A context-aware-based power supply equipment fault dynamic maintenance scheme intelligent recommendation method and system

Abstract

The application provides a context-aware-based intelligent recommendation method and system for power supply equipment fault dynamic maintenance scheme, and belongs to the field of intelligent operation and maintenance of power systems, and solves the technical problems of static scheme mismatch, lack of context awareness, low recommendation accuracy, low efficiency, difficulty in knowledge reuse and lack of closed-loop optimization. The application collects multi-source operation and maintenance data of power supply equipment and extracts multi-dimensional context features, builds a maintenance knowledge graph, uses an FCM+SVM model to complete fault diagnosis, dynamically recommends a maintenance scheme by improving a collaborative filtering algorithm, relies on DQN reinforcement learning to realize closed-loop optimization, and additionally adds a knowledge graph supplementary recommendation and equipment anomaly early warning function. The application can comprehensively perceive operation and maintenance scenes, improve the matching degree of the maintenance scheme and the personnel adoption rate, realize efficient reuse of maintenance knowledge, and improve operation and maintenance efficiency. The system has closed-loop self-optimization capability and strong robustness, can also realize predictive maintenance of equipment, and can adapt to various power supply operation and maintenance scenes.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for power systems, and in particular to an intelligent recommendation method and system for dynamic maintenance schemes for power supply equipment faults based on context awareness. Background Technology

[0002] As a core component of the power system, the operational reliability of power supply equipment directly affects the safe and stable operation of the power system. With the rapid development of urban rail transit, smart grids, and other fields, the scale of power supply equipment is becoming increasingly large and its structure increasingly complex. The suddenness and diversity of equipment failures bring enormous challenges to operation and maintenance management. When power supply equipment fails, how to quickly and accurately formulate a reasonable maintenance plan is a key factor affecting the fault recovery time and reducing power outage losses.

[0003] The existing recommended technologies for fault repair of power supply equipment have the following main shortcomings: (1) Mismatch between static solutions and actual operating conditions: Existing technologies mostly adopt maintenance solution recommendation methods based on fixed rules. The recommendation results are based on a pre-set expert rule base and do not fully consider dynamic factors such as the real-time operating status of the equipment and the current environmental conditions, resulting in a large deviation between the recommended solutions and the actual maintenance conditions. According to statistics, the solution matching rate of traditional static recommendation methods is usually no more than 70%, and even less than 50% in complex fault scenarios.

[0004] (2) Lack of multi-dimensional context awareness: Existing methods often only consider the single factor of fault type, without fully integrating multi-dimensional contextual information such as equipment attributes, operating environment, historical maintenance records, personnel skill level, and tool availability. The lack of contextual information makes it difficult for recommended solutions to meet actual field needs in terms of operability, safety, and economy.

[0005] (3) Low recommendation accuracy: Due to the lack of a personalized recommendation mechanism, existing methods struggle to make accurate recommendations based on maintenance personnel's historical preferences and skill characteristics, resulting in a low adoption rate of maintenance solutions. Survey data shows that the adoption rate of solutions using traditional recommendation methods is usually no more than 60%, and a large number of recommended solutions are discarded because they do not conform to the actual situation on site.

[0006] (4) Low efficiency: Traditional methods rely on maintenance personnel manually consulting maintenance manuals, historical cases, and other materials, with an average search time of no less than 20 minutes. In emergency fault scenarios, the inefficient solution search process seriously affects the fault recovery speed, increases power outage losses and safety risks.

[0007] (5) Maintenance knowledge is difficult to reuse: A large amount of valuable maintenance experience is scattered in personal notes, oral records or unstructured documents, lacking a unified knowledge management and reuse mechanism. With the retirement of senior maintenance personnel, a large amount of tacit knowledge is at risk of being lost.

[0008] (6) Lack of closed-loop optimization mechanism: Existing recommendation systems are usually open-loop architectures, lacking mechanisms for tracking and feedback on the effectiveness of solution implementation and for iterative optimization of the model. The performance of recommendation models gradually degrades over time and cannot adapt to changes brought about by equipment aging and technological updates.

[0009] In recent years, artificial intelligence (AI) technology has been widely applied in the field of power system operation and maintenance. Machine learning-based fault diagnosis methods (such as support vector machines, neural networks, and deep learning) have made significant progress in fault classification accuracy; collaborative filtering-based recommendation algorithms have demonstrated excellent personalized recommendation capabilities in e-commerce and content recommendation; and reinforcement learning technology has shown great potential in sequence decision optimization. However, how to deeply integrate these advanced technologies with the professional knowledge of power equipment maintenance to build a context-aware, dynamic, intelligent recommendation system for maintenance solutions remains a pressing technical challenge. Summary of the Invention

[0010] This invention proposes an intelligent recommendation method and system for dynamic maintenance schemes for power equipment faults based on context awareness. By integrating multi-source operation and maintenance data, extracting multi-dimensional contextual features, constructing a structured maintenance knowledge base, using an FCM+SVM hybrid model for fault diagnosis, utilizing an improved collaborative filtering algorithm for dynamic recommendation, and continuously optimizing based on reinforcement learning, this invention addresses the technical problems in existing technologies such as static scheme mismatch, lack of context awareness, low recommendation accuracy, low efficiency, difficulty in knowledge reuse, and lack of closed-loop optimization.

[0011] A context-aware intelligent recommendation method for dynamic maintenance solutions for power supply equipment faults includes the following steps: S1. Multi-source operation and maintenance data acquisition: Through sensor networks, SCADA systems, historical operation and maintenance databases and enterprise resource planning systems deployed at the power supply equipment site, real-time acquisition of equipment basic parameters, operating status data, historical maintenance records, environmental parameters and maintenance personnel information is carried out. The data update frequency is no less than 1Hz, and the acquired multi-source heterogeneous data is cleaned, denoised, time-series aligned and standardized preprocessed. S2. Context Feature Extraction: A 12-dimensional context feature vector is extracted from six dimensions: equipment attributes, operating environment, historical maintenance, real-time status, personnel skills, and resource availability. These 12-dimensional context features include: equipment type, equipment model, years of operation, ambient temperature, ambient humidity, historical failure frequency, historical maintenance methods, real-time operating status, real-time load rate, maintenance personnel skill level, available tool list, and spare parts inventory status. The 12-dimensional context feature vector is then normalized to obtain a standard context feature vector. ; S3. Maintenance Knowledge Base Construction: Standardize and encode maintenance plans using structured description templates, including fault phenomenon description, fault cause analysis, maintenance operation steps, required tools and materials list, personnel skill requirements, safety precautions, estimated maintenance time and historical success rate. At the same time, knowledge graph technology is used to establish semantic relationships between fault types, equipment categories, maintenance plans, and tool resources to form a reasonable maintenance knowledge network. S4. Fault diagnosis and maintenance requirement analysis: A hybrid model combining FCM clustering and SVM is adopted. First, the FCM algorithm is used to perform fuzzy clustering analysis on equipment fault samples, and then the SVM classifier is used to identify fault types and determine maintenance requirements based on the clustering results. The clustering objective function of the FCM algorithm is:

[0012] The membership update formula is:

[0013] In the formula, For clustering loss function, The total number of samples, The number of cluster centers. For the first The nth sample pair Membership degree of each cluster center For fuzzy coefficients and , For the first The feature vector of each sample For the first Cluster center vectors; S5. Dynamic Recommendation of Repair Solutions: An improved collaborative filtering algorithm based on context weights is adopted to comprehensively consider user bias, solution bias, neighboring user ratings and solution similarity, calculate the recommendation score of the repair solution, and generate a Top-K recommendation list based on the score. The formula for calculating the recommendation score is as follows:

[0014] In the formula, For users Regarding the plan Recommended score, The overall average score. For users Scoring bias, For the plan Scoring bias, For users The set of neighboring users For neighboring users Context-adaptive weights For users Regarding the plan Historical ratings For the plan With the plan The similarity measure is calculated using cosine similarity or Pearson correlation coefficient; S6. Solution Feedback and Continuous Optimization: Based on the DQN reinforcement learning framework, a reward function is constructed according to the actual execution results of the maintenance solution. The recommendation model parameters are continuously optimized through iterative training. The reward function is defined as follows:

[0015] In the formula, A reward will be given for successful repairs. Penalty for delayed repair time, Penalty for maintenance costs, This is the time delay weighting coefficient. This is the cost weighting coefficient.

