Power supply service management method and system based on multi-source data

Abstract

The invention relates to the field of power supply work order service management, in particular to a power supply service management method and system based on multi-source data. The method comprises the following steps: collecting a current fault repair work order based on a power supply service platform, carrying out deep work order text semantic analysis and standard fault instance modeling, and constructing a global fault instance map; recognizing a client real-time repair call audio stream, carrying out time window voice analysis one by one, and carrying out intelligent fault work order filling, thereby obtaining a real-time audio intelligent fault work order; performing fault demand decomposition on the real-time audio intelligent fault work order and the global fault instance map, performing power supply service resource dynamic configuration regulation and control, and constructing a service resource regulation and control engine; and performing global fault work order real-time processing based on the service resource regulation and control engine, performing overdue work order early warning, and generating a service timeliness early warning strategy. Through dynamic resource allocation and preventive work order creation, the power supply service efficiency, the power grid reliability and the intelligent level are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power supply work order service management, and in particular to a power supply service management method and system based on multi-source data. BACKGROUND

[0002] With the continuous expansion of the power system scale and the wide application of intelligent technology, the operation environment of the power supply network is increasingly complex, the types and number of faults have significantly increased, which brings great challenges to power supply service management. As the key bridge connecting user demand and power maintenance action, the management efficiency and service quality of power failure repair work order directly affect the safe and stable operation of the power grid and user satisfaction. The traditional failure repair work order management relies on manual scheduling and experience judgment, lacks deep mining and intelligent analysis of massive multi-source data, resulting in long work order processing period, unreasonable resource allocation, fault response lag and other problems, which is difficult to meet the requirements of modern smart grid for efficient and accurate service.

[0003] With the rapid development of Internet of Things, big data and artificial intelligence technology, a large amount of multi-source heterogeneous data emerges in the power supply system, including failure repair work order data, real-time monitoring data, user feedback information and environmental meteorological data, etc. These data contain rich fault characteristics and service rules, providing a solid data foundation for improving the intelligent control ability of failure repair work order. However, most of the current power supply service management systems have not fully utilized the fusion analysis ability of multi-source data, and still remain at the level of isolated data processing and simple rule judgment, which is difficult to realize accurate identification, intelligent scheduling and dynamic resource optimization of failure work order. In addition, the diversity and timeliness of power failure require the fault handling process to have high intelligence and real-time response capability. The traditional work order management mode is difficult to adapt to the rapidly changing fault environment, and cannot realize fault warning, operation path optimization and dynamic resource allocation, resulting in low fault handling efficiency, decreased user satisfaction, and even causing more widespread power grid risks. In the face of these challenges, it is a key technical requirement to build a power failure repair work order service intelligent management method and system based on multi-source data fusion, deep semantic analysis and intelligent scheduling algorithm, which promotes the transformation and upgrading of power supply service, and improves the reliability and service level of power grid. SUMMARY

[0004] To solve the above technical problems, the present application provides a power supply service management method and system based on multi-source data to solve at least one of the above technical problems.

[0005] To achieve the above purpose, the present application provides a power supply service management method based on multi-source data, comprising the following steps: Step S1: Collecting current failure repair work order based on power supply service platform, and performing deep work order text semantic analysis and standard fault instance modeling to construct global fault instance graph. Step S2: identifying the customer real-time repair call audio stream, performing voice analysis of each time window, and performing intelligent fault work order filling, thereby obtaining a real-time audio intelligent fault work order; Step S3: decomposing fault demand based on the real-time audio intelligent fault work order and the global fault instance graph, and dynamically configuring and regulating power supply service resources to build a service resource regulation engine; Step S4: real-time processing of global fault work orders based on the service resource regulation engine, and overdue work order early warning, generating service timeliness early warning strategies; Step S5: regional fault frequency distribution analysis based on the service resource regulation engine, and dynamic device fault risk prediction, obtaining multiple fault risk prediction points; Step S6: intelligent creation of preventive work orders based on the multiple fault risk prediction points, and fault operation timing and path planning based on the service timeliness early warning strategy, building an intelligent power supply service management model.

[0006] In the present specification, a power supply service management system based on multi-source data is provided for executing the power supply service management method based on multi-source data as described above, comprising: A fault instance module for collecting current fault repair work orders based on the power supply service platform, and performing deep work order text semantic analysis and standard fault instance modeling to build a global fault instance graph; An audio analysis module for identifying the customer real-time repair call audio stream, performing voice analysis of each time window, and performing intelligent fault work order filling, thereby obtaining a real-time audio intelligent fault work order; A resource regulation module for decomposing fault demand based on the real-time audio intelligent fault work order and the global fault instance graph, and dynamically configuring and regulating power supply service resources to build a service resource regulation engine; A timeliness early warning module for real-time processing of global fault work orders based on the service resource regulation engine, and overdue work order early warning, generating service timeliness early warning strategies; A fault risk prediction module for regional fault frequency distribution analysis based on the service resource regulation engine, and dynamic device fault risk prediction, obtaining multiple fault risk prediction points; An intelligent service management module for intelligent creation of preventive work orders based on the multiple fault risk prediction points, and fault operation timing and path planning based on the service timeliness early warning strategy, building an intelligent power supply service management model.

[0007] The beneficial effects of the present application are as follows: through natural language processing (NLP) technology for deep semantic analysis of historical work order text, long and unstructured descriptions can be converted into standard fields such as "fault type + performance characteristics + location + equipment type", realizing the structured modeling of power supply fault knowledge. A large number of historical repair work orders are clustered and modeled according to fault modes, affected areas, and treatment schemes, forming a graph-based knowledge network, which makes it possible to quickly compare similar faults, automatically fill out forms, and intelligently recommend treatment strategies. Through standard instance modeling, the fault data format under multiple power supply companies and different business systems can be unified, and the cross-regional and cross-system data fusion efficiency can be improved. Through automatic analysis of call content by speech recognition (ASR), semantic understanding (NLU), and speech enhancement technologies, key information such as "trip", "spark", "distribution box", and "third floor west side" can be automatically identified during user calls, and structured work orders can be generated in real time, avoiding repeated communication and manual input. The traditional manual recording and sorting of fault work orders are transferred to AI processing, greatly reducing the labor input and the probability of incorrect filling, and realizing the assistance or automatic processing of the agent. The processing method based on time window analysis helps to capture the continuity and context of the call semantics, improves the understanding of complex fault descriptions, and is especially suitable for multi-round and ambiguous customer repair. Combined with the real-time generated audio work order and graph information, the system can identify the skill types involved in this fault (such as high voltage, low voltage, and meter box), the required time limit, tools or vehicles, and realize the precise matching of service resources. Through the dynamic allocation of work order personnel and equipment by the resource regulation engine, idle resources can be dispatched or scheduling can be performed, which is suitable for high-concurrency repair scenarios and improves the timeliness of response. The engine supports multi-dimensional priority evaluation according to fault urgency, customer level, and historical response records, realizing the transformation from "passive scheduling" to "active cooperation". The regulation engine monitors the status of all assigned work orders in real time, predicts whether there is a service lag trend based on the current processing node time consumption, and automatically triggers the alarm mechanism to intervene in advance to avoid service violations. A closed-loop system is formed to assist dispatch personnel in real-time monitoring of the entire life cycle of work orders, improving the scientificity of scheduling. In the case of work order overtime or risk escalation warning, the system can automatically notify the superior command center, outsourcing units, and other linkage mechanisms for rapid cooperation, improving the coordination of emergency disposal. Through spatial, temporal, and equipment dimension analysis of fault work orders, the system can identify the trend of high recent fault frequency in a small area, block or distribution box, and make trend predictions based on historical graph models to identify high-risk points in advance. The system introduces time series modeling, clustering analysis, or deep learning prediction (such as LSTM) to predict the hotspots and equipment that may appear in the next time window, providing opportunities for resource arrangement and maintenance. According to the regional frequency distribution map and risk prediction points, it can be used to guide the priority of equipment replacement, grid-based operation resource allocation, and transformation plan development. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the steps of a power supply service management method based on multi-source data according to the present invention. Figure 2 This is a detailed flowchart illustrating the implementation steps of step S1. Figure 3 This is a detailed flowchart illustrating the implementation steps of step S2; Figure 4 This is a flowchart illustrating the detailed implementation steps of step S3. Detailed Implementation

[0009] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0010] This application provides a power supply service management method and system based on multi-source data. The implementing entities of the power supply service management method and system based on multi-source data include, but are not limited to, mechanical equipment, data processing platforms, cloud server nodes, network upload devices, etc., which can be considered general computing nodes of this application. The data processing platform includes, but is not limited to, at least one of an audio / image management system, an information management system, and a cloud data management system.

[0011] Please see Figures 1 to 4 This invention provides a power supply service management method based on multi-source data, comprising the following steps: Step S1: Collect current fault repair work orders based on the power supply service platform, and perform in-depth work order text semantic analysis and standard fault instance modeling to construct a global fault instance map; Step S2: Identify the real-time repair call audio stream of the customer, perform voice analysis for each time window, and fill in the intelligent fault work order to obtain the real-time audio intelligent fault work order; Step S3: Decompose the fault requirements of the real-time audio intelligent fault work order and the global fault instance map, and dynamically configure and control the power supply service resources to build a service resource control engine. Step S4: Based on the service resource control engine, perform real-time processing of global fault work orders, issue warnings for overdue work orders, and generate service timeliness warning strategies; Step S5: Analyze the regional fault frequency distribution based on the service resource regulation engine, and perform dynamic equipment fault risk prediction to obtain multiple fault risk prediction points; Step S6: Create preventative work orders intelligently based on multiple fault risk prediction points, and plan fault operation sequence and path based on service timeliness early warning strategy to build an intelligent power supply service management model.

