A fault maintenance framework construction method and system based on fault big data and equipment BOM

Abstract

The application discloses a fault maintenance framework construction method and system based on combination of fault big data and equipment BOM, and relates to the field of intelligent operation and maintenance of a power system. The method solves the problems of insufficient utilization of historical data, insufficient utilization of equipment BOM information, low fault prediction accuracy, poor maintenance strategy precision, low resource scheduling efficiency and lack of closed-loop feedback optimization mechanism in the existing fault maintenance technology, and comprises the following steps: distributed nodes collect multi-dimensional fault data and pre-process the multi-dimensional fault data; a component hierarchical tree BOM knowledge graph is built according to equipment data; an improved FP-Growth algorithm is used to mine frequent item sets and association rules between faults and components; an improved LSTM (Long Short-Term Memory) fusing BOM features is used for fault time series prediction, and a maintenance strategy is generated by matching a knowledge base; an improved genetic algorithm is used to optimize multi-objective scheduling of personnel, spare parts and tools; and based on deep Q network reinforcement learning, each module parameter is iteratively optimized in combination with maintenance feedback.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance of power systems, specifically to a method and system for constructing a fault repair framework based on a combination of fault big data and equipment BOM. Background Technology

[0002] With the rapid development of urban rail transit and smart grids, the scale of power equipment is becoming increasingly large and the types of equipment are becoming more complex. Traditional manual inspection and passive maintenance methods can no longer meet the dual requirements of modern power systems for high reliability and low operation and maintenance costs. At present, some data-driven intelligent technologies have emerged in the field of power equipment fault maintenance, mainly including fault diagnosis methods based on statistical analysis, fault prediction methods based on machine learning, and maintenance decision-making methods based on expert systems.

[0003] Although existing technologies have improved the level of intelligence in the operation and maintenance of power equipment to some extent, the following technical shortcomings still exist: (1) Insufficient use of historical data: Existing methods usually only use a small amount of recent fault data for simple statistical analysis, failing to fully explore the deep-seated patterns contained in massive historical fault data. The utilization rate of historical data is usually less than 30%, resulting in a low fault prediction accuracy (generally not exceeding 80%).

[0004] (2) Equipment BOM information is not fully utilized: Existing fault repair methods generally ignore the valuable information contained in the equipment BOM, such as the hierarchical relationship of components, component attributes and related relationships, and fail to effectively integrate equipment structure knowledge with data-driven methods, resulting in inaccurate fault location and a lack of targeted repair strategies.

[0005] (3) Low accuracy of fault prediction: Traditional prediction models (such as ARIMA, support vector machine, etc.) are difficult to effectively capture the long-term dependencies and nonlinear characteristics in fault time series data. The accuracy of fault prediction for complex power equipment is generally low (70%~80%), with high false alarm rate and false alarm rate.

[0006] (4) Poor accuracy of maintenance strategies: Existing maintenance strategies rely heavily on the personal experience of maintenance personnel, lack data support and system optimization, and the strategy matching rate is generally no more than 70%, often resulting in over-maintenance or under-maintenance.

[0007] (5) Low resource scheduling efficiency: The allocation of maintenance resources (personnel, spare parts, tools) is mostly done by manual experience without considering multi-objective optimization constraints. The resource utilization rate is generally no more than 60%, the maintenance waiting time is long, and the operation and maintenance cost is high.

[0008] (6) Lack of closed-loop feedback optimization mechanism: Most existing systems are open-loop architectures, and the maintenance execution results are not effectively fed back to the front-end model for continuous optimization. The model performance gradually degrades over time, and the adaptation period for new equipment types is long (usually no less than 2 weeks).

[0009] In summary, existing fault repair technologies suffer from several technical problems, including insufficient utilization of historical data, underutilization of equipment BOM information, low accuracy of fault prediction, poor precision of maintenance strategies, low efficiency of resource scheduling, and a lack of closed-loop feedback optimization mechanisms. Summary of the Invention

