inspection path is planned. The corresponding intelligent agent then performs autonomous inspections according to the optimal inspection path.

[0006] A centralized control resource intelligent inspection system based on multi-source data fusion includes the following modules: data acquisition, data analysis, and inspection path calculation. The data acquisition module is used to assign a unique ID and coordinate unified label to the historical full-dimensional monitoring data of substation equipment and environment in a dynamic semantic map; The data analysis module is used to preprocess historical full-dimensional monitoring data with unique IDs and coordinate unified labels to obtain a fused data package of historical full-dimensional monitoring data. The modal features of the fused data package are extracted by a feature extraction network. The modal features of the fused data package are input into a graph attention network to generate a final fused feature vector. The operating status data of the substation equipment is output based on the final fused feature vector. The inspection path calculation module is used to divide the substation equipment into equipment levels by combining the operating status data of the equipment with the equipment criticality. Based on the divided equipment levels, the dynamic inspection cycle is determined by a dynamic inspection cycle algorithm that integrates equipment health and predicted fault probability, and the optimal inspection path is planned. The corresponding intelligent agent then performs autonomous inspections according to the optimal inspection path.

[0007] Compared with the prior art, the present invention has the following beneficial effects: This invention helps to overcome existing technological bottlenecks, improve the intelligence level of centralized control systems, and achieve efficient management and optimized allocation of resources. Furthermore, exploring typical application scenarios such as intelligent inspection and intelligent operation and maintenance can provide intelligent solutions for the daily operation and maintenance of centralized control centers, reducing operation and maintenance costs and improving efficiency and quality, which is of great significance for promoting the intelligent development of the power industry. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the structure of the present invention; Detailed Implementation

[0009] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0010] Example 1 The first embodiment of the present invention provides a method for intelligent inspection of centralized control resources based on multi-source data fusion, which includes the following steps: Historical, multi-dimensional monitoring data of substation equipment and environment are assigned a unique ID (identity number) and coordinates in a dynamic semantic map; Historical full-dimensional monitoring data with unique IDs and coordinate unified labels is preprocessed to obtain a fused data package of historical full-dimensional monitoring data. The modal features of the fused data package are extracted by a feature extraction network. The modal features of the fused data package are input into a graph attention network to generate a final fused feature vector. The operating status data of the substation equipment is output based on the final fused feature vector. The centralized control system resource intelligent management platform divides substation equipment into levels by combining the operating status data of the equipment with the criticality of the equipment. Based on the divided equipment levels, the dynamic inspection cycle is determined by a dynamic inspection cycle algorithm that integrates equipment health and predicted fault probability, and the optimal inspection path is planned. The corresponding intelligent agent then performs autonomous inspections according to the optimal inspection path.

[0011] A method for intelligent inspection of centralized control resources based on multi-source data fusion, wherein the historical full-dimensional monitoring data of substation equipment and environment includes: Visible light and infrared images historically collected by the inspection intelligent agent; video data historically collected by intelligent video surveillance; and time-series data historically collected by fixed sensors.

[0012] The inspection agent traverses the inspection area and uses data from LiDAR and IMU (Inertial Measurement Unit) to construct a geometric map of the central control center's computer room environment through an improved SLAM (Simultaneous Localization and Mapping) algorithm. The intelligent agent identifies QR code / Near Field Communication (NFC) tags on the device through a visual sensor, or the inspector scans the tag with a mobile terminal to bind the tag's metadata, such as device ID, model, parameters, and maintenance records, to its physical location on the geometric map; The data interfaces and location information of fixed sensors (such as temperature, humidity, and vibration sensors) are also injected into the geometric map. Finally, a dynamic semantic map is constructed, whose data structure can be represented as: Map_Node = {Pose, Geometry, Device_ID, Sensor_Data_Stream_Pointer, Maintenance_History_Pointer}.