[0016] Furthermore, the multi-source operation and maintenance data collection mentioned in S1 includes: Basic equipment parameters, including equipment type, equipment model, rated voltage, rated current, rated power, manufacturer, commissioning date, and years of operation; Operating status data, including real-time voltage, real-time current, active power, reactive power, power factor, equipment temperature, and switch status, are collected at a frequency of no less than 1Hz. Historical maintenance records include the time of the historical failure, the type of failure, the symptoms of the failure, the maintenance measures, the maintenance time, the maintenance cost, and the maintenance personnel information; Environmental parameters, including ambient temperature, ambient humidity, altitude, air quality index, and weather conditions; Maintenance personnel information includes personnel skill level, professional qualifications, historical maintenance success rate, and current workload.

[0017] Furthermore, the process of extracting contextual features described in S2 is as follows: Feature engineering is performed on the collected multi-source raw data, including feature selection, feature transformation, and feature dimensionality reduction; The 12-dimensional contextual features are normalized using the Z-Score normalization method:

[0018] In the formula, For the first dimensional original eigenvalues, For the first The sample mean of the dimensional feature. For the first The sample standard deviation of the dimensional feature; Principal component analysis was performed on the normalized 12-dimensional eigenvectors to reduce dimensionality, and the top eigenvectors with a cumulative contribution rate of not less than 85% were selected. One principal component is used to obtain the dimensionality-reduced context feature vector. .

[0019] Furthermore, the knowledge graph construction process of the maintenance knowledge base described in S3 includes: Define the ontology model of the knowledge graph, including entity types: equipment, fault type, fault phenomenon, maintenance plan, tool, personnel, and environmental conditions; Define the types of relationships between entities: belonging to, causing, manifesting as, applicable to, requiring, possessing, and limited to; The Neo4j graph database is used to store the knowledge graph, and the Cypher query language is used to implement maintenance scheme reasoning based on graph traversal. The TransE or RotatE knowledge graph embedding algorithm is used to vectorize entities and relations, enabling semantic retrieval based on vector similarity.

[0020] Furthermore, the training and inference process of the FCM+SVM hybrid model described in S4 includes: The historical fault samples were fuzzy clustered using the FCM algorithm to obtain... Each fault cluster and the membership matrix of each sample to each cluster. ; The membership matrix is ​​concatenated with the original eigenvectors to obtain the enhanced eigenvectors. ; An SVM classifier with the RBF kernel function is used to train the enhanced feature vectors to obtain a fault type classification model; For a new fault sample, first calculate its membership degree in each cluster, construct an enhanced feature vector, and then input it into the SVM classifier to output the fault type and the corresponding maintenance requirement level.

[0021] Furthermore, the improved collaborative filtering algorithm described in S5 uses context-adaptive weights. Dynamically calculated based on the current context feature vector:

[0022] In the formula, This is the context feature vector of the current fault scenario. For neighboring users The context feature vector corresponding to the fault scenario, This is the bandwidth parameter of the Gaussian kernel function.

[0023] Furthermore, the model structure of DQN reinforcement learning described in S6 includes: state space It consists of the context feature vector of the currently faulty device, the fault type code, and the available resource code; Action space : This is a discrete set of all maintenance solutions in the maintenance knowledge base, with one maintenance solution recommended for each action; Q-network: It adopts a fully connected neural network with two hidden layers. The input layer dimension is the state space dimension, the number of hidden layer neurons are 256 and 128 respectively, the activation function is ReLU, and the output layer dimension is the action space dimension. Target network: Same structure as the Q network, parameters are updated using a soft update strategy. ,in This is a soft update coefficient; Experience replay: A priority experience replay mechanism is adopted to store transferred samples. Sampling is performed based on the TD error priority.

[0024] Furthermore, the dynamic recommendation of maintenance solutions described in S5 also includes: when the confidence level of the collaborative filtering recommendation result is lower than a preset threshold, triggering knowledge graph inference recommendation as a supplement, namely: Retrieve relevant maintenance solution nodes from the knowledge graph based on the current fault type; Perform multi-hop path traversal along the knowledge graph relationship edges to find a repair solution that matches the current equipment type, fault cause, and available tools; By combining historical success rates and estimated maintenance times stored in the knowledge graph, the solutions are sorted to generate a list of recommended alternatives.

[0025] Furthermore, the method also includes anomaly detection and early warning steps: The isolated forest algorithm is used to detect anomalies in the context feature vector extracted in step S2 and identify equipment states that deviate from normal operating mode. When the abnormal score exceeds the preset threshold, an early warning message is automatically generated and pushed to the operation and maintenance management personnel to achieve predictive maintenance.

[0026] A context-aware intelligent recommendation system for dynamic maintenance solutions of power supply equipment faults, applied to the aforementioned context-aware intelligent recommendation method for dynamic maintenance solutions of power supply equipment faults, includes: The multi-source operation and maintenance data acquisition module is used to collect equipment parameters, operating status, historical maintenance records, environmental parameters and personnel information in real time through sensor networks, SCADA systems and historical operation and maintenance databases, and to perform data preprocessing. The context feature extraction module is used to extract 12-dimensional context feature vectors from six dimensions: equipment attributes, operating environment, historical maintenance, real-time status, personnel skills, and resource availability, and then perform normalization and dimensionality reduction processing. The maintenance knowledge base construction module is used to standardize the coding of maintenance solutions using structured description templates and to establish semantic relationships between fault types, equipment categories, maintenance solutions, and tool resources using knowledge graph technology. The fault diagnosis and maintenance requirement analysis module is used to perform cluster analysis and classification identification of equipment faults using an FCM+SVM hybrid model to determine the fault type and maintenance requirement level. The maintenance plan dynamic recommendation module is used to calculate maintenance plan recommendation scores using an improved collaborative filtering algorithm based on context weights, and generate a Top-K recommendation list based on the scores. The solution feedback and continuous optimization module is used to construct reward signals based on the actual execution feedback of maintenance solutions using the DQN reinforcement learning framework, and to continuously optimize the recommendation model parameters through iterative training; and The knowledge graph reasoning module is used to trigger semantic reasoning recommendations based on knowledge graphs when the confidence of the collaborative filtering recommendation results is lower than a preset threshold, serving as a supplement to collaborative filtering recommendations.

[0027] Compared with the prior art, the present invention achieves significant beneficial effects through the above technical solution: (1) High accuracy of solution matching: By integrating 12-dimensional contextual features to fully characterize the fault scenario, and combining the accurate diagnosis of the FCM+SVM hybrid model with the personalized recommendation of improved collaborative filtering, the solution matching rate reaches more than 92%, which is more than 22 percentage points higher than the traditional static recommendation method.

[0028] (2) Comprehensive perception of multi-dimensional context: Innovatively extracting 12-dimensional context features from six dimensions, namely equipment attributes, operating environment, historical maintenance, real-time status, personnel skills, and resource availability, to achieve comprehensive perception of maintenance scenarios and ensure that the recommended solution is technically feasible, safe and controllable, and resource-configurable.

[0029] (3) High recommendation accuracy: By introducing a context-adaptive weight mechanism and knowledge graph semantic reasoning, personalized recommendations based on scene similarity are realized. The adoption rate of maintenance personnel reaches more than 85%, which is more than 25 percentage points higher than the traditional method.

[0030] (4) Significantly improved recommendation efficiency: The fully automated data collection, feature extraction, fault diagnosis and solution recommendation process takes no more than 2 minutes from the occurrence of the fault to the generation of the recommendation list, which is more than 10 times more efficient than the traditional manual search method (average more than 20 minutes).

[0031] (5) Effective reuse of maintenance knowledge: The combination of structured maintenance knowledge base and knowledge graph technology will systematize, standardize and digitize the storage of scattered maintenance experience, support intelligent retrieval and reasoning based on semantic association, and realize the continuous accumulation and efficient reuse of maintenance knowledge.

[0032] (6) Closed-loop continuous optimization: The feedback optimization mechanism based on the DQN reinforcement learning framework enables the recommendation model to continuously learn and evolve according to the actual performance. The recommendation accuracy continues to improve at a rate of no less than 2% per week, and the model has adaptability.

[0033] (7) The system is robust: When the confidence of the collaborative filtering recommendation result is low, the system automatically triggers knowledge graph reasoning as a supplementary recommendation mechanism. The two recommendation strategies serve as backups for each other, ensuring that reliable recommendation results can be provided in various scenarios.