[0010] This invention addresses the technical problems in existing fault repair technologies, such as insufficient utilization of historical data, inadequate use of equipment BOM information, low fault prediction accuracy, poor precision of maintenance strategies, low resource scheduling efficiency, and lack of closed-loop feedback optimization mechanisms. Therefore, this invention proposes a fault repair framework construction method and system based on the combination of fault big data and equipment BOM. The aim is to construct an intelligent fault repair system covering the entire chain of "data acquisition—knowledge modeling—correlation analysis—fault prediction—resource scheduling—closed-loop optimization" by integrating multi-source fault big data and equipment BOM structural knowledge, thereby achieving accurate prediction of power equipment faults, automatic generation of maintenance strategies, and optimized scheduling of maintenance resources. This invention achieves the above technical problems through the following technical solutions: Option 1: This invention proposes a method for constructing a fault repair framework based on a combination of fault big data and equipment BOM. The method includes the following steps: Step 1: Construct a multi-dimensional fault dataset; perform data cleaning, format standardization, missing value handling, and outlier detection and removal on the collected multi-dimensional fault dataset; Step 2: Construct a hierarchical relationship model of equipment components, using a tree structure to represent the subordinate relationships between equipment, subsystems, components, and parts, forming a BOM knowledge graph; Step 3: Use the improved FP-Growth algorithm to perform itemset mining on the multi-dimensional fault dataset to obtain the association rules between fault modes and related components; Step 4: Use an improved long short-term memory network that integrates BOM features for fault prediction, and construct a prediction function for an improved LSTM. Step 5: Construct the fitness function of the improved genetic algorithm, and use the improved genetic algorithm to perform multi-objective optimization scheduling of personnel, spare parts and tool resources required for maintenance; Step 6: Construct a closed-loop feedback mechanism based on the deep Q-network reinforcement learning framework, calculate the reward function according to the actual results of maintenance execution, and continuously update the parameters of the correlation analysis model, fault prediction model and resource scheduling model through iterative optimization to complete the construction of a fault maintenance model based on the combination of fault big data and equipment BOM.

[0011] Furthermore, a preferred embodiment is provided, wherein the BOM knowledge graph described in step 2 has the following layers: the first layer is the overall equipment layer, the second layer is the functional subsystem layer, the third layer is the component layer, the fourth layer is the part module layer, and the fifth layer is the component layer; each node includes part number, name, specifications, supplier information, expected lifespan, inventory quantity, and replacement cycle attributes; a graph database (such as Neo4j) is used to store the BOM knowledge graph, supporting relationship-based path queries and subgraph retrieval, enabling fault impact range analysis and component association traceability.

[0012] Furthermore, a preferred embodiment is provided, wherein the method for obtaining the association rules between fault modes and related components by using the improved FP-Growth algorithm to perform itemset mining on the multi-dimensional fault dataset in step 3 is as follows: A distributed FP-Growth algorithm based on MapReduce is used to process large-scale fault datasets, and the frequent itemset mining process is accelerated by constructing conditional FP-trees in parallel; a minimum support threshold is set. and minimum confidence threshold The frequently mined itemsets are sorted and filtered according to their lift, and the lift calculation formula is as follows: Only retain The positive correlation rules.

[0013] Furthermore, a preferred embodiment is provided, wherein the improved long short-term memory network described in step 4 has an input layer receiving length of... Time series historical fault data ,in It includes equipment operating status vector, environmental feature vector, and time encoding vector; the LSTM hidden layer adopts a bidirectional LSTM (Bi-LSTM) structure, with 128 hidden units, 2 layers, and a dropout rate of 0.2; the BOM feature encoding layer adopts a 3-layer MLP with 64, 32, and 16 hidden units respectively, and the activation function is ReLU; the output layer outputs the probability distribution of fault occurrence and the predicted fault type through a fully connected layer and a sigmoid activation function.

[0014] Furthermore, a preferred embodiment is provided, wherein the prediction function of the improved LSTM constructed in step 4 is:

[0015] In the formula, This is the vector of fault prediction results. It is the Sigmoid activation function. This is the output layer weight matrix. This is the output layer bias vector. Historical fault data Temporal feature vector encoded by LSTM network, BOM characteristics The structural feature vector encoded by a multilayer perceptron (MLP), This indicates the feature vector concatenation operation.

[0016] Furthermore, a preferred embodiment is provided, wherein the improved genetic algorithm described in step 5 is: The maintenance resource scheduling scheme is encoded using real-number encoding, with each chromosome containing a maintenance personnel allocation sequence, a spare parts call sequence, and a tool configuration sequence. The selection operation adopts the tournament selection method, the crossover operation adopts the partial mapping crossover operator, and the mutation operation adopts the Gaussian mutation operator. The population size is set to 100, the crossover probability is 0.85, the mutation probability is 0.05, and the maximum number of iterations is 200. An elite retention strategy is introduced in the evolution process, where the 10% of individuals with the highest fitness are retained in each generation and directly enter the next generation.

[0017] Furthermore, a preferred embodiment is provided, in which step 6 a deep Q-network is trained by minimizing the TD error, i.e. .

[0018] Option 2: A fault repair framework construction system based on the combination of fault big data and equipment BOM, the system comprising: The multi-source fault big data acquisition module is used to construct a multi-dimensional fault big data dataset; it performs data cleaning, format standardization, missing value handling, and outlier detection and removal on the acquired multi-dimensional fault big data dataset. The equipment BOM data modeling module is used to build a hierarchical relationship model of equipment components. It uses a tree structure to represent the subordinate relationships between equipment, subsystems, components, and parts, forming a BOM knowledge graph. The fault-component association analysis module is used to perform itemset mining on the multi-dimensional fault dataset using the improved FP-Growth algorithm to obtain the association rules between fault modes and associated components. The fault prediction and maintenance strategy generation module is used to perform fault prediction using an improved long short-term memory network that incorporates BOM features, and to construct a prediction function for an improved LSTM. The maintenance resource optimization and scheduling module is used to construct the fitness function of the improved genetic algorithm and to perform multi-objective optimization scheduling of personnel, spare parts and tools required for maintenance using the improved genetic algorithm. The closed-loop feedback and continuous optimization module is used to build a closed-loop feedback mechanism based on the deep Q-network reinforcement learning framework. It calculates the reward function based on the actual results of maintenance execution, and continuously updates the parameters of the correlation analysis model, fault prediction model and resource scheduling model through iterative optimization, so as to complete the construction of a fault maintenance model based on the combination of fault big data and equipment BOM.