[0013] Here, Pose is the pose, representing the node's position and orientation in space; Geometry is the geometry, representing the node's physical outline and collision volume; Device_ID is the device identifier, a string that uniquely identifies the sensor or hardware device; Sensor_Data_Stream_Pointer is the sensor data stream pointer, pointing to a reference or memory address of the real-time sensor data stream; and Maintenance_History_Pointer is the maintenance history pointer, pointing to the node's maintenance records, including update, correction, and fault handling history. The core function of assigning unique IDs and coordinate labels to historical, multi-dimensional monitoring data of substation equipment and environment within a dynamic semantic map is to establish a unified spatial and identity association benchmark for multi-source heterogeneous data. This aims to break down data silos and support intelligent inspection throughout the entire process. On one hand, the unique ID accurately binds data from different sources and types, such as visible light / infrared images collected by the inspection agent, video data from intelligent video surveillance, and time-series data from fixed sensors, to specific equipment. Simultaneously, the coordinate labels clearly define the equipment's location in physical space, providing a foundation for spatiotemporal alignment in subsequent data preprocessing, ensuring that (device ID, T_event) is used as the basis for the process. On the one hand, the tag enables precise and efficient operations such as window data filtering and timestamp alignment, ensuring the integrity and relevance of the fused data packets. On the other hand, the tag enables deep integration of data with dynamic semantic maps, allowing information such as historical monitoring data, maintenance records, and real-time status of equipment to be indexed and retrieved on the map through IDs and coordinates. This provides spatial and identity basis for graph attention networks to extract equipment topology relationships and generate context-enhanced feature vectors, and also provides accurate equipment positioning and data association support for subsequent equipment hierarchy division, dynamic inspection cycle calculation, and optimal path planning. Ultimately, it achieves intelligent collaboration throughout the entire process from data collection and analysis to inspection execution.

[0014] The specific method for preprocessing historical full-dimensional monitoring data into a fused data package of historical full-dimensional monitoring data, which is assigned a unique ID and a unified coordinate label, according to a method for intelligent inspection of centralized control resources based on multi-source data fusion, is as follows: The centralized control system resource intelligent management platform defines historical diagnostic events based on historical abnormal alarms of substation equipment. The event center is the equipment ID, and the event time is T_event. Define a window centered on the device ID and event time T_event(device ID, T_event), locking the diagnostic event device ID and its direct neighbors in the topology graph. Define the interval [T_event - Δt, T_event + Δt], where Δt is the window radius set according to the device's dynamic characteristics; The historical full-dimensional monitoring data within the set window is standardized and timestamp aligned using the device ID as the key. The aligned data is packaged into a fused data package of historical full-dimensional monitoring data; A method for intelligent inspection of centralized control resources based on multi-source data fusion, wherein the specific method for extracting the modal features of the fused data packets through a feature extraction network is as follows: The image data from the fused data package is input into a convolutional neural network (CNN) to process the visible light image and obtain the first visual feature vector V_rgb, and the infrared image is processed to obtain the second visual feature vector V_ir. The video data from the fused data package is input into a 3D-Convolutional Neural Network (3D-CNN) to process the video segments and obtain the dynamic feature vector V_video. The temporal data from the fused data package is input into a Long Short-Term Memory (LSTM) network to process the multivariate temporal data and obtain the temporal feature vector V_ts.

[0015] The specific method for inputting the image data from the fused data packet into the CNN, processing the visible light image to obtain the first visual feature vector V_rgb, and processing the infrared image to obtain the second visual feature vector V_ir is as follows: The image data is preprocessed by scaling all images to a fixed size and scaling pixel values ​​from the range of [0, 255] to the range of [0, 1] or [-1, 1].

[0016] CNNs simulate the mechanism of the human visual cortex, extracting features layer by layer through abstraction. Convolutional layers use multiple learnable convolutional kernels to slide and scan across the input image, obtaining a series of feature maps. Following the convolutional layers is a Rectified Linear Unit (ReLU) activation function, introducing non-linearity to the network and enabling it to learn more complex function mappings. Downsampling of the feature maps, such as max pooling which only takes the maximum value within a 2x2 window, is then performed. By stacking the "convolution, activation, pooling" structure multiple times, the network can extract visual features from the original pixels layer by layer, from low-level to high-level.

[0017] After multiple convolutions and pooling operations, a series of feature maps containing high-level semantic information are obtained. Finally, the visible light image is converted into the first visual feature vector V_rgb, and the infrared image is converted into the second visual feature vector V_ir.

[0018] The specific method for inputting the video data from the fused data package into a 3D-CNN to process the video segment and obtain the dynamic feature vector V_video is as follows: The video data is preprocessed by dividing the long video into fixed-length video segments, and normalizing the size and values ​​of each frame within the segment in the same way as the image.

[0019] By leveraging the natural temporal extension of 3D-CNN, spatial and temporal features are captured. A series of spatiotemporal feature maps are obtained by sliding 3D convolutional kernels across video segments. Pooling layers perform downsampling simultaneously in both spatial and temporal dimensions.