[0034] (8) Wide range of applications: The method of the present invention is not only applicable to urban rail transit power supply systems, but can also be extended to various power supply equipment operation and maintenance scenarios such as substations, distribution networks, and industrial power distribution, and has good versatility and scalability. Attached Figure Description

[0035] Figure 1 This is an overall flowchart of the intelligent recommendation method for dynamic maintenance schemes of power equipment faults based on context awareness proposed in this invention; Figure 2 This is a flowchart of multi-source operation and maintenance data acquisition and preprocessing provided in an embodiment of the present invention; Figure 3 A schematic diagram of the 12-dimensional context feature extraction framework provided in an embodiment of the present invention; Figure 4 This is a two-layer architecture diagram of the maintenance knowledge base provided in an embodiment of the present invention; Figure 5 This is a structural diagram of the FCM+SVM hybrid fault diagnosis model provided in an embodiment of the present invention; Figure 6The flowchart of the improved collaborative filtering recommendation algorithm provided in the embodiments of the present invention is shown below; Figure 7 This is a diagram of the DQN reinforcement learning closed-loop optimization framework provided in an embodiment of the present invention; Figure 8 This is a diagram illustrating the overall architecture of an intelligent recommendation system provided in an embodiment of the present invention. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Reference Figures 1-8 As shown, a context-aware intelligent recommendation method for dynamic maintenance solutions for power supply equipment faults includes the following steps: S1: Multi-source operation and maintenance data collection By deploying sensor networks, SCADA (Supervisory Control and Data Acquisition) systems, historical maintenance databases, and Enterprise Resource Planning (ERP) systems at power supply equipment sites, comprehensive multi-source heterogeneous data related to equipment maintenance decisions is collected. The collected data categories include: basic equipment parameters (equipment type, model, rated parameters, years of operation, etc.), operating status data (real-time voltage, current, power, temperature, switch status, etc., frequency not less than 1Hz), historical maintenance records (fault time, type, phenomenon, measures, costs, etc.), environmental parameters (temperature, humidity, altitude, weather, etc.), and maintenance personnel information (skill level, qualifications, success rate, load, etc.). The collected raw data undergoes cleaning (outlier removal, missing value imputation), denoising (moving average, wavelet threshold denoising), time-series alignment (timestamp-based multi-source data synchronization), and standardization preprocessing (dimensional unification, numerical normalization) to provide a high-quality data foundation for subsequent analysis.

[0038] S2: Contextual Feature Extraction Extracting 12-dimensional contextual feature vectors from six dimensions comprehensively characterizes the contextual information of fault repair scenarios: (1) Equipment attribute dimension: equipment type (circuit breaker / transformer / cable, etc.), equipment model, and years of operation; (2) Operating environment dimension: ambient temperature and ambient humidity; (3) Historical maintenance dimensions: frequency of historical failures and historical maintenance methods; (4) Real-time status dimension: real-time operating status (normal / alarm / fault), real-time load rate; (5) Personnel skills dimension: Maintenance personnel skill level (basic / intermediate / advanced / expert); (6) Resource availability dimension: list of available tools and inventory status of spare parts.

[0039] The 12-dimensional context features are Z-score normalized to eliminate dimensional differences and obtain a standard context feature vector. Based on this, principal component analysis (PCA) is used for feature dimensionality reduction, selecting principal components with a cumulative contribution rate of no less than 85%, thereby reducing computational complexity while retaining key information.

[0040] S3: Building a Maintenance Knowledge Base The maintenance knowledge base is constructed using a two-tier architecture of "structured description templates + knowledge graph association". Structured Description Layer: Standardized coding templates are developed for each repair plan, including fault phenomenon description (text description + structured tags), fault cause analysis (root cause classification + probability weight), repair operation steps (step-by-step description + estimated man-hours), required tools and materials list (tool name / specification / quantity), personnel skill requirements (minimum skill level + recommended number of personnel), safety precautions (power outage area / safe distance / protective measures), estimated repair time (optimistic / pessimistic / most likely estimate), and historical success rate (cumulative number of executions / number of successes / success rate).

[0041] Knowledge Graph Association Layer: Constructs a semantic network with devices, faults, solutions, tools, personnel, and environment as entity nodes, and "belongs to," "leads to," "manifests as," "applies to," "needs," "possesses," and "restricted to" as relation types. It uses the Neo4j graph database for storage and leverages knowledge graph embedding algorithms (TransE / RotatE) to achieve vectorized representations of entities and relations, supporting intelligent retrieval based on semantic similarity and reasoning recommendation based on graph traversal.

[0042] S4: Fault Diagnosis and Repair Needs Analysis A hybrid model combining FCM (Fuzzy C-Means) clustering and SVM (Support Vector Machine) is used for fault diagnosis. The FCM algorithm, by introducing the concept of fuzzy membership, allows a fault sample to belong to multiple fault categories to varying degrees, making it more suitable for handling the fuzzy and uncertain characteristics of power supply equipment faults.

[0043] The clustering objective function of the FCM algorithm is:

[0044] The membership update formula is:

[0045] In the formula, For clustering loss function, The total number of samples, The number of cluster centers. For the first The nth sample pair The membership degree of each cluster center satisfies , For fuzzy coefficients and (usually taken) ), For the first The feature vector of each sample For the first Cluster center vectors.

[0046] The SVM classifier uses the RBF (Radial Basis Function) kernel function to fuse the membership features obtained from FCM clustering with the original features before inputting them into the SVM for fault type identification. The advantage of the hybrid model is that FCM clustering captures the distribution structure of fault samples, while the SVM classifier constructs the optimal classification hyperplane in the high-dimensional feature space. The combination of the two effectively improves the accuracy of fault diagnosis.

[0047] Step S5: Dynamic Recommendation of Repair Solutions An improved collaborative filtering algorithm based on context weights is used to recommend maintenance solutions. Compared with traditional collaborative filtering algorithms, the improvement of this invention lies in the introduction of a context-adaptive weight mechanism, which enables the recommendation results to be dynamically adjusted according to the specific context of the current fault scenario.

[0048] The recommended score is calculated using the following formula:

[0049] In the formula, For users (i.e., the maintenance requirements corresponding to the current fault scenario) Regarding the solution Recommended score, The overall average score. For users The rating bias (reflecting how much the user's rating habits deviate from the average level). For the plan The scoring deviation (reflecting the degree to which the quality level of the scheme deviates from the average level). For users The set of neighboring users (determined by a similarity threshold or the K-nearest neighbor method). For neighboring users Context-adaptive weights For users Regarding the plan Historical ratings For the plan With the plan Similarity metric.

[0050] Context-adaptive weights The similarity is dynamically calculated based on the distance between the context feature vector of the current fault scenario and the context feature vectors of neighboring users' corresponding fault scenarios. The closer the distance, the higher the weight, making the ratings of neighboring users similar to the current scenario have a greater impact on the recommendation results. The cosine similarity or Pearson correlation coefficient is used to calculate the scheme, taking into account factors such as the matching degree of the fault type, equipment applicability, and operational complexity.

[0051] Based on recommended score Sort the solutions in descending order and take the top one. A recommended list of solutions is generated. The recommended repair solutions include detailed operating steps, required tools and materials, personnel configuration suggestions, safety precautions, and estimated completion time, presented to repair personnel in a structured manner.

[0052] Step S6: Solution Feedback and Continuous Optimization Based on the DQN (DeepQ-Network) reinforcement learning framework, a closed-loop optimization mechanism for maintenance plan recommendation is constructed. The maintenance plan recommendation process is modeled as a Markov decision process (MDP), where the state space consists of the context features of the current fault, the fault type, and the resource availability encoding; the action space consists of all plans in the maintenance knowledge base; and the reward function is calculated based on the actual effect of plan execution.

[0053] The reward function is defined as:

[0054] In the formula, A reward is given for successful repairs (positive value for successful repairs, negative value for failed repairs). Penalty for repair time delay (the portion of the actual repair time that exceeds the estimated time). Penalty for repair costs (the portion of actual costs exceeding the budget). This is the time delay weighting coefficient (reflecting the relative importance of time efficiency). This is the cost weighting coefficient (reflecting the relative importance of economic efficiency).

[0055] The DQN model employs a dual-network structure (Q-network and target network) and a priority experience replay mechanism, learning the optimal recommendation strategy through continuous interaction with the environment. After each maintenance task is completed, the system collects maintenance execution feedback (success or failure, actual time spent, actual cost, and maintenance personnel evaluation), updates the experience replay pool, and triggers model training, enabling continuous evolution of recommendation capabilities.