[0019] Option 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in Option 1.

[0020] Option 4: A computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the method described in Option 1.

[0021] The advantages of this invention are: The fault maintenance framework construction method based on the combination of fault big data and equipment BOM described in this invention can efficiently utilize more than 100,000 historical fault data for in-depth mining and model training through distributed multi-source data acquisition and MapReduce parallel processing framework, with a data utilization rate of over 95%, providing sufficient data support for fault prediction and maintenance decision-making.

[0022] The present invention describes a fault repair framework construction method based on the combination of fault big data and equipment BOM. This method integrates equipment BOM structural knowledge into the entire fault repair process. Through BOM knowledge graph construction, BOM feature fusion coding and component association traceability, it achieves effective integration of equipment structural knowledge and data-driven methods, significantly improving the accuracy of fault location and the pertinence of repair strategies.

[0023] The present invention describes a fault repair framework construction method based on the combination of fault big data and equipment BOM. The improved LSTM network that integrates BOM features effectively captures the fault evolution law and the influence of equipment structure by simultaneously encoding historical time series features and equipment structural features. The fault prediction accuracy reaches more than 92%, which is about 15 percentage points higher than the traditional method.

[0024] The fault repair framework construction method described in this invention, which combines fault big data with equipment BOM, automatically generates accurate repair strategy solutions based on data-driven correlation analysis and knowledge base matching. The adoption rate of repair strategies reaches over 80%, effectively avoiding the problems of over-repair and under-repair.

[0025] The fault maintenance framework construction method based on the combination of fault big data and equipment BOM described in this invention uses an improved genetic algorithm for multi-objective optimization scheduling. It comprehensively considers multiple optimization objectives such as cost, time and reliability, and the maintenance resource utilization rate reaches more than 90%, which is about 30 percentage points higher than the traditional manual scheduling method.

[0026] The fault repair framework construction method based on the combination of fault big data and equipment BOM described in this invention has built a complete closed-loop feedback mechanism based on the DQN reinforcement learning framework. The system performance is continuously optimized with the usage time, the fault prediction accuracy increases by no less than 1 percentage point per week, and the adaptation cycle of new equipment BOM is shortened to less than 3 days.

[0027] The fault repair framework construction method based on the combination of fault big data and equipment BOM described in this invention adopts a modular architecture design, decouples the BOM knowledge graph from the algorithm model, can quickly adapt to different types of power equipment, and has good versatility and scalability. Attached Figure Description

[0028] Figure 1 This is an overall architecture diagram of the general framework for fault repair based on the combination of fault big data and equipment BOM as described in Implementation Method 1.

[0029] Figure 2 This is a flowchart of the multi-source fault big data collection and preprocessing described in Implementation Method 1.

[0030] Figure 3 This is a schematic diagram of the five-layer tree structure model of the equipment BOM as described in Implementation Method 1.

[0031] Figure 4 This is a flowchart of the fault-component correlation analysis performed by the improved FP-Growth algorithm described in Implementation Method 1.

[0032] Figure 5 This is a diagram of the improved LSTM network structure that incorporates BOM features as described in Implementation Method 1.

[0033] Figure 6 This is a diagram of the closed-loop feedback optimization framework based on DQN reinforcement learning as described in Implementation Method 1. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0035] Implementation Method 1, see [link] Figures 1 to 6 This embodiment describes a method for constructing a fault repair framework based on a combination of fault big data and equipment BOM. The method includes the following steps: Step 1: Multi-source fault big data collection and preprocessing By deploying distributed acquisition nodes at multiple power equipment sites, fault records, maintenance records, operational data, environmental data, and component replacement records are collected from SCADA (Supervisory Control and Data Acquisition) monitoring systems, EMS (Energy Management System) energy management systems, online condition monitoring devices, and CMMS (Computerized Maintenance Management System) computerized maintenance management systems, thus constructing a multi-dimensional fault dataset covering the entire lifecycle of the equipment.

[0036] The fault records include: equipment number, fault code, fault type, fault description, fault occurrence time, fault severity, and affected area; the maintenance records include: maintenance work order number, maintenance start / end time, maintenance measures, maintenance personnel, and maintenance results; the operating data includes: load current, voltage, power, temperature, vibration frequency, and insulation resistance; the environmental data includes: ambient temperature, humidity, altitude, and pollution level; the parts replacement records include: replaced parts number, name, specifications, quantity, reason for replacement, and supplier information.