[0020] After multiple 3D convolutions and 3D pooling, the dynamic feature vector V_video is obtained.

[0021] The specific method for inputting the temporal data from the fused data packet into a Long Short-Term Memory (LSTM) network to process multivariate temporal data and obtain the temporal feature vector V_ts is as follows: The time series data in the fused data package is preprocessed by performing Z-score normalization or Min-Max scaling on each time series data to eliminate the influence of dimensions, and the continuous time series data is divided into input sequences of fixed length.

[0022] The input sequence is processed sequentially by time step using an LSTM unit. At each time step t, it receives the input x_t of the current time step and the hidden state h_{t-1} of the previous time step, and then updates its own cell state C_t and outputs the hidden state h_t through an internal gating mechanism.

[0023] After the LSTM processes the entire input sequence, the hidden state h_T output by the LSTM at the last time step is directly taken, which is the temporal feature vector V_ts.

[0024] A method for intelligent inspection of centralized control resources based on multi-source data fusion, which inputs the modal features of the fused data packets into a graph attention network to generate a final fused feature vector, and outputs the operating status data of substation equipment based on the final fused feature vector, is as follows: Define the target device node in the historical diagnostic events as v_target, and the initial node feature of the target device node v_target as X_initial, which is composed of the device's own multimodal features; X_initial = Concat(V_rgb, V_ir, ​​V_ts, V_video), where Concat() is a function used to concatenate two or more arrays or strings; Extract the device topology graph G=(V, E) from the dynamic semantic map; the node set V represents all the devices in the substation, and the edge set E represents the physical connection, pipeline flow direction or logical control relationship between the devices; The device topology graph G and the initial node features of the target device node v_target are input into a graph attention network. After propagation through multiple layers of GAT, the context-enhanced feature vector V_cnn of the target device node is obtained. The specific method is as follows: Construct a graph structure G=(V, E), where the node set V represents all equipment in the substation, and each node i corresponds to one device. The edge set E represents the physical connections, pipeline flow directions, or logical control relationships between devices. An edge (i, j) indicates that nodes i and j are connected.

[0025] Construct a feature matrix H of dimension N × F, where N is the total number of nodes, F is the initial feature dimension, and each row h_i of matrix H is the initial feature vector of the corresponding node i.

[0026] Through single-layer GAT propagation, for each node j in the target node i and its neighbor set N(i), GAT calculates the attention coefficient e_ij, and uses the softmax function to normalize the attention coefficients of all neighbors to obtain the final attention weight αij.

[0027] αij = softmax_j(e_ij) The new feature h'_i of node i is the weighted sum of the features of all its neighboring nodes.

[0028] h'_i = σ( Σ_{j ∈ N(i)} αij · W h_j ) Where W is the shared linear transformation matrix and σ is the nonlinear activation function (ReLU).

[0029] Finally, after multiple layers of GAT stacking and propagation through L layers of GAT, the features of each node are integrated with the global context information within its L-hop range, resulting in the final node feature matrix H^(L). The row vector corresponding to the target device node t (the data in the t row of the final matrix) is its context-enhanced feature vector V_cnn.

[0030] The initial node features X_initial of the target device node v_target and the context-enhanced feature vector V_cnn are input together into the multi-head cross-modal attention fusion layer; The multi-head cross-modal attention fusion layer performs a weighted summation of different feature vectors to obtain the final fused feature vector V_fused. The specific method is as follows: The multiple input feature vectors {V_rgb, V_ir, ​​V_ts, V_video, V_cnn} are considered as a set X = {x_1, x_2, x_3, x_4, x_5}, where each x_1 is the first visual feature vector V_rgb, x_2 is the second visual feature vector V_ir, ​​x_3 is the temporal feature vector V_ts, x_4 is the dynamic feature vector V_video, and x_5 is the context-enhanced feature vector V_cnn.

[0031] For each vector x_i in the input set, three new vectors are generated through three different, learnable linear transformation matrices (W_Q, W_K, W_V): The query vector q_i = x_i · W_Q represents the information that the current modality (e.g., V_rgb) obtains from all other modalities.

[0032] The key vector k_i = x_i · W_K represents the "label" or "index" of the information contained in the current modality (such as V_ir).

[0033] Value vector: v_i = x_i · W_V, representing the "information content" that is actually contained in the current modality (such as V_ts) and is to be extracted.

[0034] The attention weights are calculated by taking the dot product s_ij = q_i · k_j^T of the query vector q_i of the modality and the key vectors k_j of all modalities, where ^T is the transpose sign, which quantifies the correlation between them.