[0012] In the embodiment of the present application, referring to Figure 1 A step flow diagram of a power supply service management method based on multi-source data of the present application, in the present example, the steps of the power supply service management method based on multi-source data include: Step S1: Collect the current fault repair work order based on the power supply service platform, and perform deep work order text semantic analysis and standard fault instance modeling to construct a global fault instance graph; In this embodiment, real-time collection of current fault repair work order data is performed through the power supply service platform interface. The collected content includes structured and unstructured fields such as work order number, repair time, fault description, fault equipment, fault location, and user information. The platform uses an API interface polling mechanism to obtain new or updated work orders every 5 minutes, ensuring real-time data. In the experimental environment, about 10,000 work orders are collected daily, with a data completeness rate of over 98%. The system preprocesses the work order data, including data cleaning and outlier filtering, to remove duplicate or incomplete information. For redundant fields, a rule-based field mapping and deduplication algorithm is used to ensure the consistency and accuracy of the input data, laying a foundation for subsequent semantic analysis. For the preprocessed work order text description, natural language processing (NLP) techniques are used for deep semantic analysis. First, the Chinese work order text is segmented using a segmentation tool, and stop words are removed. Then, the text is semantically encoded using a pre-trained BERT (Bidirectional Encoder Representations from Transformers) model to obtain a multi-dimensional semantic vector representation. In the experimental parameter setting, the BERT model uses the base version, the maximum text length is set to 256, the batch size is 32, and the training is fine-tuned to adapt to the power supply fault corpus. Based on semantic encoding, combined with named entity recognition (NER) technology, key fields such as fault equipment name, fault type, and fault location are automatically extracted. The conditional random field (CRF) model is used for accurate labeling of entities, with an identification accuracy of 92%. In addition, dependency syntax analysis is used to further extract event relationships, clearly defining the scope of the fault and the user group. After extracting multi-dimensional semantic features, the system uses similarity clustering algorithms (such as K-means) to classify and optimize work orders, eliminating duplicate repair records for the same fault and improving data utilization efficiency. Experiments show that the clustering accuracy is 85%, effectively reducing redundant work orders by about 20%. Based on the results of semantic analysis, the system constructs a standardized fault instance model. This model includes fault equipment type, fault location coordinates, fault impact range, repair timestamp, and user affected group, among other core dimensions. Using multi-dimensional data fusion technology, the system integrates the features of each dimension into a unified fault information matrix. For a large number of fault instances, the system uses graph database technology (such as Neo4j) to construct a global fault instance graph. The nodes in the graph represent fault equipment and locations, and the edges represent the temporal or spatial relationships between faults. Through graph neural networks (GNN), the attributes of nodes and edges are learned, enabling deep mining of fault relationships and simulation of fault propagation patterns. Based on the fault graph constructed from about 180,000 work orders over a six-month period, the node coverage rate is 95%, and the graph is updated dynamically every 30 minutes, supporting real-time queries and intelligent inference. The graph is applied in fault prediction and resource scheduling, improving the fault response accuracy by about 10%.

[0013] Step S2: Identify the customer real-time repair call audio stream, perform voice analysis for each time window, and perform intelligent fault work order filling, so as to obtain a real-time audio intelligent fault work order; In this embodiment, the call center system of the power supply service platform accesses the customer's real-time repair call audio stream in real time. The system uses high-quality audio acquisition equipment and protocols to ensure a sampling rate of at least 16 kHz and 16-bit quantization accuracy, ensuring the clarity and detail of the sound signal. The call audio is preprocessed by an echo cancellation, noise suppression, and signal gain adjustment module to improve the accuracy of subsequent speech recognition. The accessed audio stream is transmitted in real time and continuously using streaming technology, with a delay of less than 200 ms, providing stable audio input for the next time window analysis. For continuous audio streams, the system defines a fixed-length time window, commonly 1 second, with a sliding step of 0.5 seconds, achieving a 50% overlap between time windows. This division method ensures continuous coverage of audio content, reduces semantic information gaps, and balances computational efficiency. The sliding window technique divides the audio stream into multiple overlapping time segments, facilitating individual speech recognition and semantic analysis for each time segment. This process handles approximately 10 hours of audio data in real-time, with a delay of less than 1 second. For each time window, the system uses an end-to-end automatic speech recognition (ASR) model for transcription. The Transformer-ASR model based on deep neural networks (DNN) and attention mechanisms is used, which has been trained on over 5000 hours of power domain speech corpus, including various accents and noise scenarios, with a word error rate (WER) of around 8%. The ASR output text data is further processed by a natural language processing module for word segmentation, entity recognition, and relationship extraction. Key extracted semantic features include fault location (e.g., "transformer," "distribution box"), fault type (e.g., "trip," "power outage"), fault condition description (e.g., "smoke," "abnormal noise"), on-site environment, and personnel casualty information. The BiLSTM-CRF model is used for entity recognition, with an accuracy of over 90%. After vectorizing the semantic features, the context information is combined for multi-round dialogue understanding, further improving the completeness and accuracy of fault description, ensuring that the work order content fully reflects the on-site situation. The extracted multi-dimensional semantic features are automatically mapped to the standard fault work order fields, enabling intelligent filling. The system uses a combination of rule engines and machine learning to automatically match work order template fields for different semantic information, ensuring the standardization of work order structure and the accuracy of information. Real-time audio intelligent fault work orders include work order number, fault location, fault type, fault detailed description, on-site situation, and personnel casualty information, and are dynamically updated to support real-time monitoring and subsequent fault handling. In experimental testing, this process generates approximately 15 real-time audio intelligent fault work orders per minute, with an information completeness of 92% and an accuracy of 88%, significantly improving fault response speed and information quality.

[0014] Step S3: decompose the fault demand of real-time audio intelligent fault work order and global fault instance atlas, and perform dynamic configuration and regulation of power supply service resources, to construct a service resource regulation engine; In this embodiment, the real-time generated audio intelligent fault work order is fused with the historical fault information in the global fault instance graph to form a complete view of the fault demand. During the fusion process, the key fields in the work order such as fault type, fault location, fault impact range, and user affected group are matched and associated with the corresponding nodes and relationship data in the graph to ensure the accuracy and timeliness of the demand information. The fault demand decomposition adopts a method combining rules and semantic analysis. On the rule level, according to the power industry fault handling specification, the complex fault demand is decomposed into specific sub-demands (such as equipment maintenance, fault positioning, personnel scheduling, etc.); on the semantic level, natural language understanding (NLU) is used to perform fine-grained semantic segmentation on the work order description to capture the implicit service demand and urgency. This process uses a multi-label classification model based on BERT to assist decision-making, and the model accuracy reaches 90% in experiments, greatly improving the comprehensiveness and fineness of demand decomposition. After completing the fault demand decomposition, the system further converts the demand into multiple executable sub-task sequences. Each sub-task corresponds to a specific fault handling step, such as on-site investigation, fault equipment maintenance, backup equipment allocation, etc. Task Graph modeling is used to structure the dependency relationship, execution order, and priority between tasks. For a batch of complex fault work orders, an average of 5-8 sub-tasks are generated, and the task graph construction success rate reaches 95%. Sub-task subdivision helps improve the modular management and scheduling efficiency of fault handling. The system quantitatively calculates the processing scheduling type and resource utilization for each sub-task based on the current service resource state and historical resource scheduling data. The resource portrait includes human resources (such as maintenance personnel skill level, idle state), material resources (backup equipment, maintenance tools), time resources (available time window), etc. A multi-dimensional weighted scoring model is used to match and score resource attributes and task demands to dynamically generate a fault handling resource portrait. In the experimental parameter setting, the resource matching algorithm response time is controlled within 100ms, and the matching accuracy reaches 88%, ensuring the timeliness and rationality of resource allocation. Based on the resource portrait of sub-tasks, the system introduces a reinforcement learning (RL) method to dynamically optimize resource allocation strategies. The regulation engine constructs a state space (current resource distribution, task urgency, etc.), an action space (resource allocation scheme), and a reward function (task completion timeliness, resource utilization, etc.) to perform policy iteration training, achieving the optimal balance of resource scheduling. The deep Q network (DQN) algorithm is used for training, and the training data covers multiple fault scenarios. The training period is about 100,000 steps, and after convergence, the resource utilization rate is improved by 10%, and the task processing timeliness is improved by 15%. The regulation engine supports real-time input, with a response time of less than 200ms, suitable for actual fault scheduling needs.