[0037] Comprehensive data quality preprocessing was performed on the collected raw data, including: data cleaning (removing duplicate records, correcting erroneous data, and filling missing values), format standardization (unifying data from different sources into standard field formats), and outlier detection and removal (using 3D-based methods). (A hybrid detection strategy combining statistical methods of the criteria and the isolated forest algorithm). After preprocessing, the effective data volume is no less than 100,000 records.

[0038] Step 2: Equipment BOM Data Modeling Based on equipment design drawings, technical manuals, parts lists, and equipment ledger data, a hierarchical relationship model of equipment parts is constructed. A tree structure is used to represent the subordinate relationships between equipment, subsystems, components, and parts, forming a structured BOM knowledge graph.

[0039] The BOM tree structure consists of five layers: the first layer is the overall equipment layer (such as main transformers, circuit breakers, and cable lines); the second layer is the functional subsystem layer (such as cooling systems, insulation systems, and operating systems); the third layer is the component assembly layer (such as radiators, bushings, and contact assemblies); the fourth layer is the part module layer (such as sealing rings, springs, and contacts); and the fifth layer is the component layer (such as resistors, capacitors, and sensors). The BOM tree depth does not exceed 5 layers, and the number of component nodes is no less than 100.

[0040] Each BOM node contains the following attributes: unique part number, part name, specifications, equipment type, supplier information, expected design life, current inventory quantity, recommended replacement cycle, historical failure count, and mean time to repair. A graph database (such as Neo4j) is used to store the BOM knowledge graph. The graph's nodes and relationships visually represent the assembly relationships, functional dependencies, and fault propagation paths between parts. It supports relationship-based path queries and subgraph retrieval, facilitating fault impact analysis and part association tracing.

[0041] Step 3: Fault-Component Correlation Analysis An improved FP-Growth (Frequent PatternGrowth) algorithm is used to perform frequent itemset mining on the multi-dimensional fault dataset, automatically discovering the potential correlation between fault modes and related components from massive historical fault records.

[0042] The improved FP-Growth algorithm employs a MapReduce-based distributed computing framework. It accelerates the frequent itemset mining process by constructing conditional FP-trees in parallel, enabling efficient processing of large-scale fault datasets exceeding 100,000 itemsets. The algorithm sets a minimum support threshold. and minimum confidence threshold Only frequent itemsets and association rules that meet the threshold are retained.

[0043] The confidence score of the association rule is calculated using the following formula:

[0044] In the formula, Indicates the fault mode Related components The confidence level of the association rules between them. Indicates the fault mode and components Simultaneous support, Indicates the fault mode Support To include fault modes and parts The number of transactions, For including fault modes The total number of transactions. The closer the confidence value is to 1, the stronger the failure mode. When the parts appear The greater the likelihood of a malfunction or the need for replacement.

[0045] Furthermore, the mined association rules are sorted and filtered according to their lift, calculated using the following formula:

[0046] Only keep We identify positive correlation rules and eliminate invalid and negative correlation rules to ensure the validity and reliability of the correlation analysis results.

[0047] Step 4: Fault Prediction and Maintenance Strategy Generation An improved Long Short-Term Memory (LSTM) network incorporating BOM (Bill of Materials) features is employed for fault timing prediction. Traditional LSTM networks rely solely on historical time-series data for prediction, making it difficult to effectively utilize device structural knowledge. The improved LSTM proposed in this invention significantly enhances fault prediction accuracy by simultaneously fusing historical fault timing features and BOM structural features through a parallel encoder architecture.

[0048] The prediction function of the improved LSTM is:

[0049] In the formula, Let be the vector of fault prediction results, where Number of fault types; The Sigmoid activation function maps the output value to... The interval represents the probability of occurrence for each type of fault. This is the output layer weight matrix; This is the output layer bias vector; Historical fault data The temporal feature vector encoded by a bidirectional LSTM network, where This represents the number of hidden units in the LSTM. BOM characteristics The structural feature vector encoded by a multilayer perceptron (MLP), where Output dimension for MLP; This represents the concatenation operation of feature vectors along the feature dimension.

[0050] Based on the fault prediction results, combined with the maintenance knowledge base and the association rules obtained in step 3, a maintenance strategy plan is automatically generated, including the following: - Maintenance priority ranking: The urgency level is determined based on the fault prediction probability value, and high-probability faults are handled first; - Fault root cause analysis: Combining the BOM knowledge graph and association rules, the root cause of the fault and possible impact paths are traced; - Recommended maintenance process: The optimal maintenance process route and operation steps are matched from the maintenance knowledge base; - Required spare parts list: Based on the confidence level of association rules and BOM node attributes, a list of parts to be replaced and their quantities are automatically generated; - Estimated maintenance time: Estimated maintenance time based on historical maintenance data statistics and the current fault complexity.