[0056] Specifically, this invention collects multi-source operation and maintenance data related to power supply equipment in real time and performs preprocessing tasks such as cleaning, noise reduction, time-series alignment, and standardization. It then extracts a twelve-dimensional contextual feature vector based on six dimensions: equipment attributes, operating environment, historical maintenance, real-time status, personnel skills, and resource availability, and performs normalization processing. This allows for the complete capture of various dynamic information at the fault site, overcoming the limitations of traditional technologies that rely solely on single fault information to formulate solutions. It effectively solves the problem of mismatch between static maintenance plans and actual on-site conditions. Furthermore, this invention uses structured description templates combined with knowledge graph technology to build a maintenance knowledge base. It standardizes and encodes various maintenance plans and establishes semantic relationships between different entities and information, integrating scattered maintenance experience into a searchable and reasonable knowledge network. This achieves the accumulation and efficient reuse of operation and maintenance knowledge, preventing the loss of valuable implicit operation and maintenance experience. This invention utilizes a hybrid model built on fuzzy C-means clustering and support vector machines to perform fault diagnosis. This model accurately identifies fault types and determines corresponding maintenance needs. An improved collaborative filtering algorithm incorporating context-adaptive weights is then used to calculate maintenance solution recommendation scores and generate a recommendation list. Personalized recommendations are made based on the similarity of different fault scenarios, effectively improving the matching degree between maintenance solutions and actual scenarios, and increasing the willingness of maintenance personnel to adopt recommended solutions. The entire process of data collection, feature extraction, fault diagnosis, and solution recommendation is fully automated, eliminating the need for maintenance personnel to manually consult maintenance materials and significantly improving the efficiency of solution acquisition during fault handling. Furthermore, this invention leverages a deep Q-network reinforcement learning framework to build a closed-loop optimization mechanism, continuously updating model parameters and optimizing recommendation strategies based on the actual execution results of maintenance solutions. This allows the model to continuously adapt to real-world situations such as equipment aging and changes in the maintenance environment, improving the performance degradation problem of recommendation models under traditional open-loop architectures. The entire technical solution is adaptable to various power supply maintenance scenarios, including urban rail transit power supply systems, substations, and distribution networks, demonstrating strong versatility and scalability.

[0057] Furthermore, the multi-source operation and maintenance data collection mentioned in S1 includes: Basic equipment parameters, including equipment type, equipment model, rated voltage, rated current, rated power, manufacturer, commissioning date, and years of operation; Operating status data, including real-time voltage, real-time current, active power, reactive power, power factor, equipment temperature, and switch status, are collected at a frequency of no less than 1Hz. Historical maintenance records include the time of the historical failure, the type of failure, the symptoms of the failure, the maintenance measures, the maintenance time, the maintenance cost, and the maintenance personnel information; Environmental parameters, including ambient temperature, ambient humidity, altitude, air quality index, and weather conditions; Maintenance personnel information includes personnel skill level, professional qualifications, historical maintenance success rate, and current workload.

[0058] Specifically, this invention refines the content of multi-source operation and maintenance data collection, clearly defining five major categories: basic equipment parameters, operating status data, historical maintenance records, environmental parameters, and maintenance personnel information. For operating status data, a collection frequency of no less than 1Hz is set, enabling comprehensive and detailed collection of various raw information related to power supply equipment operation and maintenance, ensuring data integrity, real-time performance, and richness from the source. Detailed basic equipment parameters fully present inherent attributes such as equipment model, rated operating conditions, commissioning and service status, providing relevant equipment-specific evidence for subsequent feature extraction and fault analysis. High-frequency collected operating status data can track dynamic operating indicators such as voltage, current, power, equipment temperature, and switch status in real time, accurately reflecting the current operating condition of the equipment. Comprehensive historical maintenance records integrate past faults, handling processes, time spent, and costs, which can be used to analyze fault patterns and learn from historical operation and maintenance experience. Various environmental parameters accurately recreate the external environmental conditions at the time of the fault. Maintenance personnel information covers skill qualifications, work performance, and current workload, fully considering the actual situation of on-site human resources. Based on clearly categorized and detailed collected data, subsequent steps such as contextual feature extraction, fault diagnosis, and maintenance plan recommendation can be carried out by comprehensively considering multiple factors such as equipment, operating conditions, environment, personnel, and historical maintenance. This effectively avoids analytical biases caused by incomplete data collection and missing information, and further improves the problem that traditional technologies rely on only a small amount of information to formulate maintenance plans and are difficult to adapt to complex on-site conditions.

[0059] Furthermore, the process of extracting contextual features described in S2 is as follows: Feature engineering is performed on the collected multi-source raw data, including feature selection, feature transformation, and feature dimensionality reduction; The 12-dimensional contextual features are normalized using the Z-Score normalization method:

[0060] In the formula, For the first dimensional original eigenvalues, For the first The sample mean of the dimensional feature. For the first The sample standard deviation of the dimensional feature; Principal component analysis was performed on the normalized 12-dimensional eigenvectors to reduce dimensionality, and the top eigenvectors with a cumulative contribution rate of not less than 85% were selected. One principal component is used to obtain the dimensionality-reduced context feature vector. .

[0061] Specifically, this invention performs a series of feature engineering processes on the collected multi-source raw data, including feature selection, feature transformation, and feature dimensionality reduction. Simultaneously, it employs Z-Score standardization to normalize the twelve-dimensional context features, effectively eliminating dimensional differences between features of different dimensions. This ensures that various feature data have a unified computational standard, avoiding interference from inconsistent numerical scales in subsequent model calculations and feature analysis. After normalization, this invention further reduces the dimensionality of the feature vectors through principal component analysis, selecting principal components with a cumulative contribution rate of no less than 85% to construct new feature vectors. While preserving as much effective information as possible from the original context, such as equipment attributes, operating environment, personnel skills, and maintenance history, this invention simplifies feature dimensions, reduces data redundancy, and effectively lowers the computational overhead of subsequent fault diagnosis and algorithm calculations. After standardization and dimensionality reduction optimization, the feature vectors can more accurately depict the complete operation and maintenance scenario, enhance the system's ability to perceive multi-dimensional contextual information, improve the problem of inaccurate scenario restoration caused by coarse feature processing, feature redundancy or dimensional imbalance in traditional technologies, and provide high-quality input data for fault diagnosis models combining FCM and SVM and improved collaborative filtering recommendation algorithms. From the feature level, it ensures the accuracy of fault identification and the degree of matching between maintenance plans and actual working conditions.

[0062] Furthermore, the knowledge graph construction process of the maintenance knowledge base described in S3 includes: Define the ontology model of the knowledge graph, including entity types: equipment, fault type, fault phenomenon, maintenance plan, tool, personnel, and environmental conditions; Define the types of relationships between entities: belonging to, causing, manifesting as, applicable to, requiring, possessing, and limited to; The Neo4j graph database is used to store the knowledge graph, and the Cypher query language is used to implement maintenance scheme reasoning based on graph traversal. The TransE or RotatE knowledge graph embedding algorithm is used to vectorize entities and relations, enabling semantic retrieval based on vector similarity.

[0063] Specifically, this invention constructs a knowledge graph ontology model adapted to power supply equipment operation and maintenance scenarios, clearly defining entity categories such as equipment, fault types, fault phenomena, maintenance plans, tools, personnel, and environmental conditions, and defining the corresponding relationships between entities. This clearly outlines the internal logic of various operation and maintenance elements. The knowledge graph is then stored in the Neo4j graph database, and combined with the Cypher query language, it enables graph traversal-based maintenance plan reasoning. It can deeply mine the connections between different operation and maintenance knowledge based on the relationship links between entities. Simultaneously, this invention uses TransE or RotatE knowledge graph embedding algorithms to vectorize entities and relationships, and relies on vector similarity to complete semantic retrieval, breaking through the limitations of traditional data retrieval that relies solely on keyword matching, making knowledge search more aligned with actual semantic needs. This knowledge graph, combined with standardized coded maintenance solutions, forms a fully functional maintenance knowledge base. It integrates previously scattered maintenance experience, fault handling methods, and resource usage standards into an interconnected semantic network, enabling the systematic storage, accumulation, and inheritance of maintenance knowledge. This effectively improves the problems of knowledge dispersion, difficulty in reuse, and easy loss of senior personnel experience in traditional operation and maintenance models. It also provides solid knowledge support for subsequent fault diagnosis and maintenance solution recommendation. Even in scenarios where the confidence of conventional recommendation algorithms is insufficient, supplementary reasoning and recommendation can be completed with the help of the knowledge graph, further ensuring the reliability of operation and maintenance decisions and the overall efficiency of the process.