[0035] Weighted summation is performed on the value vectors v_j of all modalities according to the attention weights αij, and z_i = Σ(αij · v_j) is obtained to obtain the new feature representation after fusion.

[0036] Multi-head review, from different perspectives, involves running multiple independent attention computation units (heads) in parallel. Each head has its own set of W_Q, W_K, W_V to capture the characteristic interaction patterns of different subspaces.

[0037] The output vectors of all attention heads are concatenated, then integrated through the final linear transformation layer, and finally aggregated again to obtain the final and unique fused feature vector V_fused.

[0038] The final fused feature vector V_fused is input into a fully connected classifier, which outputs the operating status data of the substation equipment (such as normal, overheating, abnormal vibration, loose connection, etc.).

[0039] The steps involve preprocessing historical full-dimensional monitoring data with unique IDs and coordinate labels to obtain a fused data package. A feature extraction network extracts the modal features of this fused data package, which are then input into a graph attention network to generate a final fused feature vector. The significance of outputting the substation equipment operating status data based on this final fused feature vector lies in overcoming the technical bottleneck of fragmented multi-source data. It transforms massive, isolated, and low-value raw monitoring data into a highly condensed intelligent decision-making basis that reflects the overall health level and potential risks of the equipment. Through systematic data processing and feature fusion, it achieves accurate diagnosis of equipment operating status, providing reliable data support for subsequent intelligent inspections, and ultimately realizing accurate assessment and prediction of equipment operating status. The preprocessing uses equipment ID and event time as the core to lock in the associated data window. Standardization and timestamp alignment eliminate format differences and time misalignments between multi-source data, constructing a complete and consistent fused data package. This solves the problem of traditional data being difficult to analyze collaboratively, transforming raw data into usable data. The feature extraction network, tailored to the characteristics of different modal data, utilizes CNN, 3D-CNN, and LSTM. Visual, dynamic, and temporal feature vectors are extracted separately to accurately capture key information from multiple dimensions, such as equipment appearance, temperature, operational dynamics, and long-term status changes, transforming usable data into structured information. Graph attention networks introduce equipment topology relationships, fusing the multimodal features of a single device with contextual information from surrounding related devices to generate a final fused feature vector that combines its own characteristics with global correlations. This overcomes the limitations of single-dimensional data analysis and transforms structured information into systematic knowledge. Finally, a fully connected classifier outputs equipment operating status data, enabling intelligent and accurate judgment of equipment status. This transforms systematic knowledge into executable intelligence, solving the problems of missed detections and misjudgments in traditional manual inspections. It also provides a core basis for subsequent equipment hierarchy division and dynamic inspection cycle calculation, driving the centralized control system to shift from passive maintenance to proactive prediction.

[0040] A method for intelligent inspection of centralized control resources based on multi-source data fusion, wherein the specific method for classifying equipment hierarchically by combining the operating status data of substation equipment with the equipment criticality is as follows: The operating status data of substation equipment includes Business Impact Factor (BIF), Historical Failure Rate (HFR), Maintenance Cost Factor (MCF), and Location Criterion (LOC). The criticality index (DCI) of the equipment is calculated using the following formula: DCI = w1·BIF + w2·HFR + w3·MCF + w4·LOC; In the formula: w1 is the weight coefficient of Business Impact Factor (BIF), w2 is the weight coefficient of Historical Failure Rate (HFR), w3 is the weight coefficient of Maintenance Cost Factor (MCF), and w4 is the weight coefficient of Location Criterion (LOC); w1, w2, w3, and w4 are dynamically adjusted through machine learning. The system automatically classifies substation equipment into different levels based on DCI values; for example, DCI > 0.8 is the critical level, 0.4 < DCI ≤ 0.8 is the important level, and DCI ≤ 0.4 is the general level. This classification changes dynamically as the DCI value is updated.