[0015] Step S4: Real-time processing of global fault work order based on service resource regulation engine, and overdue work order early warning, generating service timeliness early warning strategy; In this embodiment, the service resource regulation engine as the core scheduling module, continuously receives task data from the global fault instance atlas and real-time audio intelligent fault work order, dynamically allocates manpower, materials and scheduling path, realizes real-time processing of work order. Based on the resource allocation strategy obtained by reinforcement learning in the previous step, the engine adjusts the resource allocation in real time to respond to the change of fault demand. In the processing process, the system continuously tracks the state information of the work order, including "to be processed", "processing", "processed" and other nodes, and uses the state machine model to manage the life cycle of the work order. In the experiment, the work order state update frequency is set to 1 minute, which ensures the timeliness of the information. The response delay of the scheduling engine is controlled within 200ms, ensuring fast resource response. The system analyzes the work order state and the processing log filled in the work order in real time to evaluate the service progress of each work order. A progress estimation model based on time series analysis is used to compare the actual processing time of the work order with the estimated standard time, identify the current service stage and the remaining workload. The progress prediction model trained based on the historical work order processing time has an average prediction error of less than 8%, which can accurately reflect the real-time progress of work order processing and effectively support subsequent time management. In order to reasonably allocate limited service resources, the system designs a multi-factor work order priority calculation model. This model considers the fault influence range (number of users, key equipment), fault urgency (such as safety hazards), fault type and historical response time performance. The priority calculation uses a weighted scoring method, and the weights are determined through expert scoring and historical data optimization. For example, the safety hazard weight is set to 0.4, the user influence weight is 0.3, the historical overtime record weight is 0.2, and the fault type weight is 0.1. In the experiment, through the priority sorting of 10,000 work orders, the model effectively improves the response speed of high-risk fault work orders, and the customer satisfaction is improved by about 12%. Based on the priority calculation results, the system combines the pre-set standard processing time of the work order (such as within 24 hours), and performs service time analysis. Using the time difference algorithm, the system calculates the difference between the work order processing time and the time threshold in real time, and dynamically judges whether there is an overdue risk. When the remaining time of the work order is less than the warning threshold (for example, 2 hours remaining), the warning process is triggered. In the experimental environment, the system tests 50,000 historical work orders, and the overdue identification accuracy reaches 94%, effectively avoiding the omission of delayed work orders. For the identified overdue work orders, the system automatically generates multi-level service time warning strategies, including SMS push, background alarm and resource priority scheduling adjustment. The warning strategy dynamically adjusts the resource allocation strategy according to the work order priority and historical response data to ensure that the overdue risk work order is given priority. The system supports self-learning optimization of the warning strategy. Through the feedback loop, the system analyzes the work order processing effect after the warning, and continuously adjusts the warning threshold and response measures. In the experiment, after applying the self-learning strategy, the average processing time of the work order is improved by about 15%, and the overdue rate is reduced by 20%.

[0016] Step S5: According to the service resource regulation engine, the regional fault frequency distribution is analyzed, and the dynamic device fault risk prediction is carried out, and a plurality of fault risk prediction points are obtained; In this embodiment, based on the whole network fault work order data output by the service resource regulation engine, the system accurately locates the fault space position. Combined with the geographic information system (GIS) technology, the power supply network is divided into a plurality of adjacent fault regions, and the region division is based on the power grid topology and geographical adjacency. A typical division unit is a substation coverage area or a distribution line section. In the experimental parameters, the area of the smallest unit is controlled within 0.5 square kilometers to ensure the spatial resolution. For each division region, the fault type and fault frequency in a historical time window (such as the past 30 days) are counted to form a regional fault frequency distribution matrix. The time sliding window technology is adopted to dynamically update the fault frequency data, and the time window size and step are set to 7 days and 1 day respectively, so that the fault occurrence trend is continuously monitored. The statistical results are visualized as a heat map to reflect the fault intensive area and the fault type distribution characteristics. In order to enrich the analysis depth of the frequency distribution, the system combines the meteorological monitoring data (such as temperature, humidity, wind speed and lightning activity) with the historical fault data to extract multi-dimensional features. The association rule mining algorithm (such as Apriori) is used to find the potential association between regional meteorological factors and fault occurrence, and to extract regional meteorological-fault related features. Experimental data shows that a specific wind speed threshold (such as wind speed exceeding 15 m / s) is significantly positively correlated with the fault rate, and the correlation degree is more than 0.6. The system combines the device operation data (load rate, maintenance period) with the regional fault frequency, and reduces the dimension through principal component analysis (PCA), extracts the key risk indicators, and provides input features for subsequent fault risk prediction. Based on the above extracted regional fault frequency and multi-dimensional features, the system uses a time series prediction model to dynamically predict the device fault risk. The specific method uses a long short-term memory network (LSTM) model, which is suitable for capturing nonlinear and long-term dependence characteristics in time series. The model input includes historical fault frequency sequence, meteorological data sequence and device state features. In the training process, the fault and meteorological data in the past year are used, and the ratio of training set to test set is 8:2. The model evaluation indicators are mean square error (MSE) and accuracy. In the experiment, the LSTM model MSE is controlled within 0.015, and the accuracy is more than 87%, showing strong prediction ability. The model prediction result is mapped back to the power supply network space according to the geographical region, and the region with significantly increased risk value is identified and marked as a fault risk prediction point. The prediction point not only covers the fault high-risk area, but also can identify the potential risk increasing area, and realizes early warning. In practical application, the system sets a risk threshold, such as a risk score exceeding 0.75 is defined as a high-risk prediction point. The threshold is determined by ROC curve optimization according to historical risk and fault occurrence data, and the false positive rate and false negative rate are considered.

[0017] Step S6: According to a plurality of fault risk prediction points, an intelligent preventive work order is created, and a fault operation time sequence and path planning are carried out based on a service timeliness warning strategy, and an intelligent power supply service management model is constructed.

[0018] In this embodiment, these high-risk geographic locations and related equipment information are automatically converted into preventive maintenance work orders. The process uses a rule engine combined with a machine learning classifier to comprehensively evaluate the risk level of the risk points, equipment types, and historical failure records to determine the priority and maintenance content of the preventive work order. The rule engine automatically triggers the creation of a work order based on a risk score threshold (such as a risk value exceeding 0.75), and generates targeted maintenance tasks in combination with equipment operation manuals and past maintenance cases. The system scans risk prediction points every hour and automatically generates work orders to ensure the timeliness of preventive maintenance. In practical applications, the coverage rate of preventive work orders has increased by about 20%, avoiding failures in advance. The generated preventive work orders coexist with real-time work orders, and the system applies the service timeliness warning strategy to preventive work orders as well. Using the previously constructed multi-factor priority model (including risk level, failure impact range, maintenance resource shortage degree, etc.), the service priority of each preventive work order is calculated. The service timeliness threshold is set according to the type of work order, for example, the timeliness of high-risk equipment preventive work orders is controlled within 48 hours. The system dynamically adjusts the priority of work orders to ensure that high-risk preventive work orders are not overwhelmed by real-time fault work orders, optimizing overall maintenance timeliness. After the implementation of the warning strategy, the response time of preventive work orders is shortened by an average of 15% in experiments, greatly improving maintenance efficiency. After the generation of work orders, the system uses scheduling algorithms to plan the timing of fault operations. For all work orders (real-time and preventive), a heuristic algorithm (such as genetic algorithm, ant colony algorithm) combined with constraint optimization technology is used to consider work order priority, resource availability, job personnel skill matching, and safety specifications to determine the optimal operation sequence. The scheduling algorithm takes an average of 3 minutes to calculate in a million-level work order data test, ensuring the real-time nature of timing planning. The scheduling results are verified by simulation, with an operation completion rate of about 18% and a resource utilization rate of 12%. To reduce operation costs and response time, the system uses path planning algorithms (such as Dijkstra algorithm and dynamic path adjustment mechanism) to optimize the travel routes of maintenance personnel in combination with GIS map data and real-time traffic information. The algorithm considers geographic distance, traffic congestion, and operation timeliness requirements to achieve multi-objective path minimization. Path planning uses a 10-second-level response update to dynamically adjust the route in conjunction with the scheduling results, reducing the average travel time of maintenance personnel by 20%, greatly improving fault response efficiency. The above four modules of preventive work order creation, service timeliness warning, operation timing planning, and path optimization are integrated to form an intelligent power supply service management model. Based on multi-source data fusion, machine learning, and optimization algorithms, the model realizes closed-loop management of power supply services from fault prediction to operation execution. The model supports real-time data feedback and dynamic adjustment, and has self-learning and continuous optimization capabilities.

[0019] In this embodiment, refer to Figure 2For the detailed implementation step flowchart of step S1, in this embodiment, the detailed implementation step of step S1 includes: Collecting a current fault repair work order based on the power supply service platform; Performing abnormal redundant data filtering on the current fault repair work order, and extracting a filtered and optimized fault repair work order; Performing key field identification on the filtered and optimized fault repair work order, and performing deep work order text semantic analysis to extract multi-dimensional work order semantic features; Performing fault feature extraction based on the multi-dimensional work order semantic features to obtain a fault device type, a fault location, a fault influence, and a fault user group, and fitting to generate a multi-dimensional information matrix of the fault work order; Extracting a fault reporting timestamp based on the filtered and optimized fault repair work order; Modeling a standard fault instance on the multi-dimensional information matrix of the fault work order according to the fault reporting timestamp, and constructing a global fault instance graph.

[0020] In this embodiment, in the power supply service management system, first, the fault repair work order data submitted by the user needs to be collected from various power supply service platforms (such as 95598 power service platform, WeChat public number, enterprise exclusive repair interface, etc.). This step is the starting point of the entire early warning model data chain, and the data collection quality directly affects the subsequent processing effect. The collection content mainly includes work order number, user number, repair time, equipment number, geographic location information (latitude and longitude or address), user described fault content, access channel, etc. Usually, distributed data collection tools such as Kafka or Flume are used to realize real-time access to platform data, and structured (such as form fields) and unstructured (such as text description, voice transcription) two data types need to be compatible. The collection frequency is controlled to be once every 30 seconds to ensure real-time. In a day of data, about 20,000 work order data can be accessed on average. For multi-platform and multi-channel data, a unified data specification standard is used for access adaptation, including data cleaning, field mapping and unified coding processing, to ensure the consistency of subsequent data processing. In addition, to avoid data redundancy or repeated collection, the work order number and access timestamp are verified for uniqueness to ensure data accuracy. This step not only collects single repair information, but also links multiple source data such as power grid equipment operating state and historical work order library to provide multi-dimensional support for subsequent analysis. Repeat work order identification, low-quality description identification and empty field filtering. First, the repeat work order identification uses the hash value comparison method and the time-space clustering method to generate work order signatures by combining and encoding the repair time, address, user ID and other fields. If the consecutive work order signatures are highly consistent and the time interval is less than 5 minutes, it is marked as repeated; secondly, for low-quality descriptions, the information entropy and word frequency density of the work order text are calculated to identify hollow, too short or template type invalid descriptions; thirdly, for work orders with missing key information fields such as no address or equipment number, the system marks them as low-confidence work orders and removes them.