[0064] Furthermore, the training and inference process of the FCM+SVM hybrid model described in S4 includes: The historical fault samples were fuzzy clustered using the FCM algorithm to obtain... Each fault cluster and the membership matrix of each sample to each cluster. ; The membership matrix is ​​concatenated with the original eigenvectors to obtain the enhanced eigenvectors. ; An SVM classifier with the RBF kernel function is used to train the enhanced feature vectors to obtain a fault type classification model; For a new fault sample, first calculate its membership degree in each cluster, construct an enhanced feature vector, and then input it into the SVM classifier to output the fault type and the corresponding maintenance requirement level.

[0065] Specifically, this invention utilizes the fuzzy C-means algorithm to perform fuzzy clustering on historical fault samples, obtaining the corresponding fault clusters and the membership matrix of each sample. The membership matrix is ​​then concatenated with the original feature vectors to form an enhanced feature vector, thereby enriching the feature dimensions and fully exploring the implicit distribution characteristics of the fault samples. A support vector machine equipped with radial basis functions and kernel functions is then used to train the model, constructing a reliable fault type classification model. When faced with new fault samples, the system can first calculate the membership of each cluster corresponding to the sample and construct the enhanced feature vector, then input it into the trained model for analysis, ultimately accurately outputting the fault type and the corresponding maintenance requirement level. This processing method, combining fuzzy clustering and support vector machines, can adapt to the fuzziness and uncertainty inherent in power supply equipment faults. By using fuzzy clustering to capture the overall distribution structure of fault samples and then using support vector machines to construct the optimal classification hyperplane in the high-dimensional feature space, the accuracy of fault diagnosis is significantly improved, effectively addressing the problems of inaccurate fault identification and ambiguous maintenance requirement determination in traditional operation and maintenance methods.

[0066] Furthermore, the improved collaborative filtering algorithm described in S5 uses context-adaptive weights. Dynamically calculated based on the current context feature vector:

[0067] In the formula, This is the context feature vector of the current fault scenario. For neighboring users The context feature vector corresponding to the fault scenario, This is the bandwidth parameter of the Gaussian kernel function.

[0068] Specifically, this invention relies on a Gaussian kernel function to dynamically calculate the context-adaptive weights corresponding to neighboring users. It combines the spatial distance between the context feature vector of the current fault scenario and the context feature vectors of past fault scenarios of neighboring users to complete weight allocation. The bandwidth parameter of the Gaussian kernel function can flexibly adjust the influence of feature distance on the weight values, allowing neighboring users with higher similarity to the current maintenance scenario to receive a higher weight ratio, fully leveraging the reference value of similar historical maintenance cases. This dynamic weight setting method changes the limitation of fixed weights in traditional collaborative filtering algorithms. It can closely follow the real-time changes in multi-dimensional contexts such as on-site equipment status, environmental conditions, personnel skills, and resource configuration, allowing the calculation process of maintenance plan recommendation scores to fully incorporate the differentiated features of different scenarios, compensating for the shortcomings of traditional recommendation methods that ignore contextual information and have rigid recommendation logic. The optimized weight mechanism makes the recommendation logic of the improved collaborative filtering algorithm more closely aligned with the actual scenario of power supply equipment fault maintenance, effectively reducing recommendation bias caused by differences in maintenance scenarios, further improving the matching degree between maintenance plans and on-site working conditions, and making the output recommendation results more consistent with the actual operational needs of maintenance personnel, thereby increasing maintenance personnel's willingness to adopt recommended plans.

[0069] Furthermore, the model structure of DQN reinforcement learning described in S6 includes: state space It consists of the context feature vector of the currently faulty device, the fault type code, and the available resource code; Action space : This is a discrete set of all maintenance solutions in the maintenance knowledge base, with one maintenance solution recommended for each action; Q-network: It adopts a fully connected neural network with two hidden layers. The input layer dimension is the state space dimension, the number of hidden layer neurons are 256 and 128 respectively, the activation function is ReLU, and the output layer dimension is the action space dimension. Target network: Same structure as the Q network, parameters are updated using a soft update strategy. ,in This is a soft update coefficient; Experience replay: A priority experience replay mechanism is adopted to store transferred samples. Sampling is performed based on the TD error priority.

[0070] Specifically, this invention relies on a deep Q-network to build a complete reinforcement learning model architecture. The context feature vector of the faulty equipment, the fault type encoding, and the available resource encoding are combined to form the state space, which can comprehensively represent the actual on-site working conditions corresponding to each maintenance decision. Simultaneously, all maintenance solutions in the maintenance knowledge base are set as discrete action spaces, fully covering all possible maintenance methods and ensuring the comprehensiveness of the decision-making scope. The Q-network in the model adopts a fully connected neural network structure with two hidden layers, and with the ReLU activation function, it can efficiently mine the intrinsic correlation between state information and maintenance solutions, improving feature fitting ability. The corresponding target network adopts the same structure as the Q-network and completes parameter iteration through a soft update strategy, effectively alleviating the parameter oscillation problem that occurs during model training, making the overall training process more stable and achieving better convergence. This invention also employs a priority experience replay mechanism to store various transfer samples during the operation and maintenance process, and conducts sample sampling according to TD error priority, enabling the model to prioritize learning historical operation and maintenance data with higher reference value, significantly improving learning efficiency and learning effect. With the aforementioned robust deep Q-network architecture, the system can continuously receive feedback information after the actual execution of maintenance plans, iteratively optimize various parameters of the recommendation model, and build a closed-loop optimization system. This completely improves the problem that traditional operation and maintenance recommendation systems are open-loop architectures that cannot adjust themselves according to actual usage, leading to a gradual degradation of model performance over time. This allows the recommendation strategy to continuously adapt to various actual changes such as equipment aging, changes in operation and maintenance resources, and updates in on-site working conditions, ensuring that the accuracy and rationality of maintenance plan recommendations remain at a good level.

[0071] Furthermore, the dynamic recommendation of maintenance solutions described in S5 also includes: when the confidence level of the collaborative filtering recommendation result is lower than a preset threshold, triggering knowledge graph inference recommendation as a supplement, namely: Retrieve relevant maintenance solution nodes from the knowledge graph based on the current fault type; Perform multi-hop path traversal along the knowledge graph relationship edges to find a repair solution that matches the current equipment type, fault cause, and available tools; By combining historical success rates and estimated maintenance times stored in the knowledge graph, the solutions are sorted to generate a list of recommended alternatives.

[0072] Specifically, when the confidence level of the recommendations output by collaborative filtering does not meet the preset requirements, this invention automatically initiates knowledge graph reasoning to complete supplementary recommendations. First, it retrieves related maintenance solution nodes in the knowledge graph based on the current fault type. Then, it performs multi-hop path traversal along the relationship edges in the knowledge graph to filter out maintenance solutions that match the current equipment type, fault cause, and available tools on-site. This fully leverages the semantic association and graph traversal reasoning capabilities of the knowledge graph, broadening the scope of solution retrieval and matching. Based on this, it ranks the filtered solutions by combining the historical success rate and estimated maintenance time stored in the knowledge graph, ultimately generating a standardized candidate recommendation list. This ensures that the supplementary recommendations simultaneously consider solution reliability and on-site operational efficiency. This dual-mode recommendation approach addresses the limitations of single recommendation algorithms in terms of applicable scenarios and the instability of results. By combining an improved collaborative filtering algorithm with a knowledge graph reasoning mechanism, the two recommendation methods serve as backups for each other, effectively enhancing the overall system's robustness. Even in complex fault scenarios or when the reliability of conventional recommendation algorithm outputs is low, the system can still output reasonable maintenance solutions that are suitable for the on-site working conditions, further improving the overall matching effect of maintenance solutions. It also allows the established maintenance knowledge network to be fully utilized, continuously providing stable and reliable support for on-site operation and maintenance decisions.

[0073] Furthermore, the method also includes anomaly detection and early warning steps: The isolated forest algorithm is used to detect anomalies in the context feature vector extracted in step S2 and identify equipment states that deviate from normal operating mode. When the abnormal score exceeds the preset threshold, an early warning message is automatically generated and pushed to the operation and maintenance management personnel to achieve predictive maintenance.