[0051] Step 5: Optimize and schedule maintenance resources An improved genetic algorithm is used to perform multi-objective optimization scheduling of personnel, spare parts and tools required for maintenance. Under the conditions of meeting maintenance time window, personnel skill constraints and spare parts inventory constraints, a resource scheduling scheme that maximizes the overall benefits is generated.

[0052] The fitness function of the improved genetic algorithm is:

[0053] In the formula, The fitness function value indicates that a larger value indicates a better scheduling scheme. The total maintenance cost includes labor costs, spare parts costs, and downtime loss costs. To the maximum permissible repair cost; This refers to the total maintenance time, including preparation time, maintenance operation time, and recovery time. This is the maximum permissible repair time; To determine maintenance reliability indicators, factors such as the skill matching of maintenance personnel, the quality level of spare parts, and the completeness of tools are comprehensively considered. , , Let be the weighting coefficient, satisfying and Adjustments can be made flexibly based on actual operation and maintenance strategies (e.g., increasing the amount of fuel when cost control is emphasized). Emphasizing timeliness increases ).

[0054] Step 6: Closed-loop feedback and continuous optimization A closed-loop feedback mechanism is constructed based on the reinforcement learning framework of Deep Q-Network (DQN). The actual results of maintenance execution (whether the maintenance is successful, the actual maintenance cost, the actual maintenance time, etc.) are used as environmental feedback signals, and the model parameters of each module are optimized through continuous learning.

[0055] The reward function is defined as follows:

[0056] In the formula, A reward is given for successful repairs (value is 1 for successful repairs and 0 for failures). This is a cost penalty item (the ratio of actual cost to budgeted cost). This is a time penalty (the ratio of actual repair time to planned repair time). and This is a balancing coefficient used to adjust the weighting relationship between the various reward components.

[0057] By storing historical state-action-reward transition samples through an experience replay mechanism, and iteratively updating the Q-network parameters using a mini-batch gradient descent method, continuous collaborative optimization of the correlation analysis model, fault prediction model, and resource scheduling model is achieved, enabling the system performance to gradually improve with increasing usage time.

[0058] Example 1: Fault Repair of a 110kV Transformer in a Municipal Rail Transit Substation A main substation for a certain urban rail transit system has two 110kV / 35kV main transformers, each with a capacity of 63MVA, responsible for traction power supply and power and lighting power supply for the line. These transformers have been in operation for eight years, with a total of 156 historical fault records. This embodiment applies the method described in this invention to perform intelligent fault maintenance management on these transformers.

[0059] Step 1: Multi-source fault big data collection and preprocessing The following data were collected from the SCADA system, transformer online monitoring devices (including online monitoring of dissolved gas in oil, online monitoring of winding temperature, and online monitoring of partial discharge), and CMMS maintenance management system of the main substation: Fault Records: A total of 156 historical fault records were recorded, including 42 instances of oil chromatography abnormalities, 18 instances of bushing oil leakage, 35 instances of cooler failures, 28 instances of on-load tap changer abnormalities, and 33 instances of winding temperature rise abnormalities. Maintenance records: A total of 203 maintenance work orders were recorded, including 128 minor repairs, 2 major repairs, and 73 temporary maintenance repairs. Operating data: load current, three-phase voltage, active / reactive power, top oil temperature, winding temperature, dissolved gas concentration in oil (H2, CH4, C2H6, C2H4, C2H2, CO, CO2), core grounding current, sampling interval is 15 minutes, and the cumulative data volume is approximately 2.8 million records; Environmental data: ambient temperature and humidity, with a sampling interval of 1 hour; Parts replacement records: There are a total of 67 historical replacement records, including 12 replacements of gaskets, 8 replacements of cooling fan motors, and 15 replacements of on-load tap changer contacts.

[0060] The raw data was preprocessed as follows: data cleaning removed 3 duplicate records and 5 records with incorrect formatting; linear interpolation was used to fill missing values ​​in the operational data (missing value rate approximately 0.8%); and the Isolation Forest algorithm was used to detect and remove approximately 1200 abnormal operational data points. The effective data volume after preprocessing was: 153 fault records, approximately 2.79 million operational data records, 203 maintenance records, and 67 parts replacement records.

[0061] Step 2: Equipment BOM Data Modeling Based on the design drawings and technical manual of the 110kV main transformer, a five-layer BOM tree structure model was constructed: First layer (overall equipment layer): 110kV / 35kV main transformer (1 unit) The second layer (functional subsystem layer) includes: oil tank and body, cooling system, insulation system, bushing system, on-load tap changer system, and protection system (6 subsystems). The third layer (component layer): winding assembly, core assembly, radiator assembly, oil pump assembly, oil conservator assembly, high-voltage bushing assembly, low-voltage bushing assembly, voltage regulating switch assembly, etc. (24 components) Fourth layer (component module layer): sealing gaskets, cooling fans, oil pump motors, breather silicone, contact assemblies, spring mechanisms, etc. (78 component modules) Fifth layer (component layer): Seals, bearings, relays, sensors, resistors, capacitors, etc. (156 components) The BOM tree has a depth of 5 levels and a total of 265 component nodes. The Neo4j graph database is used to store the BOM knowledge graph, and various relationship types such as "belongs to", "depends on", "replaces", and "fault propagation" are established.