[0041] A method for intelligent inspection of centralized control resources based on multi-source data fusion, wherein the specific method for determining the dynamic inspection cycle based on the dynamic inspection cycle algorithm that fuses equipment health status and predicted failure probability according to the divided equipment levels is as follows: Set a basic inspection cycle T_base for each level; A dynamic adjustment factor (DAF) is introduced. The dynamic adjustment factor (DAF) is jointly determined by the device's real-time health score H (from the diagnostic model, range [0,1]) and the predictive failure probability P (range [0,1]). The calculation formula is as follows: DAF = α·(1 - H) +β·P In the formula, α is the weighting coefficient of the real-time health H of the equipment and β is the weighting coefficient of the predictive failure probability P. The larger the value of DAF, the worse the equipment condition is, and the more frequent the inspection is required. The final dynamic inspection cycle T_dynamic is calculated based on the basic inspection cycle and the dynamic adjustment factor. The calculation formula is as follows: T_dynamic = T_base / (1 + γ·DAF) Where γ is the adjustment intensity coefficient. When the equipment health decreases or the failure probability increases, DAF increases and T_dynamic decreases, thereby dynamically increasing the inspection frequency.

[0042] A method for intelligent inspection of centralized control resources based on multi-source data fusion, wherein the specific method for planning the optimal inspection path is as follows: The centralized control system's intelligent resource management platform automatically generates routine inspection tasks based on a dynamic inspection cycle. The hybrid deep learning diagnostic module of the centralized control system resource intelligent management platform generates task priorities based on the location and urgency of routine inspection tasks, using the following method: Quantify the urgency of tasks, establish a weighted scoring model, and assess the inherent risks and importance of the tasks themselves.

[0043] Urgency score = f(equipment criticality, equipment health status, real-time event alarm level) Among these, equipment criticality assigns different base risk scores to equipment at different levels. Equipment health, represented by V_fused, is input into the anomaly detection or fault prediction model, and dynamically scored based on the output risk probability. Real-time event alarm level is quantified and scored according to the severity level of the event, where f() represents a weighted summation function.

[0044] A model that quantifies location and path efficiency and evaluates the immediate cost of task execution can be used to optimize resource utilization.

[0045] Location efficiency factor = g(agent - task shortest path distance, path cooperation) Here, the agent-task shortest path distance is calculated on the dynamic semantic map, taking the minimum of the shortest path distances between the task location and all currently idle agents. Path cooperation determines whether the task point is within the "smooth path" range of the agent's planned paths. g() represents the strategic equilibrium function.

[0046] A weighted fusion strategy is adopted to combine the two quantitative indicators mentioned above to generate the final priority score.

[0047] Task priority score = η·urgency score + θ·location efficiency factor Where η is the weighting coefficient of the urgency score, θ is the weighting coefficient of the location efficiency factor, and η>>θ.

[0048] The dynamic task scheduling module of the centralized control system's intelligent resource management platform uses an improved A* algorithm to plan the optimal inspection path for the agent on a dynamic semantic map. The specific calculation method is as follows: Cost(n) = g(n) + h(n) + λ·Task_Urgency In the formula: Cost(n) is the total cost of node n; g(n) is the actual cost incurred from the starting point to node n; h(n) is the heuristic estimated cost from node n to the target state; Task_Urgency is the task priority; λ is the urgency weight coefficient.

[0049] When an inspection task (such as "inspection of the front-end application server") is triggered, the resource intelligence collaborative task allocation engine of the centralized control system first decomposes it into a series of atomic sub-tasks, such as: {taking pictures of the equipment appearance, checking for abnormal noises, analyzing video anomalies}.

[0050] Establish an agent-task capability matching matrix M, where M(i, j) represents the suitability score (range [0,1]) of agent i (robot, human, video) in completing task j. The scoring criteria include accuracy, efficiency, cost, etc.

[0051] A greedy algorithm is employed, assigning the most suitable agent to each subtask based on the capability matching matrix M, thereby maximizing the overall task execution efficiency or accuracy. For example, taking pictures of the appearance is assigned to a robot, checking for abnormal noises is assigned to a human, and analyzing video anomalies is assigned to an intelligent video surveillance system.

[0052] A method for intelligent inspection of centralized control resources based on multi-source data fusion, wherein the corresponding intelligent agent autonomously inspects and collects data according to the optimal inspection path, specifically: The corresponding intelligent agent uses Simultaneous Localization and Mapping (SLAM) to perform real-time localization and obstacle avoidance based on the optimal inspection path, and autonomously moves to the target device point. Upon reaching the target device point, the corresponding intelligent agent automatically invokes the preset inspection script based on the device ID, controlling multiple sensors to work collaboratively. Visible light cameras capture information such as the equipment's appearance and indicator light status. Infrared thermal imagers detect temperature distribution and identify potential overheating risks. Sound sensors continuously collect information on the equipment's operating status. The cloud platform retrieves historical maintenance data, descriptive information (service interface) reported by mobile devices, and real-time data uploaded by fixed sensors from the semantic map index.