[0074] Specifically, this invention adds anomaly detection and early warning processing flow, utilizing the isolated forest algorithm to perform anomaly analysis on the previously extracted context feature vectors. This effectively identifies equipment operating states that deviate from normal operating modes. Once the equipment anomaly score exceeds a preset threshold, the system automatically generates corresponding early warning information and pushes it to maintenance personnel, thereby achieving predictive maintenance of power supply equipment. This method relies on complete multi-dimensional context feature mining to uncover hidden anomaly signals during equipment operation, changing the passive situation of traditional maintenance models that can only carry out emergency repairs after a fault occurs. It can detect potential equipment faults in advance, allowing maintenance personnel to intervene and maintain equipment before a fault occurs, preventing minor operational anomalies from gradually evolving into serious equipment failures. This reduces power outage losses and on-site safety risks caused by sudden faults, and also reduces the difficulty and cost of later fault repairs. At the same time, this functional module is integrated with the original processes such as data collection, feature extraction, fault diagnosis, and solution recommendation, further enriching and improving the entire intelligent maintenance system for power supply equipment. This makes equipment status monitoring, hidden danger early warning, fault diagnosis, and maintenance solution recommendation an organic whole, comprehensively improving the initiative and comprehensiveness of power maintenance work, and continuously ensuring the long-term stable operation of various power supply equipment.

[0075] A context-aware intelligent recommendation system for dynamic maintenance solutions of power supply equipment faults, applied to the aforementioned context-aware intelligent recommendation method for dynamic maintenance solutions of power supply equipment faults, includes: The multi-source operation and maintenance data acquisition module is used to collect equipment parameters, operating status, historical maintenance records, environmental parameters and personnel information in real time through sensor networks, SCADA systems and historical operation and maintenance databases, and to perform data preprocessing. The context feature extraction module is used to extract 12-dimensional context feature vectors from six dimensions: equipment attributes, operating environment, historical maintenance, real-time status, personnel skills, and resource availability, and then perform normalization and dimensionality reduction processing. The maintenance knowledge base construction module is used to standardize the coding of maintenance solutions using structured description templates and to establish semantic relationships between fault types, equipment categories, maintenance solutions, and tool resources using knowledge graph technology. The fault diagnosis and maintenance requirement analysis module is used to perform cluster analysis and classification identification of equipment faults using an FCM+SVM hybrid model to determine the fault type and maintenance requirement level. The maintenance plan dynamic recommendation module is used to calculate maintenance plan recommendation scores using an improved collaborative filtering algorithm based on context weights, and generate a Top-K recommendation list based on the scores. The solution feedback and continuous optimization module is used to construct reward signals based on the actual execution feedback of maintenance solutions using the DQN reinforcement learning framework, and to continuously optimize the recommendation model parameters through iterative training; and The knowledge graph reasoning module is used to trigger semantic reasoning recommendations based on knowledge graphs when the confidence of the collaborative filtering recommendation results is lower than a preset threshold, serving as a supplement to collaborative filtering recommendations.

[0076] Specifically, the intelligent recommendation system built by this invention implements the entire fault repair solution recommendation method into a modular entity architecture. Each functional module has a clear division of labor and works in tandem with the others, completely connecting the entire business process from raw data acquisition to continuous model optimization. The multi-source operation and maintenance data acquisition module can stably complete the collection and preprocessing of various types of operation and maintenance data, providing standardized and reliable basic data for all subsequent business processes. The context feature extraction module, based on established rules, completes multi-dimensional feature extraction, normalization, and dimensionality reduction, accurately extracting core information that can characterize the on-site working conditions. The maintenance knowledge base construction module realizes standardized coding of maintenance solutions and the construction of knowledge graphs, transforming scattered maintenance experience into a systematic knowledge resource that can be called upon and reasoned about. The fault diagnosis and maintenance requirements analysis module relies on a mature hybrid model to complete fault identification and requirement determination, accurately locate fault types, and clarify the basic requirements for maintenance work. The maintenance solution dynamic recommendation module uses an improved collaborative filtering algorithm to complete solution scoring and ranking, outputting a recommendation list that fits the actual site conditions. The solution feedback and continuous optimization module combines a deep Q-network reinforcement learning mechanism to form a closed-loop iterative system, allowing the system to continuously optimize its performance based on actual operation and maintenance results. The knowledge graph reasoning module can initiate supplementary recommendations when the credibility of conventional recommendation results is insufficient, further enhancing the system's comprehensive service capabilities. The entire system architecture is logically clear and functionally comprehensive, changing the current situation where traditional operation and maintenance tools have fragmented functions, disconnected data and knowledge, and cannot achieve autonomous iterative upgrades. It effectively solves problems such as the difficulty in reusing maintenance knowledge, the single recommendation mode, and the continuous degradation of model performance over time, comprehensively improving the automation and intelligence of power supply equipment operation and maintenance. At the same time, the system can adapt to various operation and maintenance scenarios such as urban rail transit power supply systems, substations, and distribution networks, and has good versatility and scalability, providing stable, efficient, and intelligent decision support for fault handling of different types of power supply equipment.

[0077] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0078] Example 1: Recommended Repair Solution for Circuit Breaker Faults in Urban Rail Transit Main Substations This embodiment takes the closing failure of the 110kV SPF6 circuit breaker in a main substation of a city rail transit system as an example to illustrate the practical application process of the method described in this invention.

[0079] S1: Multi-source operation and maintenance data collection After a circuit breaker failure occurs, the system automatically collects the following multi-source data: Basic equipment parameters: The equipment type is SF6 circuit breaker, model LW36-126 / T3150-40, rated voltage 126kV, rated current 3150A, rated short-circuit breaking current 40kA, manufacturer is a well-known electrical equipment company, commissioning date is March 2018, and service life is approximately 6 years.

[0080] Operational status data: Within one hour before the fault, the SCADA system recorded abnormal fluctuations in the motor current of the circuit breaker operating mechanism at a frequency of 1Hz (normal 0.8-1.2A, abnormal peak up to 3.5A), the SF6 gas pressure dropped from 0.60MPa to 0.52MPa (rated value 0.60MPa, alarm threshold 0.55MPa, lockout threshold 0.50MPa), the circuit breaker opening and closing coil resistance was normal, and there was a 10ms delay in the auxiliary switch action sequence.

[0081] Historical maintenance records: This circuit breaker has a total of 4 maintenance records in the past 3 years, including 2 SF6 gas replenishments, 1 operating mechanism lubrication and maintenance, and 1 control circuit terminal tightening.

[0082] Environmental parameters: The fault occurred at 14:30 in the summer afternoon, with an ambient temperature of 38°C, an ambient humidity of 72%, and the weather was sunny turning cloudy.

[0083] Maintenance personnel information: There are 3 maintenance personnel on duty, including 1 senior technician (with circuit breaker maintenance qualification and a historical success rate of 95%) and 2 intermediate workers (one of whom has SF6 gas handling qualification). The current workload is medium.

[0084] S2: Contextual Feature Extraction Extract a 12-dimensional contextual feature vector from the above multi-source data:

[0085] After Z-Score standardization, the standard context feature vector is obtained. After dimensionality reduction using PCA, the top 5 principal components (cumulative contribution rate 87.3%) were selected to obtain the dimensionality-reduced feature vector. .

[0086] S3: Repair Knowledge Base Search Using "SF6 circuit breaker + closing failure + low gas pressure" as search criteria, semantic retrieval was performed in the maintenance knowledge graph. Nodes in the knowledge graph matching this fault scenario included: fault type node "SF6 gas leakage", fault symptom node "closing failure + low pressure alarm", and maintenance solution nodes "gas replenishment + leak detection + seal replacement", etc. Through a graph traversal algorithm, the system retrieved three highly relevant historical maintenance paths from the knowledge graph, providing prior knowledge for subsequent recommendations.

[0087] S4: Fault Diagnosis and Repair Needs Analysis The dimensionality-reduced context feature vector is input into the trained FCM+SVM hybrid model. The FCM clustering layer calculates the membership degree of the sample to each fault cluster: 0.72 for the "SF6 gas leak" cluster, 0.21 for the "operating mechanism fault" cluster, and 0.07 for the "control loop fault" cluster. After the enhanced feature vector is input into the SVM classifier, the model outputs the fault type as "SF6 gas leak causing closing lockout," with a confidence level of 0.91 and a maintenance requirement level of "urgent" (requiring completion within 2 hours).