[0062] Step 3: Fault-Component Correlation Analysis A distributed FP-Growth algorithm based on MapReduce was used to perform correlation analysis on 153 historical fault records and 67 component replacement records. Algorithm parameters were set as follows: minimum support. minimum confidence .

[0063] A total of 45 frequent itemsets were obtained, with 32 valid association rules. Some typical association rules are as follows:

[0064] Taking the first rule as an example, the confidence level calculation process is as follows: the number of times oil chromatographic anomalies and gasket aging occurred simultaneously was 8, and the total number of times oil chromatographic anomalies occurred was 42. Therefore... That is, when an oil chromatography abnormality occurs, the probability of the sealing gasket aging is 78.4%.

[0065] Step 4: Fault Prediction and Maintenance Strategy Generation An improved LSTM network incorporating BOM features is used for fault prediction. The model parameters are set as follows: Input timing length (Corresponding to 7 days, with 24 sampling points per day); Input feature dimensions: load current, three-phase voltage, top oil temperature, winding temperature, H2 concentration, and total hydrocarbon concentration, totaling 6 dimensions; Bidirectional LSTM layer: 2 layers, 128 hidden units per direction, dropout rate 0.2; BOM Feature Encoding MLP: Input dimension 265 (corresponding to the number of BOM nodes), hidden layer [64, 32, 16], activation function ReLU; Output layer: Fault type classification (8 common faults), activation function Sigmoid.

[0066] The Adam optimizer was used for training with an initial learning rate of 0.001, a batch size of 32, 100 training epochs, and an early stopping patience of 10. The fault prediction accuracy on the test set reached 93.5%.

[0067] Based on the prediction results and association rules, the maintenance strategy is automatically generated: When the system predicts that the probability of "oil chromatography abnormality" is 0.85, it automatically associates it with the "sealant aging" rule (confidence level 0.784) and generates the maintenance strategy as follows: priority - emergency, recommended process - take oil sample for testing and confirm + check sealant status, required spare parts - sealant (2 sets), estimated time - 4 hours.

[0068] Step 5: Optimize and schedule maintenance resources For the aforementioned transformer oil chromatographic anomaly repair task, an improved genetic algorithm is used for resource scheduling. Algorithm parameters are set as follows: population size 100, crossover probability 0.85, mutation probability 0.05, maximum number of iterations 200, and weight coefficients. , , .

[0069] The optimized scheduling plan included: assigning one senior technician and one technician; utilizing two sets of spare sealing gaskets (with sufficient inventory); and providing one set of oil sample testing tools and one set of seal replacement tools. The estimated repair cost was 3200 yuan (lower than the budget of 4000 yuan), and the estimated repair time was 3.5 hours (lower than the 4-hour time window). In actual implementation, the repair resource utilization rate reached 91%.

[0070] Step 6: Closed-loop feedback and continuous optimization After the maintenance is completed, the system collects the actual execution data: Maintenance successful ( The actual cost was 3400 yuan. The actual repair time was 3 hours. Calculate the reward value. The empirical samples were stored in the experience pool, and the DQN network parameters were updated every 50 samples. After three months of continuous operation, the accuracy of transformer fault prediction improved from the initial 93.5% to 96.2%.

[0071] Example 2: Fault Repair of a 35kV Circuit Breaker in a City Rail Traction Substation A traction substation for urban rail transit is equipped with 12 35kV vacuum circuit breakers, model ZN72-40.5, which are responsible for the protection and control functions of the traction power supply system. This embodiment applies the method described in this invention to perform intelligent fault maintenance management of the circuit breakers.

[0072] Step 1: Collect 89 historical fault records, approximately 1.5 million operational data entries, 156 maintenance records, and 43 component replacement records from the SCADA system, the circuit breaker mechanical characteristic online monitoring device, and the CMMS system.

[0073] Step 2: Construct a five-layer BOM tree structure model. The BOM tree has a depth of 5 layers and a total of 178 component nodes, including key components such as operating mechanisms, vacuum interrupters, insulating rods, and contact springs.

[0074] Step 3: Using the improved FP-Growth algorithm, 28 frequent itemsets and 21 effective association rules were obtained. Typical rules include: "Abnormal mechanical characteristics". "Contact spring fatigue" (confidence level 0.731, lift 2.156), "Closing coil burnout" "Core jamming" (confidence level 0.692, lift 1.987).