[0053] The process involves classifying substation equipment into hierarchical levels based on operational status data and equipment criticality. A dynamic inspection cycle algorithm, which integrates equipment health and predicted fault probability, is then used to determine the dynamic inspection cycle and plan the optimal inspection path. The core of this intelligent inspection method is the autonomous inspection by the corresponding intelligent agent following this optimal path. Its core function is to achieve precise allocation and efficient scheduling of inspection resources, breaking the traditional inspection model. The aim is to improve the targeting of inspections, reduce maintenance costs, and ensure the stable operation of the centralized control system through intelligent means. First, the DCI index is calculated by combining multiple dimensions such as business impact and historical failure rate to classify equipment into hierarchical levels. This accurately distinguishes between critical, important, and general equipment, ensuring that inspection resources are tilted towards core assets and solving the problem of uneven allocation of traditional inspection resources. Second, a basic inspection cycle is set based on the equipment level, and a DAF (Dynamic Inspection Flow Algorithm) is constructed using health and predicted fault probability. Dynamic adjustment factors automatically shorten the inspection cycle of equipment in poor condition or with high fault risk, achieving dynamic adaptation where more frequent inspections are conducted as the condition worsens, thus addressing the pain point of delayed fault warnings. Furthermore, an improved A* algorithm incorporates task urgency to plan the optimal path, ensuring both the efficiency of the inspection path and prioritizing high-priority tasks. Combined with intelligent agents utilizing SLAM positioning and multi-sensor collaborative inspection, this replaces manual labor for repetitive and complex inspections, solving the problems of low efficiency and susceptibility to missed or misjudged inspections in manual inspections. Finally, through hierarchical classification, dynamic cycles, optimal paths, and autonomous inspection, the system transforms inspection from passive response to proactive prediction, and from comprehensive coverage to precise focus, maximizing operational efficiency and quality. This provides a solid guarantee for the stable and intelligent operation of the centralized control system, constructing a dynamic, adaptive, and autonomous intelligent inspection system that is risk-oriented and aims for optimal resource allocation. This maximizes operational efficiency and cost-effectiveness, improves resource utilization efficiency, significantly enhances equipment reliability and system security, standardizes and intelligentizes inspection quality, and comprehensively reduces operational costs while ensuring the highest level of substation safety.

[0054] A centralized control resource intelligent inspection method based on multi-source data fusion, which also includes the construction of a large model knowledge base and intelligent decision support: The specific method for constructing the large model knowledge base is as follows: The fault reports automatically generated by corresponding intelligent agents and multi-source data fusion analysis provide corpus for training the large model knowledge base and output unstructured documents; We utilize a Large Language Model (LLM) to perform deep parsing of unstructured documents, identifying entities such as devices, components, fault codes, and parameters; relationships between entities; and entity attributes, obtaining triples. The specific method is as follows: Based on prompting engineering, we define output patterns and design a machine-parseable output format, JSON Schema, to constrain the model to generate structured entity and relation data.

[0055] Create complete instructions that include character settings, task descriptions, output format constraints, and input text.

[0056] The program will prompt you to input LLM, retrieve its generated JSON output, and parse it into usable triplet data.

[0057] Unstructured documents, extracted text fragments, and triple descriptions are stored in a vector database through a text embedding model to build a large model knowledge base. The specific scheme for the intelligent assisted decision-making is as follows: Operations personnel can input questions using natural language or structured forms; The user's question is vectorized, and a similarity search is performed in the vector database to retrieve the most relevant knowledge fragments. LLM is used to identify key entities from the question. Starting from these entities, multi-hop queries are performed in the knowledge graph to find related failure modes, causes, and solutions.

[0058] Example 2 Reference Figure 1 This is a second embodiment of the present invention, which provides a centralized control resource intelligent inspection system based on multi-source data fusion, comprising the following data acquisition module, data analysis module, and inspection path calculation module: The data acquisition module is used to assign a unique ID and coordinate unified label to the historical full-dimensional monitoring data of substation equipment and environment in a dynamic semantic map; The data analysis module is used to preprocess historical full-dimensional monitoring data with unique IDs and coordinate unified labels to obtain a fused data package of historical full-dimensional monitoring data. The modal features of the fused data package are extracted by a feature extraction network. The modal features of the fused data package are input into a graph attention network to generate a final fused feature vector. The operating status data of the substation equipment is output based on the final fused feature vector. The inspection path calculation module is used to divide the substation equipment into equipment levels by combining the operating status data of the equipment with the equipment criticality. Based on the divided equipment levels, the dynamic inspection cycle is determined by a dynamic inspection cycle algorithm that integrates equipment health and predicted fault probability, and the optimal inspection path is planned. The corresponding intelligent agent then performs autonomous inspections according to the optimal inspection path.