[0088] S5: Dynamic Recommendations for Repair Solutions The system employs an improved collaborative filtering algorithm to calculate the recommended scores for each repair solution. Based on current context features, it calculates context-adaptive weights for neighboring users, with a focus on referencing historical case scores similar to the current fault scenario. The system then generates a Top-3 recommended solution. Option 1 (Recommended score 0.92): SF6 gas replenishment + infrared leak detection + seal replacement. Estimated repair time: 90 minutes. Requires 1 senior technician + 1 intermediate technician. Requires SF6 gas cylinder, infrared leak detector, seal kit, and other tools and materials. Historical success rate: 94%.

[0089] Option 2 (Recommended score 0.85): SF6 gas replenishment + soapy water leak detection. Estimated repair time: 60 minutes. Operation is relatively simple, but leak detection accuracy is lower, with a historical success rate of 88%.

[0090] Option 3 (Recommended score 0.78): Replace the entire sealing assembly of the operating mechanism. Estimated repair time: 180 minutes. While the work is extensive, it is highly thorough and has a historical success rate of 96%.

[0091] The maintenance personnel selected and implemented Solution 1 based on the actual situation. The final maintenance time was 95 minutes, a deviation of 5.6% from the estimated time, and the maintenance was successfully completed. This embodiment achieved a solution matching rate of 93.5% and a maintenance personnel adoption rate of 87%, reducing maintenance time by 35% compared to the traditional manual search method.

[0092] Example 2: Transformer Fault Repair in Urban Rail Traction Substation This example uses a 35kV / 1180V traction transformer in a city rail transit traction substation that has experienced excessive temperature rise as a case study.

[0093] S1: Multi-source operation and maintenance data collection The system collected real-time data from the transformer: top oil temperature 95°C (rated temperature rise 65K, alarm value 85°C, trip value 105°C), winding temperature 105°C, load rate 92%, cooling fan operating normally but airflow approximately 15% lower than design value. Environmental conditions: underground substation, ambient temperature 32°C, ventilation system operating normally. Historical records: This transformer has been in operation for 4 years, with 3 temperature rise alarms in the past year. The last maintenance was a cooling system cleaning 6 months ago.

[0094] S2: Contextual Feature Extraction After extracting 12-dimensional contextual features and normalizing them, the equipment type is a traction transformer, the service life is 4 years, the ambient temperature is 32°C, the ambient humidity is 65%, the historical failure frequency is 3 times / year, the historical maintenance method is cleaning and maintenance, the real-time status is temperature rise alarm, the load rate is 92%, the personnel skill level is senior technician + intermediate worker, the tool list is complete, and the spare parts inventory is sufficient.

[0095] S3: Fault Diagnosis The FCM+SVM hybrid model diagnostic result is "the cooling system's heat dissipation efficiency has decreased, leading to excessive temperature rise", with a confidence level of 0.88 and a repair requirement level of "important" (to be handled within 4 hours).

[0096] S4: Dynamic Recommendations for Repair Solutions The system recommends two repair options: Option 1 (Recommended score 0.89): Comprehensive cleaning of the cooling system + inspection of fan blades + unblocking of air ducts. Estimated repair time: 120 minutes, requiring 2 repair personnel, historical success rate: 92%.

[0097] Option 2 (Recommended score 0.82): Replace the cooling fan and clean the radiator. Estimated repair time: 180 minutes. Higher cost, but more lasting effect. Historical success rate: 96%.

[0098] The maintenance team selected Solution 1 and executed it. The repair took 115 minutes, and the temperature returned to normal. This embodiment achieved a 91.8% solution matching rate and an 85% adoption rate, improving maintenance efficiency by 30% compared to traditional methods.

[0099] Example 3: Cable Fault Repair in Urban Rail Transit Substation This embodiment takes the insulation breakdown fault of a 10kV power cable in a city rail transit substation as an example.

[0100] S1: Multi-source operation and maintenance data collection System data collected: Cable model YJV22-3×240, service life 8 years. One week before the fault, partial discharge detection showed the discharge level increased from 20pC to 150pC (alarm value 100pC). At the time of the fault, the cable's insulation resistance to ground decreased from 500MΩ to 2MΩ. Environmental parameters: Temperature inside the cable tunnel 45°C, humidity 85%. Personnel information: Two technicians with cable termination qualifications were on duty.

[0101] S2: Contextual Feature Extraction Extract 12-dimensional features: equipment type is power cable, operating years are 8 years, ambient temperature is 45°C, ambient humidity is 85%, historical fault frequency is 1 time / year (last 3 years), historical maintenance method is partial repair, current status is insulation breakdown, load rate is 75%, personnel skill level is senior technician, tool list includes special equipment such as cable fault tester, and spare parts need to be temporarily allocated.

[0102] S3: Fault Diagnosis The FCM+SVM hybrid model diagnosis result is "cable main insulation aging and breakdown", with a confidence level of 0.93 and a repair requirement level of "urgent".

[0103] S4: Dynamic Recommendations for Repair Solutions The system recommends two repair options: Option 1 (Recommended score 0.91): Cable fault location + intermediate joint replacement. The fault location is determined using the pulse reflection method. After cutting off the faulty section, two sets of intermediate joints are fabricated. The estimated repair time is 240 minutes. Requires a cable fault tester, joint fabrication materials, etc. Historical success rate is 90%.

[0104] Option 2 (Recommended score 0.86): Replace the entire cable. Highest reliability but high cost and long lead time. Estimated repair time is 480 minutes; new cable needs to be procured in advance.

[0105] The maintenance team selected Solution 1, which accurately located the fault, ensured the connector fabrication was of acceptable quality, and passed the insulation test. This embodiment achieved a 92.5% solution matching rate and an 86% adoption rate, reducing maintenance costs by 25% compared to the complete section replacement solution.

[0106] Example 4: Validation of the continuous optimization effect of the model based on DQN To verify the effectiveness of the reinforcement learning closed-loop optimization described in S6, a backtesting verification was conducted using six months of actual operating data from a certain urban rail power supply system.

[0107] Experimental setup: 326 real maintenance records from January to June 2023 were selected as training data. The DQN model was initialized with data from the first three months, and online optimization was performed using data from the last three months. The state space has 32 dimensions (5-dimensional context features + 12-dimensional fault type encoding + 15-dimensional resource encoding), and the action space consists of 86 solutions from the maintenance knowledge base. The Q-network structure is 32→256→128→86, with ReLU activation function, a learning rate of 0.001, and a discount factor. Soft update coefficient The experience playback pool has a capacity of 10,000 and a batch size of 64.

[0108] Optimization Results: After three months of online optimization, the matching rate of the recommendation model increased from 88.2% to 94.1% (an increase of 5.9 percentage points), and the adoption rate by maintenance personnel increased from 80.5% to 88.3% (an increase of 7.8 percentage points). The average recommendation response time was 1.2 seconds. The cumulative reward function showed a stable upward trend, indicating that the model continuously optimized its recommendation strategy through learning.

[0109] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for intelligently recommending dynamic maintenance solutions for power supply equipment faults based on context awareness, characterized in that, Includes the following steps: S1. Multi-source operation and maintenance data acquisition: Through sensor networks, SCADA systems, historical operation and maintenance databases and enterprise resource planning systems deployed at the power supply equipment site, real-time acquisition of equipment basic parameters, operating status data, historical maintenance records, environmental parameters and maintenance personnel information is carried out. The data update frequency is no less than 1Hz, and the acquired multi-source heterogeneous data is cleaned, denoised, time-series aligned and standardized preprocessed. S2. Context Feature Extraction: A 12-dimensional context feature vector is extracted from six dimensions: equipment attributes, operating environment, historical maintenance, real-time status, personnel skills, and resource availability. These 12-dimensional context features include: equipment type, equipment model, years of operation, ambient temperature, ambient humidity, historical failure frequency, historical maintenance methods, real-time operating status, real-time load rate, maintenance personnel skill level, available tool list, and spare parts inventory status. The 12-dimensional context feature vector is then normalized to obtain a standard context feature vector. ; S3. Maintenance Knowledge Base Construction: Standardize and encode maintenance plans using structured description templates, including fault phenomenon description, fault cause analysis, maintenance operation steps, required tools and materials list, personnel skill requirements, safety precautions, estimated maintenance time and historical success rate. At the same time, knowledge graph technology is used to establish semantic relationships between fault types, equipment categories, maintenance plans, and tool resources to form a reasonable maintenance knowledge network. S4. Fault diagnosis and maintenance requirement analysis: A hybrid model combining FCM clustering and SVM is adopted. First, the FCM algorithm is used to perform fuzzy clustering analysis on equipment fault samples, and then the SVM classifier is used to identify fault types and determine maintenance requirements based on the clustering results. The clustering objective function of the FCM algorithm is: The membership update formula is: In the formula, For clustering loss function, The total number of samples, The number of cluster centers. For the first The nth sample pair Membership degree of each cluster center For fuzzy coefficients and , For the first The feature vector of each sample For the first Cluster center vectors; S5. Dynamic Recommendation of Repair Solutions: An improved collaborative filtering algorithm based on context weights is adopted to comprehensively consider user bias, solution bias, neighboring user ratings and solution similarity, calculate the recommendation score of the repair solution, and generate a Top-K recommendation list based on the score. The formula for calculating the recommendation score is as follows: In the formula, For users Regarding the plan Recommended score, The overall average score. For users Scoring bias, For the plan Scoring bias, For users The set of neighboring users For neighboring users Context-adaptive weights For users Regarding the plan Historical ratings For the plan With the plan The similarity measure is calculated using cosine similarity or Pearson correlation coefficient; S6. Solution Feedback and Continuous Optimization: Based on the DQN reinforcement learning framework, a reward function is constructed according to the actual execution results of the maintenance solution. The recommendation model parameters are continuously optimized through iterative training. The reward function is defined as follows: In the formula, A reward will be given for successful repairs. Penalty for delayed repair time, Penalty for maintenance costs, This is the time delay weighting coefficient. This is the cost weighting coefficient.