[0075] Step 4: Fault prediction is performed using an improved LSTM that incorporates BOM features. Input features include opening and closing times, stroke curves, coil current waveforms, contact wear, etc., and input timing length. (Corresponding to 3 days). Model training parameters: 128 bidirectional LSTM hidden units, 2 layers, BOM-encoded MLP hidden layers [64, 32, 16], 80 training epochs. The fault prediction accuracy on the test set reached 92.8%.

[0076] Step 5: For circuit breaker closing anomaly faults, an improved genetic algorithm is used for resource scheduling, with weight coefficients... , , The optimized scheduling plan involves assigning two professional circuit breaker maintenance technicians, one closing coil, one bottle of lubricating oil, and one mechanical characteristic tester. The actual maintenance time is 2.5 hours, approximately 30% shorter than the traditional method.

[0077] Step 6: After two months of closed-loop feedback optimization, the circuit breaker fault prediction accuracy increased from 92.8% to 94.5%, and the maintenance strategy adoption rate increased from the initial 81% to 86%.

[0078] Example 3: Repair of a 10kV power cable fault in a certain urban rail substation A certain urban rail transit substation has multiple 10kV cross-linked polyethylene (XLPE) power cables, with a total length of approximately 15 kilometers, which are responsible for supplying power to the station's lighting and equipment. This embodiment applies the method described in this invention to perform intelligent fault maintenance management of the cable lines.

[0079] Step 1: Collect 67 historical fault records, approximately 800,000 operational data entries, 112 maintenance records, and 28 component replacement records (cable accessory replacement records) from the cable online monitoring system (including distributed fiber optic temperature measurement, partial discharge monitoring, and sheath circulation current monitoring) and the CMMS system.

[0080] Step 2: Construct a five-layer BOM tree structure model with a depth of 4 layers (the cable structure is relatively simple). The total number of component nodes is 95, including the cable body, intermediate joints, terminal heads, grounding devices, sheath protectors, etc.

[0081] Step 3: Using the improved FP-Growth algorithm, 18 frequent itemsets and 14 effective association rules were obtained. Typical rules include: "Partial discharge anomaly". Intermediate joint insulation degradation (confidence level 0.756, lift 2.324), "abnormal sheath circulation current" "The grounding box connection is loose" (confidence level 0.684, lift 1.876).

[0082] Step 4: Fault prediction is performed using an improved LSTM that incorporates BOM features. Input features include conductor temperature, sheath temperature, partial discharge quantity, sheath circulating current, dielectric loss factor, and input timing length. Model training parameters: 128 bidirectional LSTM hidden units, 2 layers, BOM-encoded MLP hidden layers [64, 32], 60 training epochs. Fault prediction accuracy reached 91.5% on the test set.

[0083] Step 5: For overheating faults at cable joints, an improved genetic algorithm is used for resource scheduling, with weight coefficients... , , The optimized scheduling plan involves assigning one senior cable technician and two technicians, utilizing one set of intermediate joint accessories, and configuring one cable fault tester and one partial discharge detector. The actual repair cost is 8,500 yuan, a reduction of approximately 25% compared to the traditional method.

[0084] Step 6: After 1.5 months of closed-loop feedback optimization, the accuracy of cable fault prediction increased from 91.5% to 93.8%, and the adoption rate of maintenance strategies remained above 80%. The results of the implementation examples are summarized below:

[0085] In summary, the verification results of the three embodiments show that the general fault repair framework based on the combination of fault big data and equipment BOM proposed in this invention has significant technical effects and practical value in the field of intelligent operation and maintenance of power systems. The fault prediction accuracy rate reaches 91.5%-93.5%, and the maintenance strategy adoption rate reaches 80%-82%, effectively improving the intelligence level and operation and maintenance efficiency of power equipment fault repair.

[0086] Those skilled in the art will understand that the above description is merely a preferred embodiment of the present invention, and the features described in the various embodiments and / or claims of this disclosure can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. This is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0087] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for constructing a fault maintenance framework based on the combination of fault big data and equipment BOM, characterized in that, The method includes the following steps: Step 1: Construct a multi-dimensional fault dataset; perform data cleaning, format standardization, missing value handling, and outlier detection and removal on the collected multi-dimensional fault dataset; Step 2: Construct a hierarchical relationship model of equipment components, using a tree structure to represent the subordinate relationships between equipment, subsystems, components, and parts, forming a BOM knowledge graph; Step 3: Use the improved FP-Growth algorithm to perform itemset mining on the multi-dimensional fault dataset to obtain the association rules between fault modes and related components; Step 4: Use an improved long short-term memory network that integrates BOM features for fault prediction, and construct a prediction function for an improved LSTM. Step 5: Construct the fitness function of the improved genetic algorithm, and use the improved genetic algorithm to perform multi-objective optimization scheduling of personnel, spare parts and tool resources required for maintenance; Step 6: Construct a closed-loop feedback mechanism based on the deep Q-network reinforcement learning framework, calculate the reward function according to the actual results of maintenance execution, and continuously update the parameters of the correlation analysis model, fault prediction model and resource scheduling model through iterative optimization to complete the construction of a fault maintenance model based on the combination of fault big data and equipment BOM.