[0021] The TF-IDF-based text quality evaluation model is used for filtering threshold setting. For example, if the proportion of valid keywords in the work order text is less than 15%, the work order will be identified as low quality. Through this processing step, about 15,000 high-quality work orders can be retained from the initial 20,000 work orders, with a filtering rate of 25%. This step is crucial to ensure the effectiveness of subsequent semantic analysis and model training, avoiding the inclusion of false or interfering information in the fault modeling process. Structured and unstructured work order data are further analyzed to extract key semantic features from the work order through a text semantic understanding model. First, key field identification is performed, including fault time, fault location, user type, device type, and fault phenomenon information. For structured fields, direct extraction is possible, while unstructured description parts (such as "the house kept tripping at night" and "the elevator in the community was out of power") rely on deep natural language processing models for analysis. The BERT (Bidirectional Encoder Representations from Transformers) pre-training model, after fine-tuning on a power industry dataset, can effectively extract multiple semantic dimensions such as fault cause, affected equipment, and user experience from sentences. In the experiment, a labeled dataset of 100,000 historical work orders was used for fine-tuning, including "device fault type," "fault emotion tendency," "impact range," and other six dimensions. After model training, the F1-score reached 87.4%, achieving high extraction accuracy on the actual test set. In addition, BiLSTM-CRF structure is used for named entity recognition (NER) to assist in identifying fault locations and equipment information, realizing entity association and context understanding in the semantic space. Finally, each work order is represented as a multi-dimensional expression containing multiple labels and semantic feature vectors, providing semantic support for subsequent feature extraction and modeling. Through further aggregation analysis, the multi-dimensional fault features corresponding to each work order are summarized, and a standardized feature matrix is formed. The matrix generally takes "work order ID" as the primary key, with horizontal dimensions including device type (transformer / switch / user meter, etc.), fault location (latitude and longitude or administrative division), impact level (single household / community / district / region), affected user group (high voltage / low voltage / commercial / residential, etc.), and fault phenomenon category (trip / voltage anomaly / power outage / flash), totaling 12 dimensions. In the fault device type identification, the TextCNN classification model is used to quickly classify semantic features, with an accuracy of 92%. For location matching, Baidu Map API or Gaode address resolution module is used to perform geographic coordinate reverse lookup and standardization on ambiguous place names in the text. In terms of impact analysis, by linking with the distribution network topology map, the fault impact level is evaluated based on the number of users reporting, geographic density, and device impact range, and the K-means clustering analysis method is used to divide the work order impact into four levels.

[0022] The constructed "fault work order multi-dimensional information matrix" as a standardized intermediate data structure is beneficial to subsequent time series analysis and provides basic data support for building a global knowledge graph. The experimental results show that the modeling time of the multi-dimensional matrix is controlled within 1 minute, which is suitable for near real-time business processing. Accurate extraction of timestamps is of key significance for fault early warning and trend analysis. In this link, the system extracts the reporting time information from the filtered and optimized fault work orders, mainly from the system generation time field and user description text. First, the "submission time", "receipt time" and "processing time" recorded in the structured field will be parsed, and the "submission time" will be used as the basic timestamp of fault triggering. However, in some platforms or scenarios, there are time information in user descriptions, such as "power outage at 3 am today" and "sudden power failure last night", etc. This kind of semantic time needs to be structured by NLP time parser. The system uses HeidelTime time recognition module to standardize the natural language time expression in Chinese and convert it into a unified UTC format timestamp. In the experimental environment, the proportion of accurately parsed semantic time is 82.3%, and the errors mainly concentrate on ambiguous time descriptions (such as "recently" or "just now", etc.), in which case the default current submission time is used as the timestamp. The system also compares the fault timestamp with the historical data of the same period to eliminate the influence of some user delayed repair. For example, for work orders whose submission time is more than 2 hours later than the device outage time, a "delayed repair" label will be added. This timestamp serves as the starting point for building time series models, laying the foundation for fault trend modeling and early warning curve generation.

[0023] In this embodiment, refer to Figure 3 For the detailed implementation step flowchart of step S2, in this embodiment, the detailed implementation steps of step S2 include: Identify the customer real-time repair call audio stream; dynamically adapt the gain of the customer real-time repair call audio stream to obtain an adaptive gain audio stream; Define an equal-length time window, and perform sliding window decomposition on the adaptive gain audio stream to obtain a multi-time window audio stream; Perform speech analysis on the multi-time window audio stream in each time window to extract the audio semantic features of each time window; Based on the audio semantic features, fault information is mined to extract real-time fault information, including fault location, fault type, fault condition, on-site condition, and personnel casualty condition; According to the real-time fault information, intelligent fault work order filling is performed to obtain a real-time audio intelligent fault work order.

[0024] In this embodiment, the real-time recognition and access of audio data during the customer call process of the power supply service hotline (such as 95598) constructs the data source basis for subsequent processing. When the customer dials the repair phone, the customer service system accesses the audio stream through the IVR (Interactive Voice Response) platform. The audio stream is usually double-channel voice data, corresponding to the customer and the agent voice respectively. In the actual environment, in order to ensure the real-time processing, the system captures the audio stream based on WebRTC or SIP protocol, and uniformly converts it to 16kHz sampling rate, 16-bit single-channel PCM format, so as to facilitate subsequent voice processing. In the stream processing system deployed in the call center of a certain province, the average audio length of each repair call is 90 seconds, and the sampling frame number is about 144000 frames. The audio recognition process needs to realize low-delay access, so Kafka stream channel is used to cooperate with SparkStreaming processing module to ensure that each audio stream enters the real-time processing queue within 3 seconds. In addition, based on VAD (Voice Activity Detection) technology, the silent section is filtered out, and only the valid voice section is retained, further improving data utilization and processing efficiency.

[0025] Due to various external interferences (such as noise, unstable telecommunication signals, inconsistent microphone pickup distance, etc.) during the customer call process, the audio signal quality is uneven, and the recognition effect of the voice signal needs to be optimized through adaptive gain processing. The core goal of this step is to realize individualized audio gain processing based on voiceprint dynamic analysis, that is, to dynamically adjust the gain amplitude combined with the customer voice feature, to improve the signal-to-noise ratio and semantic analysis accuracy of the voice. First, the voiceprint feature of the caller is extracted through the voiceprint recognition module (SpeakerEmbedding + D-vector model), and compared with the historical voiceprint library to judge the characteristics of the caller such as gender, speaking style, tone strength, etc. Then, based on the characteristics, a gain control model (Adaptive Gain Control, AGC) is built to predict the optimal gain parameter through a deep neural network. The experiment uses the AGC model open sourced by Google Research as the basic framework, and performs transfer learning on the audio samples in the power industry, finally realizes the dynamic adjustment of the gain range in +6dB to +20dB, and the average signal-to-noise ratio is improved by 12dB. In order to meet the needs of time sequence continuity and real-time response of semantic analysis, the system needs to process the audio stream in time periods. This step divides the continuous audio stream into time domain through sliding window mechanism, cuts the entire audio signal into several equal-length time windows, and extracts the time sequence semantic features. The division principle is based on the balance between voice continuity and semantic carrying capacity.

[0026] The fixed window length is defined as 5 seconds, and the sliding step is 2.5 seconds, realizing the sliding window decomposition with 50% overlap to avoid important semantic information being truncated at the window boundary. In the experimental setup, the audio with a length of 90 seconds is divided into sliding windows, and about 34 audio segments are obtained, each containing about 80000 audio sampling points. To avoid semantic conflicts, a timestamp is attached to each window to ensure that the timing position of the audio in the overall context is preserved during processing. The short-time Fourier transform (STFT) with pre-processing is used to extract the spectral features of each window audio, and the amplitude spectrum and Mel frequency coefficients are retained to provide multi-dimensional acoustic information input for subsequent semantic feature modeling. The sliding window decomposition process is implemented by the TorchAudio module of the PyTorch framework, supporting GPU parallel processing, and each window decomposition takes less than 80 milliseconds, which can meet the real-time processing requirements. Through the speech recognition and natural language understanding model, the audio segments in each time window are analyzed one by one to extract the text semantic content and the implied fault element information. This process is executed in two steps: first, speech recognition (ASR), and second, semantic feature extraction.