2. The intelligent recommendation method for dynamic maintenance schemes of power supply equipment based on context awareness according to claim 1, characterized in that, The multi-source operation and maintenance data collection mentioned in S1 includes: Basic equipment parameters, including equipment type, equipment model, rated voltage, rated current, rated power, manufacturer, commissioning date, and years of operation; Operating status data, including real-time voltage, real-time current, active power, reactive power, power factor, equipment temperature, and switch status, are collected at a frequency of no less than 1Hz. Historical maintenance records include the time of the historical failure, the type of failure, the symptoms of the failure, the maintenance measures, the maintenance time, the maintenance cost, and the maintenance personnel information; Environmental parameters, including ambient temperature, ambient humidity, altitude, air quality index, and weather conditions; Maintenance personnel information includes personnel skill level, professional qualifications, historical maintenance success rate, and current workload.

3. The intelligent recommendation method for dynamic maintenance schemes for power supply equipment faults based on context awareness as described in claim 1, characterized in that, The process of context feature extraction described in S2 is as follows: Feature engineering is performed on the collected multi-source raw data, including feature selection, feature transformation, and feature dimensionality reduction; The 12-dimensional contextual features are normalized using the Z-Score normalization method: In the formula, For the first dimensional original eigenvalues, For the first The sample mean of the dimensional feature. For the first The sample standard deviation of the dimensional feature; Principal component analysis was performed on the normalized 12-dimensional eigenvectors to reduce dimensionality, and the top eigenvectors with a cumulative contribution rate of not less than 85% were selected. The principal components are used to obtain the dimensionality-reduced context feature vector. .

4. The intelligent recommendation method for dynamic maintenance schemes for power supply equipment faults based on context awareness according to claim 1, characterized in that, The knowledge graph construction process of the maintenance knowledge base described in S3 includes: Define the ontology model of the knowledge graph, including entity types: equipment, fault type, fault phenomenon, maintenance plan, tool, personnel, and environmental conditions; Define the types of relationships between entities: belonging to, causing, manifesting as, applicable to, requiring, possessing, and limited to; The Neo4j graph database is used to store the knowledge graph, and the Cypher query language is used to implement maintenance scheme reasoning based on graph traversal. The TransE or RotatE knowledge graph embedding algorithm is used to vectorize entities and relations, enabling semantic retrieval based on vector similarity.

5. The intelligent recommendation method for dynamic maintenance schemes of power supply equipment based on context awareness according to claim 1, characterized in that, The training and inference process of the FCM+SVM hybrid model described in S4 includes: The historical fault samples were fuzzy clustered using the FCM algorithm to obtain... Each fault cluster and the membership matrix of each sample to each cluster. ; The membership matrix is ​​concatenated with the original eigenvectors to obtain the enhanced eigenvectors. ; An SVM classifier with the RBF kernel function is used to train the enhanced feature vectors to obtain a fault type classification model; For a new fault sample, first calculate its membership degree in each cluster, construct an enhanced feature vector, and then input it into the SVM classifier to output the fault type and the corresponding maintenance requirement level.

6. The intelligent recommendation method for dynamic maintenance schemes of power supply equipment based on context awareness according to claim 1, characterized in that, Context-adaptive weights of the improved collaborative filtering algorithm described in S5 Dynamically calculated based on the current context feature vector: In the formula, This is the context feature vector of the current fault scenario. For neighboring users The context feature vector corresponding to the fault scenario, This is the bandwidth parameter of the Gaussian kernel function.

7. The intelligent recommendation method for dynamic maintenance schemes for power supply equipment faults based on context awareness according to claim 1, characterized in that, The model structure of DQN reinforcement learning described in S6 includes: state space It consists of the context feature vector of the currently faulty device, the fault type code, and the available resource code; Action space : This is a discrete set of all maintenance solutions in the maintenance knowledge base, with one maintenance solution recommended for each action; Q-network: It adopts a fully connected neural network with two hidden layers. The input layer dimension is the state space dimension, the number of hidden layer neurons are 256 and 128 respectively, the activation function is ReLU, and the output layer dimension is the action space dimension. Target network: Same structure as the Q network, parameters are updated using a soft update strategy. ,in This is a soft update coefficient; Experience replay: A priority experience replay mechanism is adopted to store transferred samples. Sampling is performed based on the TD error priority.

8. The intelligent recommendation method for dynamic maintenance schemes of power supply equipment based on context awareness according to claim 1, characterized in that, The dynamic recommendation of maintenance solutions described in S5 also includes: when the confidence level of the collaborative filtering recommendation result is lower than a preset threshold, triggering knowledge graph inference recommendation as a supplement, which is: Retrieve relevant maintenance solution nodes from the knowledge graph based on the current fault type; Perform multi-hop path traversal along the knowledge graph relationship edges to find a repair solution that matches the current equipment type, fault cause, and available tools; By combining historical success rates and estimated maintenance times stored in the knowledge graph, the solutions are sorted to generate a list of recommended alternatives.

9. The intelligent recommendation method for dynamic maintenance schemes for power supply equipment faults based on context awareness according to any one of claims 1-8, characterized in that, The method also includes anomaly detection and early warning steps: The isolated forest algorithm is used to detect anomalies in the context feature vector extracted in step S2 and identify equipment states that deviate from normal operating mode. When the abnormal score exceeds the preset threshold, an early warning message is automatically generated and pushed to the operation and maintenance management personnel to achieve predictive maintenance.

10. A context-aware intelligent recommendation system for dynamic maintenance solutions of power supply equipment faults, applied to the context-aware intelligent recommendation method for dynamic maintenance solutions of power supply equipment faults as described in any one of claims 1-9, characterized in that, include: The multi-source operation and maintenance data acquisition module is used to collect equipment parameters, operating status, historical maintenance records, environmental parameters and personnel information in real time through sensor networks, SCADA systems and historical operation and maintenance databases, and to perform data preprocessing. The context feature extraction module is used to extract 12-dimensional context feature vectors from six dimensions: equipment attributes, operating environment, historical maintenance, real-time status, personnel skills, and resource availability, and then perform normalization and dimensionality reduction processing. The maintenance knowledge base construction module is used to standardize the coding of maintenance solutions using structured description templates and to establish semantic relationships between fault types, equipment categories, maintenance solutions, and tool resources using knowledge graph technology. The fault diagnosis and maintenance requirement analysis module is used to perform cluster analysis and classification identification of equipment faults using an FCM+SVM hybrid model to determine the fault type and maintenance requirement level. The maintenance plan dynamic recommendation module is used to calculate maintenance plan recommendation scores using an improved collaborative filtering algorithm based on context weights, and generate a Top-K recommendation list based on the scores. The solution feedback and continuous optimization module is used to construct reward signals based on the actual execution feedback of maintenance solutions using the DQN reinforcement learning framework, and continuously optimize the recommendation model parameters through iterative training. as well as The knowledge graph reasoning module is used to trigger semantic reasoning recommendations based on knowledge graphs when the confidence of the collaborative filtering recommendation results is lower than a preset threshold, serving as a supplement to collaborative filtering recommendations.

Images