2. The method for constructing a fault repair framework based on the combination of fault big data and equipment BOM as described in claim 1, characterized in that, The BOM knowledge graph described in step 2 has five layers: the first layer is the overall equipment layer, the second layer is the functional subsystem layer, the third layer is the component layer, the fourth layer is the part module layer, and the fifth layer is the component layer. Each node contains the part number, name, specifications, supplier information, expected lifespan, inventory quantity, and replacement cycle attribute. The BOM knowledge graph is stored in a graph database, which supports relationship-based path queries and subgraph retrieval, enabling fault impact range analysis and component association traceability.

3. The method for constructing a fault repair framework based on the combination of fault big data and equipment BOM as described in claim 1, characterized in that, Step 3 uses the improved FP-Growth algorithm to perform itemset mining on the multi-dimensional fault dataset to obtain the association rules between fault modes and related components. A distributed FP-Growth algorithm based on MapReduce is used to process large-scale fault datasets, and the frequent itemset mining process is accelerated by constructing conditional FP-trees in parallel; a minimum support threshold is set. and minimum confidence threshold ; The frequently mined itemsets are sorted and filtered according to their lift, and the lift is calculated using the following formula: Only retain The positive correlation rules.

4. The method for constructing a fault repair framework based on the combination of fault big data and equipment BOM as described in claim 1, characterized in that, The improved Long Short-Term Memory network described in step 4 has an input layer receiving length of... Time series historical fault data ,in It includes equipment operating status vector, environmental feature vector, and time encoding vector; the LSTM hidden layer adopts a bidirectional LSTM (Bi-LSTM) structure, with 128 hidden units, 2 layers, and a dropout rate of 0.2; the BOM feature encoding layer adopts a 3-layer MLP with 64, 32, and 16 hidden units respectively, and the activation function is ReLU; the output layer outputs the probability distribution of fault occurrence and the predicted fault type through a fully connected layer and a sigmoid activation function.

5. The method for constructing a fault repair framework based on the combination of fault big data and equipment BOM as described in claim 1, characterized in that, The prediction function of the improved LSTM constructed in step 4 is: In the formula, This is the vector of fault prediction results. It is the Sigmoid activation function. This is the output layer weight matrix. This is the output layer bias vector. Historical fault data Temporal feature vector encoded by LSTM network, BOM characteristics The structural feature vector encoded by a multilayer perceptron (MLP), This indicates the feature vector concatenation operation.

6. The method for constructing a fault repair framework based on the combination of fault big data and equipment BOM as described in claim 1, characterized in that, The improved genetic algorithm described in step 5 is as follows: The maintenance resource scheduling scheme is encoded using real-number encoding, with each chromosome containing a maintenance personnel allocation sequence, a spare parts call sequence, and a tool configuration sequence. The selection operation adopts the tournament selection method, the crossover operation adopts the partial mapping crossover operator, and the mutation operation adopts the Gaussian mutation operator. The population size is set to 100, the crossover probability is 0.85, the mutation probability is 0.05, and the maximum number of iterations is 200. An elite retention strategy is introduced in the evolution process, where the 10% of individuals with the highest fitness are retained in each generation and directly enter the next generation.

7. The method for constructing a fault repair framework based on the combination of fault big data and equipment BOM as described in claim 1, characterized in that, Step 6 involves training a deep Q-network by minimizing the TD error, i.e. .

8. A fault repair framework construction system based on the combination of fault big data and equipment BOM, characterized in that, The system includes: The multi-source fault big data acquisition module is used to construct a multi-dimensional fault big data dataset; it performs data cleaning, format standardization, missing value handling, and outlier detection and removal on the acquired multi-dimensional fault big data dataset. The equipment BOM data modeling module is used to build a hierarchical relationship model of equipment components. It uses a tree structure to represent the subordinate relationships between equipment, subsystems, components, and parts, forming a BOM knowledge graph. The fault-component association analysis module is used to perform itemset mining on the multi-dimensional fault dataset using the improved FP-Growth algorithm to obtain the association rules between fault modes and associated components. The fault prediction and maintenance strategy generation module is used to perform fault prediction using an improved long short-term memory network that incorporates BOM features, and to construct a prediction function for an improved LSTM. The maintenance resource optimization and scheduling module is used to construct the fitness function of the improved genetic algorithm and to perform multi-objective optimization scheduling of personnel, spare parts and tools required for maintenance using the improved genetic algorithm. The closed-loop feedback and continuous optimization module is used to build a closed-loop feedback mechanism based on the deep Q-network reinforcement learning framework. It calculates the reward function based on the actual results of maintenance execution, and continuously updates the parameters of the correlation analysis model, fault prediction model and resource scheduling model through iterative optimization, so as to complete the construction of a fault maintenance model based on the combination of fault big data and equipment BOM.

9. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-7.

10. A computer device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the method of any one of claims 1-7.

Images