[0027] The voice recognition part adopts an end-to-end voice recognition model based on a Conformer structure, which is fine-tuned on 100,000 Chinese power repair call samples to improve the recognition accuracy of industry terms such as "trip", "unstable voltage", "secondary line", etc. Experiments show that the model can achieve a WER (Word Error Rate) of 9.7% on the adaptive gain voice stream, which is about 15% higher than the traditional RNN model. The recognition result is output to the subsequent module in text form. The semantic features of the recognized text result are extracted, and a BERT-BiLSTM-CRF joint model is used for named entity recognition (NER) of the time window content to extract semantic labels such as fault phenomenon, device name, time point, user feedback, and danger prompt. At the same time, an emotion tendency analysis module is introduced to identify the urgency, anxiety or on-site risk prompt in the repairer's voice. Each time window outputs a multi-dimensional vector semantic feature set containing entity labels, keyword weights, and emotion values. Through the fusion analysis of multi-window semantic information, the specific fault content mentioned by the customer in the call is comprehensively judged. The core goal is to convert the natural language description into structured five-category fault dimension information: fault location, fault type, fault condition, on-site condition, and personnel casualty condition. First, a combination of rule guidance and deep learning is used for semantic fusion and context linkage judgment. For example, through the sliding window semantic graph construction mechanism, the keywords, location description, and time expression in multiple time windows are semantically associated to identify that "house tripping" and "smoke from the fire downstairs" belong to the same event chain. Second, a relation extraction model (RE, Relation Extraction) is introduced to identify semantic pairs between "device-fault type", "location-impact range", and "personnel-casualty status" to further structure into event triples. By introducing the Transformer-RE architecture model for training, the model's extraction accuracy for typical fault information is above 90%, especially in identifying key events such as personnel casualties and fires. In addition, the system introduces an "information confidence scoring mechanism" to label the confidence level of each fault information according to the extraction accuracy, improving the overall data reliability.

[0028] In the embodiment, the specific steps of dynamically adapting the voiceprint gain of the customer's real-time repair call audio stream to obtain the adaptive gain audio stream are as follows: Performing short-time Fourier transform on the customer's real-time repair call audio stream to obtain an audio spectrum graph; Performing abnormal frequency mutation detection on the audio spectrum graph to mark environmental noise points; Calculating the frequency mutation amplitude of the environmental noise points; Performing high-frequency filtering and denoising on the customer's real-time repair call audio stream according to the frequency mutation amplitude to obtain a high-frequency filtered and denoised audio stream; A user speech speed of the adaptive de-noising audio stream is calculated, semantic intelligibility is evaluated, multi-band voiceprint feature recognition is obtained, and a multi-band voiceprint vector is obtained. A dynamic voiceprint adaptive gain is performed based on the multi-band voiceprint vector, and an adaptive gain audio stream is obtained.

[0029] In this embodiment, the original time-domain audio signal is converted into a frequency-domain signal for subsequent frequency feature analysis and noise mutation detection. Since the customer call content often contains background noise, telecommunication transmission interference, static electricity sound, and other non-semantic information, a short-time Fourier transform (STFT) is required for time-frequency joint analysis. The STFT method divides the entire audio into several overlapping segments (frames), performs Fourier transform on each frame to obtain the frequency components corresponding to the frame. In the experimental environment, the frame length is set to 25 ms, the frame shift is 10 ms, and a Hamming window function is used for windowing processing to maintain spectral smoothness. After STFT processing of the 16 kHz audio, about 100 frames per second are obtained, and each frame spectrum contains 256 frequency points, finally forming a two-dimensional matrix-like audio spectrum graph (time x frequency dimension). The spectrum graph reflects the changes of each frequency component in the audio over time, and is the key basis for identifying noise, speech features, and mutation abnormalities. In the image presentation, the vertical axis of the spectrum graph is the frequency, the horizontal axis is the time, and the pixel intensity represents the amplitude of the frequency component. This graph will be an important input for the next step of mutation detection and noise identification, and is the core step of frequency domain analysis in the audio signal processing chain. First, the frequency dimension of the spectrum graph is subjected to first-order differential analysis to calculate the rate of change of each frequency component in the time dimension. Mutation points are usually manifested as a significant increase or decrease in the energy of certain frequency components within a very short period of time. In the experiment, the mutation threshold is set to a change in energy of a certain frequency band greater than 2.5 times the average value, and the existence of more than 2 frames in succession, which is determined as an "abnormal frequency mutation point". When the customer call background appears TV, fan, car horn, etc., the corresponding frequency band will instantaneously rise sharply, which is obviously different from the continuous frequency distribution pattern of normal speech. In the spectrum graph, these abnormal points are marked as "environmental noise points" and their time position, frequency range, energy peak value, etc. are stored for subsequent noise amplitude evaluation and filtering processing. In 500 sample calls, about 17% have significant mutation frequency bands. By accurately identifying these mutation points, the de-noising effect of the subsequent speech enhancement model can be effectively improved, and noise interference semantic recognition can be avoided. After marking the environmental noise points, the mutation amplitude needs to be quantitatively calculated to provide parameter support for filter design. This step measures the time gradient change rate of the noise point spectrum energy to evaluate its disturbance intensity and persistence. At the mutation point, the energy values of the previous and subsequent 3 frames are taken, and the difference ratio of the center frame and the surrounding frames is calculated. If the average energy of a certain frequency band in the previous and subsequent frames is E1, and the mutation frame energy is E2, the mutation amplitude is calculated as ΔE = (E2 - E1) / E1. In the experiment, the mutation amplitude level standard is set as follows: ΔE > 2 for strong noise (such as impact sound), 1 < ΔE ≤ 2 for medium noise (such as household appliance sound), and ΔE ≤ 1 for weak interference.In addition, the noise frequency range (e.g. 2500Hz-4000Hz), the number of continuous frames, the trend of change, etc. are recorded to form a noise description vector, which facilitates the dynamic adjustment of the parameters of the high-pass or band-pass filter in subsequent filtering. The quantification of the mutation amplitude helps to establish a dynamic filtering weight mechanism, making the filtering operation more targeted and with better speech protection capability. In the experimental results, the system analyzed 200 call records with different noise contents, and the mutation frequency was concentrated in the range of 2800Hz-5000Hz, with an average mutation amplitude of 1.8, indicating that the high frequency band is the main source of interference. This feature data will be used to accurately set the filter cutoff frequency, thereby removing non-semantic noise to the maximum extent while maintaining clear speech. After identifying the frequency mutation point and its amplitude, the next step is to perform high-frequency filtering and noise reduction, focusing on removing non-speech high-frequency interference signals. This step uses a dynamic high-pass filter or a band-pass filter to filter out the contaminated audio frequency band while preserving the speech frequency band (usually 300Hz-3400Hz). The filter design is based on the frequency mutation amplitude and frequency location extracted in the previous step: for strong noise segments with high ΔE values and frequencies above 4000Hz, a high-pass filter is used to remove them entirely; for frequency bands with moderate amplitude but obvious interference (such as television noise, with frequencies between 2000Hz and 3500Hz), a band-pass filter is used for local filtering. The filter function uses FIR design, and the passband edge is dynamically adjusted to adapt to the specific noise frequency. In actual testing, the system dynamically generates a frequency response graph for each audio segment to ensure that the filtering operation does not harm critical speech frequency bands. After filtering, the SNR (signal-to-noise ratio) of the audio is improved by an average of 8-10dB, and the speech intelligibility MOS score is improved to 4.3, with significant weakening effects on sudden strong noise such as knocking and electronic interference signals. The audio data after high-frequency filtering has a relatively high speech purity, and at this time it can be subjected to multi-dimensional acoustic analysis to extract personalized speech features such as voiceprint and speech rate. Speech rate is an important indicator for evaluating customer emotional state and call expression quality; voiceprint is used for individual identification and personalized processing optimization.

[0030] The speech rate evaluation method is based on the alignment of speech recognition results and timestamps: the number of recognized words per unit time (e.g., every 10 seconds) is counted and normalized by averaging. In the experiment, the normal speech rate range is set to 2.5-3.5 words per second, and exceeding this range will be marked as "fast speech" or "slow speech" and used as one of the semantic clarity evaluation factors. The semantic clarity evaluation combines ASR recognition confidence, syntactic integrity, and keyword coverage to give a comprehensive score, with a result between 0 and 1, with a higher value indicating more complete voice information and more explicit semantic expression. The voiceprint recognition part uses a multi-frequency voiceprint extraction technology based on the ECAPA-TDNN model, using the Mel frequency spectrum of the voice, multi-frequency energy distribution, phase spectrum, etc. to construct a 30-dimensional voiceprint vector, capturing the individual characteristics of the customer's timbre, pitch, and intonation. In the experimental setup, the voiceprint model is trained using the power customer service voice dataset, with an identification accuracy of over 96%, especially suitable for repeated repair identification and voice enhancement strategy adaptation. Based on the customer voice characteristics revealed by the voiceprint feature vector in the previous step, dynamic voiceprint-driven gain control is performed to maximize the analyzability and semantic quality of the audio stream. This process is equivalent to personalized voice optimization for each customer's call sound. The system introduces a DNN regression model to predict the optimal gain value range based on the voiceprint vector input. The model output includes the overall gain amplitude (e.g., +8dB) and the frequency band priority gain (e.g., +2dB for the 1000-2000Hz band), and adjusts the gain distribution in combination with the speech rate and clarity parameters. For example, for customers with fast speech and low voice energy, the model will output a gain strategy that enhances high-frequency bands and stabilizes low-frequency bands, avoiding semantic compression distortion. In the test phase, the system performs dynamic gain processing on audio with different speech rates and voiceprint types, with an identification rate improvement of about 11% compared to no gain processing, especially in poor channel or weak voice scenarios, where dynamic voiceprint gain shows obvious advantages. The final "adaptive gain audio stream" serves as the key input for subsequent fault semantic recognition and intelligent work order filling, not only preserving user individual characteristics but also greatly improving the overall robustness and intelligence of the voice signal processing chain, making it an important audio processing achievement for building a highly reliable power supply early warning system.