[0059] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent inspection of centralized control resources based on multi-source data fusion, characterized in that, It includes the following steps: Historical, multi-dimensional monitoring data of substation equipment and environment are assigned unique IDs and coordinates in a dynamic semantic map; Historical full-dimensional monitoring data with unique IDs and coordinate unified labels is preprocessed to obtain a fused data package of historical full-dimensional monitoring data. The modal features of the fused data package are extracted by a feature extraction network. The modal features of the fused data package are input into a graph attention network to generate a final fused feature vector. The operating status data of the substation equipment is output based on the final fused feature vector. The substation equipment operation status data is combined with the equipment criticality to divide the equipment into levels. Based on the divided equipment levels, the dynamic inspection cycle is determined by a dynamic inspection cycle algorithm that integrates equipment health and predicted fault probability, and the optimal inspection path is planned. The corresponding intelligent agent then performs autonomous inspections according to the optimal inspection path.

2. The intelligent inspection method for centralized control resources based on multi-source data fusion as described in claim 1, characterized in that: The historical, comprehensive monitoring data of the substation equipment and environment includes: Visible light and infrared images historically collected by the inspection intelligent agent; video data historically collected by intelligent video surveillance; and time-series data historically collected by fixed sensors.

3. The intelligent inspection method for centralized control resources based on multi-source data fusion as described in claim 1, characterized in that: The specific method for preprocessing historical full-dimensional monitoring data with unique IDs and unified coordinate labels to obtain the fused data package of historical full-dimensional monitoring data is as follows: The centralized control system resource intelligent management platform defines historical diagnostic events based on historical abnormal alarms of substation equipment. The event center is the equipment ID, and the event time is T_event. Define a window centered on the device ID and event time T_event, locking the diagnostic event device ID and its direct neighbors in the topology graph; define the interval [T_event - Δt, T_event + Δt], where Δt is the window radius set according to the device's dynamic characteristics; The historical full-dimensional monitoring data within the set window is standardized and timestamp aligned using the device ID as the key. The aligned data is packaged into a fused data package of historical full-dimensional monitoring data.

4. The intelligent inspection method for centralized control resources based on multi-source data fusion as described in claim 1, characterized in that: The specific method for extracting the modal features of the fused data packets through a feature extraction network is as follows: The image data in the fused data package is input into a convolutional neural network to process the visible light image to obtain the first visual feature vector V_rgb and the infrared image to obtain the second visual feature vector V_ir; the video data in the fused data package is input into a three-dimensional convolutional neural network to obtain the dynamic feature vector V_video; and the temporal data in the fused data package is input into a long short-term memory network to obtain the temporal feature vector V_ts.

5. The intelligent inspection method for centralized control resources based on multi-source data fusion as described in claim 4, characterized in that: The specific method for inputting the modal features of the fused data packets into a graph attention network to generate the final fused feature vector, and then outputting the substation equipment operating status data based on the final fused feature vector, is as follows: Define the target device node in the historical diagnostic events as v_target, and the initial node feature of the target device node v_target as X_initial, which is composed of the device's own multimodal features; X_initial = Concat(V_rgb, V_ir, ​​V_ts, V_video), where Concat() is a function used to concatenate two or more arrays or strings; Extract the device topology graph G=(V, E) from the dynamic semantic map; the node set V represents all devices, and the edge set E represents the physical connection, pipeline flow direction or logical control relationship between devices; The device topology graph G and the initial node features of the target device node v_target are input into the graph attention network and propagated through multiple layers of GAT to obtain the context-enhanced feature vector V_cnn of the target device node. The initial node features X_initial of the target device node v_target and the context-enhanced feature vector V_cnn are input together into the multi-head cross-modal attention fusion layer; The multi-head cross-modal attention fusion layer performs a weighted summation of different feature vectors to obtain the final fused feature vector V_fused; The final fused feature vector V_fused is input into a fully connected classifier, which outputs the operating status data of the substation equipment.