[0031] In this embodiment, reference Figure 4 The detailed implementation steps of step S3 include: The real-time audio intelligent fault work order and the global fault instance graph are decomposed to obtain real-time fault demand information of each work order; According to the real-time fault demand information, multiple sub-task sequences are generated; According to the multiple sub-task sequences, processing scheduling type and resource utilization are quantitatively calculated, so as to obtain the fault processing resource portrait of each sub-task; Based on the fault processing resource image, power supply service resource dynamic configuration and regulation are performed, and a service resource regulation engine is constructed.

[0032] In this embodiment, the real-time generated audio intelligent fault work order is jointly analyzed with the existing global fault instance graph, and the specific disposal requirements behind each work order are mined. Due to the diverse sources of work order information, both natural language audio analysis results and knowledge such as historical similar events and device node attributes in the graph are combined, so the demand decomposition needs to integrate semantic understanding and knowledge matching in two dimensions. The system performs structured semantic analysis on real-time work orders, including extracting key fields such as “fault type”, “fault location”, “user density”, “response time limit”, “danger level” and the like. Then, the work order content is compared with the known instances in the graph (such as “transformer overload trip” and “user concentrated area power supply interruption”), a BERT embedding + cosine similarity model is used, and a similarity threshold of 0.85 or above is set as “highly similar requirements”, so as to further deduce the demand granularity with the help of historical cases. For example, for an audio work order description “no electricity in the user's home, and sparks at the lower power switch”, the system identifies its high-risk characteristics, and finds the processing path of the past “low-voltage switch short circuit” event through the graph, thereby deducing requirements including “electrical detection”, “cable replacement”, “distribution cabinet maintenance” and the like. After completing the semantic extraction of the work order fault requirements, the system needs to convert these requirements into executable task units, i.e. “multi-task decomposition”. The core of multi-task decomposition is to divide the complex fault disposal process into a sequence of sub-tasks with time sequence correlation, clear roles and resource matching, so as to ensure that the dispatching system can be efficiently executed. The decomposition strategy adopts a “task decomposition model based on operation dependency graph”. First, the system internally builds a power supply service operation knowledge graph, which defines typical fault disposal processes such as “fault determination → employee verification → device power-off → component replacement → power restoration → user follow-up” and the like. Then, the real-time requirement information is mapped to the graph structure, the nodes and paths are matched, and the operation flowchart of the specific work order is constructed. If the requirement is “transformer burnout”, the decomposed tasks include “electrical test”, “temporary power supply configuration”, “burned equipment removal” and “new equipment installation”, each of which is further refined into a number of sub-steps (such as test insulation, cable wiring, transportation scheduling and the like). The system uses DFS (depth-first search) to generate task paths, and sets task priority (high / medium / low) and time limit (such as some tasks requiring completion within 2 hours). In the experimental parameter setting, each fault work order is decomposed into 4-7 sub-tasks, and the decomposition time does not exceed 2 seconds. The task dependency relationship is represented by DAG (directed acyclic graph), which supports advanced scheduling logic such as concurrent execution judgment and task conflict verification. The completion of multi-task decomposition enables the power supply fault to be transformed from “macroscopic description” to “scheduling execution unit”, and establishes a clear task basis for subsequent resource image and regulation configuration.

[0033] After the sequence of sub-tasks is determined, the system enters the resource assessment and matching phase, which requires quantitative analysis of the processing capacity, tool equipment, human roles, and other resources required for each task, i.e., establishing a "fault handling resource portrait". The core of this process is to assess the execution requirements of the task and map them with the current available resources. The resource portrait is composed of multiple dimensions: ① personnel resources (such as high-voltage electricians, repair technicians, and engineering vehicle drivers); ② equipment resources (such as insulation testers, power cut-off switches, and cable coils); ③ time resources (the number of hours required to complete the task); ④ geographical resources (the response radius from the fault location); and ⑤ response level (such as night repair, rain day construction conditions, etc.). The system constructs a "task-resource consumption" matrix by associating task characteristics with historical data and estimates resource requirements based on a multi-dimensional regression model. For example, for the "device replacement" task, the model predicts that 2 people are needed for 1.5 hours, 1 crane and 1 roll of cable are needed, the estimated working hours are 1.8 hours, and the emergency level is medium. The training set of the resource quantification model contains 50,000 historical dispatching data from the past 5 years, with an error control of ±7%. In addition, the system also analyzes the resource consumption patterns of different task types based on K-Means clustering, establishes standardized resource templates, and further improves the efficiency of portrait construction. Finally, each sub-task output is a set of structured resource portrait vectors, including resource items, quantities, time, urgency, and replaceability, which are the direct input for resource scheduling.

[0034] After obtaining the resource image of all sub-tasks, the system enters the final stage: according to the task priority, resource occupancy, scheduling type and other parameters, the dynamic configuration and control of global power supply resources is implemented. The goal of this step is to build a "service resource control engine" that responds to real-time work orders, optimizes resource utilization, and improves fault handling efficiency. The control engine is based on a multi-factor scheduling optimization algorithm, and the core model is a reinforcement learning scheduling network with constraints. First, according to the time limit, geographical distribution and emergency level of the task, a scheduling priority queue is constructed for all sub-tasks. Then, combined with the resource image library, a heuristic matching algorithm (such as the Hungarian algorithm or multi-objective genetic optimization) is used to find the optimal configuration combination of current resources. If there are multiple concurrent tasks competing for the same type of device (such as emergency generators), the system constructs a priority coefficient by multiplying the "task urgency x resource occupancy rate", and real-time retrieves the surrounding schedulable resource pool (GIS positioning combined with resource state monitoring), to achieve optimal allocation. In addition, the system also introduces a simulation mechanism to evaluate the impact of scheduling results on the overall service load, such as predicting that a certain scheduling strategy will improve the overall processing efficiency by 12.4% within the next 2 hours. The engine is deployed in the power supply service unified scheduling center platform, and resource state updates are performed every 5 minutes, and supports immediate rescheduling in the case of emergencies. Experimental data shows that in the disaster weather scenario, this control mechanism shortens the average response time of peak faults from 53 minutes to 38 minutes, and the resource utilization rate is increased by more than 17%.

[0035] In this embodiment, step S4 includes the following steps: Based on the service resource control engine, real-time processing of global fault work orders is performed, and real-time state information of fault work orders is collected; According to the real-time state information of fault work orders, fault service progress recognition is performed to obtain the real-time service progress of each work order; The real-time fault demand information is subjected to work order service priority calculation to obtain the work order service priority; According to the work order service priority, service timeliness analysis is performed to generate the service timeliness of each work order; Based on the service timeliness, the real-time service progress is subjected to timeliness overdue work order recognition, and overdue work order warning is performed to generate a service timeliness warning strategy.

[0036] In this embodiment, the execution of the fault work order is promoted by the service resource regulation engine, and various dynamic state data in the processing process is collected and processed in real time to form a closed-loop data feedback mechanism. The resource regulation engine has matched manpower, equipment, timeliness and task path according to the fault resource portrait. This step is mainly responsible for tracking and managing the real-time landing execution of these configuration schemes. After the system pushes the fault work order to the power supply dispatching subsystem, it starts to receive state update events in real time through the work order execution interface. The state information includes task reception, on-site check-in, job start, device replacement completion, power recovery, user follow-up and other key nodes. Each node is reported in real time through a mobile job terminal (PDA, smart watch, operation and maintenance APP) and a GIS positioning module, and is marked with a timestamp and a geographic label. The system also collects resource state changes in the execution process, such as personnel work hour records, device usage, construction environment feedback, etc., to form a complete "work order state flow". To ensure data stability, the system sets a state refresh every 60 seconds, and supports event-driven data push. Data transmission is stable and encrypted through a 5G edge gateway to ensure low latency and high availability. The system tracks 2000 work orders in real time, with a state node coverage rate of 96.8%, and a time delay control within 2.3 seconds, significantly improving the visualization and feedback efficiency of work order flow. It is the basic data source for realizing service progress identification and risk prediction. After completing the real-time state data collection, the system analyzes and models it to calculate the accurate service progress percentage of the current work order as a key indicator for subsequent priority adjustment and early warning judgment. Progress identification is not simply a state jump statistic, but a comprehensive consideration of task complexity, subtask time consumption, execution quality and other multi-dimensional factors to form a more refined "progress portrait". The system first sets a standard task path according to the task DAG graph (generated by the task decomposition module) contained in the work order, and then combines the state feedback and completion time of each task node to calculate its weight completion rate. For example, a work order includes 5 subtasks, each subtask is assigned a different progress weight according to resource consumption and risk level (such as task 1 accounts for 20%, task 2 accounts for 30% …), and the current completion tasks 1 and 2, the service progress is 20% + 30% = 50%. At the same time, the system introduces the "progress offset" indicator to evaluate the progress ahead of schedule, normal or delay by comparing the current work order progress with the average progress of historical similar work orders (based on graph historical work order samples). The offset formula is ΔP = Pcurrent - Phistorical average, positive value means ahead of schedule, negative value warns delay risk.

[0037] The determination of service priority directly affects the resource reallocation strategy and task execution order, and is the key parameter of service scheduling intelligence. The priority calculation is based on a multi-factor fusion model, covering dimensions such as fault nature, impact range, risk level, user density, task progress, and resource tightness, and is quantified into numerical levels (such as 1-5 levels) through a weighted decision method. The model design uses the analytic hierarchy process (AHP) to build a priority factor weight system, with specific weight settings as follows: fault type (25%), user impact (20%), on-site danger (20%), task progress lag (15%), resource scarcity (10%), and historical emergency response records (10%). The system standardizes each factor score (e.g., 0-1), and calculates the final priority score through weighted total score. For example, a certain work order is "residential area main cable fault", affecting more than 200 users, with a high risk level, a current progress lag of 15%, and resources in a high occupancy state, resulting in a final score of 0.85 (high priority), which will be marked as a first-level repair task. The priority model has a consistency of 91.7% when compared with artificial expert scoring. This index is widely used in resource occupation scheduling, early warning threshold setting, and other scenarios, and is an important decision basis for dynamically adjusting power supply scheduling strategies. Based on the priority confirmation, the system further conducts service time efficiency analysis to predict the remaining processing time of each work order and determine whether there is an overdue risk. This step not only focuses on time itself, but also on the impact of task structure, resource allocation, and external environment on processing efficiency. The time efficiency analysis model uses a time series prediction algorithm based on LSTM (Long Short-Term Memory), with input including current work order progress state sequence, priority level, resource scheduling log, historical work order processing samples, etc., and output including predicted remaining processing time and risk confidence. At the same time, a service level standard system is introduced: first-level repair (≤1 hour), second-level fault (≤4 hours), and third-level non-urgent task (≤24 hours). The system compares the LSTM prediction time with the service standard to calculate the "time efficiency deviation value" TΔ = Tpredicted - Tstandard, and sets a threshold strategy, such as TΔ>30 minutes triggering the time efficiency risk label. The system actively identifies and issues a risk warning for work orders that may exceed the processing time limit, to implement early intervention, scheduling reconstruction, and resource augmentation, thereby avoiding actual power supply interruption escalation or user complaints. The overdue identification logic is based on a comparative analysis model: if the "predicted completion time" Tpredicted of the work order is greater than the "regulation service time" Tstandard, and the current progress growth rate is less than the average growth rate × 0.8, it is determined as a "potential overdue work order". In addition, the system also introduces a service behavior anomaly monitoring mechanism, such as long-term stagnation of state updates, inconsistency between on-site check-in and actual location, etc., which will also trigger an "overdue warning signal".The early warning strategy adopts a hierarchical processing mechanism: slight overage (≤ 30 minutes) is optimized through scheduling rearrangement; moderate overage (30-60 minutes) pushes resource supplement suggestions to dispatchers; severe overage (≥ 60 minutes) automatically triggers emergency response schemes, including enabling a backup resource pool, adjusting the order of other work orders, deploying a remote support team, etc. The system presents the overage warning results in a visual layer superimposed manner on the dispatch console through the service timeliness early warning strategy engine. The warning information includes: work order number, location, overage level, recommended response mode, etc. According to the feedback of dispatch personnel and the system execution record, the early warning response efficiency is improved by 36%, and the average user satisfaction is increased by 15%.

[0038] In this embodiment, step S5 includes the following steps: According to the service resource regulation engine, the fault space position and fault type of all work orders are calculated; The adjacent area of the fault space position is divided to obtain a plurality of fault work order areas; Current meteorological monitoring data is collected according to the plurality of fault work order areas; According to the fault type and the current meteorological monitoring data, the regional meteorological-fault correlation analysis of the plurality of fault work order areas is performed to obtain the regional meteorological-fault correlation characteristics; According to the fault type, the fault type frequency statistics are performed, and according to the plurality of fault work order areas, the regional fault frequency distribution analysis is performed to generate a regional fault frequency distribution graph; Based on the regional meteorological-fault correlation characteristics, the dynamic equipment fault risk prediction is performed on the regional fault frequency distribution graph to obtain a plurality of fault risk prediction points.

[0039] In this embodiment, by deploying a geographic information analysis component in the service resource regulation engine, the spatial coordinates and fault types of each fault work order are extracted to realize the basis of subsequent regional analysis. After the engine receives each real-time work order, it first calls the GIS positioning module to parse the structured fields (such as detailed address, place name in user feedback) in the work order into latitude and longitude coordinates. The parsing uses the Gaode / Baidu map API or the built-in geographic coding system, and combines with the context semantic correction (such as "a community power supply room" is associated with the enhancement of the high-voltage line segment), and the parsing accuracy can reach ± 50 meters. The system identifies the fault type in the work order, such as "transformer fault", "line short circuit", "voltage drop", etc., which comes from the classification field provided by the previous work order intelligent filling module. The fault type data is encoded in a standard enumeration form, such as Type 1 = "circuit breaker action", Type 2 = "distribution box fault", Type 3 = "trip", etc.

[0040] For 6000 cloud work orders in a certain provincial power grid, the resolution rate reached 98.8%, and the success rate of high-precision coordinates was 95.3%. The accuracy rate of fault type mapping reached 94.5% through artificial sampling inspection. Thus, the system can automatically establish a fault space database containing the four-tuple of "work order ID-longitude-latitude-fault type", providing core data support for subsequent spatial clustering and risk prediction. After obtaining the spatial positions of all work orders, spatial clustering analysis is needed to form geographical partitions for regional data statistics and risk modeling. In this process, the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) and grid division combined method are mainly used for partition clustering of spatial points. According to the geographical distribution density, DBSCAN clustering is performed on all fault coordinate points, with ε (neighborhood radius) set to 0.5 kilometers and MinPts (minimum clustering points) set to 5, to identify fault hot spot areas with better aggregation. At the same time, for areas with insufficient density but important administrative boundaries (such as township boundary ranges), the system is supplemented with administrative area layers for slicing to ensure a coverage rate of 100%.

[0041] In the preliminary spatial clustering, a total of 270 DBSCAN clusters were generated from 6000 work orders, with an average of 22 fault records per cluster. After combining the boundary completion, a total of 310 valid fault work order areas (including hot spot areas and administrative areas) were generated. The average area of the region is about 3 km², which can balance the precision and statistical stability. The clustering process takes less than 2.2 seconds, which is suitable for real-time processing requirements at the provincial level. After completing the regional division, the system needs to collect the corresponding weather conditions for each region in real time to realize the correlation analysis of “weather-fault”. This step involves dynamic retrieval of multiple data sources, including national meteorological station data, regional public weather API, and local high-frequency Internet of Things (IoT) weather sensor network. The system generates geographic center coordinates for each region and calls the weather interface based on this to obtain multiple weather indicators such as temperature, humidity, precipitation, wind speed, and lightning probability; for regions with IoT sensors installed, high-resolution (every 5 minutes) local data such as instantaneous wind speed and humidity drop information are also collected. The weather collection frequency used is every 10 minutes to ensure timeliness. The results of collecting 14 days of daily data for the 310 regions show that the API response rate is 99.2%, the proportion of IoT abnormal data is less than 3%, and the available data coverage rate is more than 96%. Through the weather database interface, the temperature error is controlled within ±0.5℃, the precipitation error is controlled within ±0.3mm, and the wind speed sensor error is less than 0.2m / s. High-quality weather data provides a reliable environmental basis for regional fault correlation analysis. To explore the correlation between weather factors and fault types in each region, we refine the correlation features for use in the prediction model. In implementation, the system first establishes a feature matrix of fault types and weather variables, with row dimensions as regions and column dimensions as fault type frequencies and weather indicators. Statistical analysis and machine learning are combined to mine the correlation. The specific process is as follows: first, use the Pearson correlation coefficient to conduct preliminary linear analysis of fault frequency and weather factors in each region; for indicators with high significance (p-value <0.05, |r|>0.3), record them as main correlation items. Further introduce random forest regression to evaluate the influence of weather features on the frequency of a certain fault type, and extract the feature importance (Gini Importance) index.

[0042] The linear correlation coefficient between regional wind speed and line short-circuit fault Type 3 is 0.42, and the correlation coefficient between humidity and trip accident Type 1 is 0.38, both of which are significantly higher than the statistical threshold. Random forest model further evaluation shows that: wind speed explains 26% of short-circuit fault, and humidity explains 18% of trip fault, both of which are higher than the redundancy threshold of 15%. The "regional meteorological-fault correlation characteristics" formed therefrom include, such as "wind speed short-circuit risk index", "humidity trip risk ratio", etc., which can be used for subsequent risk prediction models. The visualization analysis step of the overall regional fault status not only needs to statistically analyze the overall trend of the fault type, but also needs to show the fault density distribution of each region. Based on the regional work order data, the system calculates the regional frequency of each fault type, and constructs a frequency distribution heat map for different types. The system classifies faults by type, and counts the number of fault records in each region in the last N days (such as 14 days), and then normalizes the area to calculate the fault density per unit area. The result is generated into a contour heat map, a color block risk map, and a hot spot marker bubble map through the GIS visualization platform. The color coding is mapped according to the standard gradient, with blue-yellow-red corresponding to low frequency-medium frequency-high frequency regions. The "short-circuit" type in the hilly area with strong wind reaches a frequency of 8 times / km², which is mapped as red; the short-circuit frequency in the urban center is 3 times / km², which is mapped as orange; and the other regions are in the low frequency area. The fault frequency distribution map is updated in real time in the dispatcher console, and the model is refreshed every 6 hours to assist decision makers in quickly identifying high-risk areas. Combined with the meteorological-fault correlation characteristics in the region and the current meteorological state, the potential fault risk in the future is spatially predicted. The system constructs a prediction model based on Gradient Boosting Machine (GBM), and the input features include: current meteorological variables, historical fault frequency, regional characteristics (grid density, equipment type proportion), and the above correlation characteristics. The model training uses 14 days of historical data to predict the probability of each region and each fault type in the next 1-3 days. The output layer is set to the predicted fault "hot spot probability score", such as the probability of a certain region occurring a short-circuit fault in the next 24 hours is 0.27. Using AUC, precision and recall rate evaluation indexes, 36 experimental samples can be obtained, AUC reaches 0.82, recall rate is 78%, and precision rate is 65%, which is better than the historical average benchmark model (AUC=0.75). The prediction result is presented in the form of a risk bubble in the heat map, and the bubble size is proportional to the probability value, effectively providing a clear target area for early warning dispatch. The system can automatically generate "fault risk prediction points" and send them to the dispatch center and mobile repair terminal, forming a closed-loop mechanism of "early warning-preparation of resources-advance inspection", which maximizes the reduction of the impact of meteorologically driven risks on the power supply system.