6. The intelligent inspection method for centralized control resources based on multi-source data fusion as described in claim 1, characterized in that: The specific method for classifying substation equipment hierarchically by combining the operating status data of substation equipment with the equipment criticality is as follows: The operating status data of substation equipment includes Business Impact Factor (BIF), Historical Failure Rate (HFR), Maintenance Cost Factor (MCF), and Location Criterion (LOC). The criticality index (DCI) of the equipment is calculated using the following formula: DCI = w1·BIF + w2·HFR + w3·MCF + w4·LOC; In the formula: w1 is the weight coefficient of Business Impact Factor (BIF), w2 is the weight coefficient of Historical Failure Rate (HFR), w3 is the weight coefficient of Maintenance Cost Factor (MCF), and w4 is the weight coefficient of Location Criterion (LOC). w1, w2, w3, and w4 are dynamically adjusted through machine learning. The substation equipment is automatically classified into different levels based on the DCI value.

7. The intelligent inspection method for centralized control resources based on multi-source data fusion as described in claim 1, characterized in that: The specific method for determining the dynamic inspection cycle based on the divided equipment hierarchy and the dynamic inspection cycle algorithm that fuses equipment health and predicted failure probability is as follows: Set a basic inspection cycle T_base for each equipment level; A dynamic adjustment factor (DAF) is introduced. The dynamic adjustment factor (DAF) is jointly determined by the real-time health score H and the predictive failure probability P of the device. The real-time health score H comes from the intelligent diagnostic model and ranges from [0,1]. The predictive failure probability P ranges from [0,1]. The formula for calculating the dynamic adjustment factor (DAF) is as follows: DAF =α·(1 - H) +β·P In the formula, α is the weighting coefficient of the real-time health of the equipment H and β is the weighting coefficient of the predictive failure probability P. The final dynamic inspection cycle T_dynamic is calculated based on the basic inspection cycle and the dynamic adjustment factor. The calculation formula is as follows: T_dynamic = T_base / (1 + γ·DAF) Where γ is the intensity adjustment coefficient.

8. The intelligent inspection method for centralized control resources based on multi-source data fusion as described in claim 1, characterized in that: The specific method for planning the optimal inspection path is as follows: Regular inspection tasks are automatically generated based on the dynamic inspection cycle T_dynamic. Generate task priorities based on the location and urgency of routine inspection tasks; The improved A* algorithm is used to plan the optimal inspection path for the agent on the dynamic semantic map. The specific calculation method is as follows: Cost(n) = g(n) + h(n) + λ·Task_Urgency In the formula: Cost(n) is the total cost of node n; g(n) is the actual cost incurred from the starting point to node n; h(n) is the heuristic estimated cost from node n to the target state; Task_Urgency is the task priority; λ is the urgency weight coefficient.

9. The intelligent inspection method for centralized control resources based on multi-source data fusion as described in claim 1, characterized in that: The specific method for the corresponding intelligent agent to autonomously inspect according to the optimal inspection path is as follows: The corresponding intelligent agent uses SLAM for real-time localization and obstacle avoidance based on the optimal inspection path, and autonomously moves to the target device point. Upon reaching the target device point, the corresponding intelligent agent automatically invokes the preset inspection script based on the device ID, controlling multiple sensors to work collaboratively. Visible light images are captured using a visible light camera; infrared images are captured using an infrared thermal imager; and equipment operating status information is continuously collected using a sound sensor. The cloud platform retrieves historical maintenance data, descriptive information reported by mobile devices, and real-time data uploaded by fixed sensors from the semantic map index.

10. A centralized control resource intelligent inspection system based on multi-source data fusion, characterized in that: It includes the following modules: data acquisition, data analysis, and inspection path calculation: The data acquisition module is used to assign a unique ID and coordinate unified label to the historical full-dimensional monitoring data of substation equipment and environment in a dynamic semantic map; The data analysis module is used to preprocess historical full-dimensional monitoring data with unique IDs and coordinate unified labels to obtain a fused data package of historical full-dimensional monitoring data. The modal features of the fused data package are extracted by a feature extraction network. The modal features of the fused data package are input into a graph attention network to generate a final fused feature vector. The operating status data of the substation equipment is output based on the final fused feature vector. The inspection path calculation module is used to divide the substation equipment into equipment levels by combining the operating status data of the equipment with the equipment criticality. Based on the divided equipment levels, the dynamic inspection cycle is determined by a dynamic inspection cycle algorithm that integrates equipment health and predicted fault probability, and the optimal inspection path is planned. The corresponding intelligent agent then performs autonomous inspections according to the optimal inspection path.

Images