[0043] In the embodiment, step S6 includes the following steps: According to the preventive work order intelligent creation of multiple fault risk prediction points, a fault prediction preventive work order is obtained; Based on the real-time audio intelligent fault work order, the global fault instance atlas and the fault prediction preventive work order, service area fault information mapping is performed, and a power supply network digital twin model is constructed; Based on the service timeliness early warning strategy and the service resource regulation engine, the power supply network digital twin model is subjected to fault operation timing and path planning, and an intelligent power supply service management model is constructed.

[0044] In this embodiment, the risk prediction points obtained based on the regional meteorological-fault correlation analysis are converted into executable preventive work orders, and the active maintenance strategy of "early warning before occurrence and processing before fault" is realized. The system first decodes the attributes of each fault risk prediction point, including geographic location coordinates, risk level, fault type, and predicted risk time window. The intelligent creation of preventive work orders adopts a combination of rule-driven and machine learning. Specifically, according to historical fault data and fault prediction probability, high-risk points are matched with corresponding preventive maintenance action templates (such as "line inspection", "device reinforcement", "transformer insulation detection", etc.). The template content includes work order number, task description, priority setting, and estimated completion time. The system further applies a Bayesian inference model to correct the confidence of the predicted risk, combines work order execution history and resource consumption data, and dynamically adjusts work order priority and resource allocation suggestions. In practical application, for the line short circuit risk caused by wind speed warning in a certain area, the system automatically creates a "temporary line maintenance" preventive work order, with priority set to "high" and estimated operation time set to 48 hours before the risk period. Through the intelligent creation of preventive work orders for 300 risk prediction points, the creation accuracy rate reaches 92%, the dispatch response time is shortened by an average of 30%, and the preventive maintenance rate is increased by 18%. This mechanism significantly improves the proactivity of fault response and resource utilization, effectively reducing sudden power outage events.

[0045] Digital twin model construction is a key step to realize intelligent management of power supply network. The system integrates multi-source data: real-time audio intelligent fault work order provides dynamic fault semantics and on-site situation information, global fault instance graph brings historical fault correlation and spatial and temporal characteristics, and fault prediction and prevention work order preposes potential risk points and maintenance plans. First, the three types of data are unified in time and space semantic fusion. The time-space data alignment algorithm is used to map the work order information at different time points to the power supply network topology graph, and the GIS system is used to map the fault points, device locations, and line topology to graph nodes and edges. The graph database (such as Neo4j) is used to store and manage the topology network. Based on the graph neural network (GNN), the correlation between fault nodes is deeply mined to capture the time-space fault propagation rules and influence range. The model input includes node attributes (device type, current fault state, warning level), edge weight (voltage level, line distance, device connection strength), and work order semantic embedding vector. Through training, GNN can predict the fault propagation path and potential impact area, and realize dynamic fault simulation. Using about 20,000 fault work orders and real-time audio work order data in the past six months, a digital twin model is constructed, with a node prediction accuracy of 87% and a fault propagation path prediction recall rate of 75%. The system updates the fault mapping in real time, refreshing every 5 minutes to ensure the real-time and accuracy of the digital twin model, providing strong support for dispatching decisions.

[0046] By integrating the dynamic fault mapping information of the digital twin model and the existing service timeliness warning strategy and resource regulation engine, intelligent dispatching management is realized. This model aims to optimize the execution order and path planning of fault handling operations, improve maintenance efficiency and ensure service quality. It includes two aspects: job time sequence optimization and path planning. Job time sequence optimization is based on the service timeliness warning strategy, which uses priority queue and task dependency to build a time sequence scheduling model. The system input includes work order priority, estimated completion time, on-site personnel and device availability. Heuristic scheduling algorithms such as genetic algorithm and simulated annealing are used to calculate the optimal job sequence to ensure that high-priority overdue risk work orders are handled first. Path planning is based on the spatial distribution of fault nodes and network topology information in the digital twin model, and uses improved Dijkstra or A* algorithm combined with real-time traffic and traffic prediction for shortest path planning, supporting multi-task path joint optimization. Path planning also considers the current positioning of maintenance personnel and traffic time consumption, and dynamically adjusts the route. Experimental comparison shows that in 1000 work order dispatching tests, the average maintenance response time is shortened by 25%, the work order overdue rate is reduced by 15%, and the work order overdue rate is reduced by 15%. The system can dynamically adjust the dispatching scheme according to the real-time resource state, supporting fast response to faults and efficient allocation of resources in the entire network.

[0047] In the embodiment, a power supply service management system based on multi-source data is provided for executing the power supply service management method based on multi-source data as described above, comprising: a fault instance module configured to collect current fault repair work orders based on the power supply service platform, and perform deep work order text semantic analysis and standard fault instance modeling, and construct a global fault instance graph; an audio analysis module configured to identify customer real-time repair call audio streams, perform voice analysis in each time window, and perform intelligent fault work order filling, thereby obtaining real-time audio intelligent fault work orders; a resource regulation module configured to perform fault demand decomposition on the real-time audio intelligent fault work orders and the global fault instance graph, and perform dynamic configuration and regulation of power supply service resources, and construct a service resource regulation engine; a time limit early warning module configured to perform real-time processing of global fault work orders based on the service resource regulation engine, and perform overdue work order early warning, and generate a service time limit early warning strategy; a fault risk prediction module configured to perform regional fault frequency distribution analysis based on the service resource regulation engine, and perform dynamic device fault risk prediction, thereby obtaining a plurality of fault risk prediction points; an intelligent service management module configured to perform preventive work order intelligent creation based on the plurality of fault risk prediction points, and perform fault operation timing and path planning based on the service time limit early warning strategy, and construct an intelligent power supply service management model.

[0048] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, the scope of the present application being defined by the appended claims and not by the above description, therefore all variations falling within the meaning and scope of the equivalent elements of the application file are intended to be included in the present application.

[0049] The above description is merely that of specific embodiments of the application, enabling a person skilled in the art to understand or implement the application. Numerous modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the application. Therefore, the present application shall not be limited to these embodiments shown herein, but shall conform to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A power supply service management method based on multi-source data, characterized in that, The method comprises the following steps: Step S1: collecting current fault repair work orders based on a power supply service platform, and performing deep work order text semantic analysis and standard fault instance modeling to construct a global fault instance graph; Step S2: identifying customer real-time repair call audio streams, performing voice analysis of each time window, and performing intelligent fault work order filling to obtain real-time audio intelligent fault work orders; Step S3: decomposing fault requirements of real-time audio intelligent fault work orders and the global fault instance graph, and performing dynamic configuration and regulation of power supply service resources to construct a service resource regulation engine; Step S4: performing real-time processing of global fault work orders based on the service resource regulation engine, and performing overdue work order early warning to generate service timeliness early warning strategies; Step S5: performing regional fault frequency distribution analysis based on the service resource regulation engine, and performing dynamic device fault risk prediction to obtain multiple fault risk prediction points; Step S6: performing intelligent creation of preventive work orders based on the multiple fault risk prediction points, and performing fault operation timing and path planning based on the service timeliness early warning strategies to construct an intelligent power supply service management model. 2.The multi-source data based power service management method of claim 1, wherein, The specific steps of step S1 are as follows: Collecting current fault repair work orders based on a power supply service platform; Filtering abnormal redundant data of the current fault repair work orders to extract filtered and optimized fault repair work orders; Identifying key fields of the filtered and optimized fault repair work orders, and performing deep work order text semantic analysis to extract multi-dimensional work order semantic features; Extracting fault features based on the multi-dimensional work order semantic features to obtain fault device types, fault locations, fault influences, and fault user groups, and fitting to generate a multi-dimensional information matrix of fault work orders; Extracting fault reporting time stamps based on the filtered and optimized fault repair work orders; Modeling standard fault instances based on the fault reporting time stamps and the multi-dimensional information matrix of fault work orders to construct a global fault instance graph. 3.The multi-source data based power service management method of claim 1, wherein, The specific steps of step S2 are as follows: Identifying customer real-time repair call audio streams; Performing dynamic voiceprint adaptive gain on the customer real-time repair call audio streams to obtain an adaptive gain audio stream; Defining an equal-length time window, and performing sliding window decomposition on the adaptive gain audio stream to obtain a multi-time window audio stream; Performing voice analysis of each time window of the multi-time window audio stream to extract audio semantic features of each time window; Performing fault information mining based on the audio semantic features to extract real-time fault information, wherein the real-time fault information includes fault locations, fault types, fault conditions, on-site conditions, and personnel casualty conditions; Performing intelligent fault work order filling based on the real-time fault information to obtain real-time audio intelligent fault work orders.

4. The multi-source data based power service management method of claim 3, wherein, The specific steps of performing dynamic voiceprint adaptive gain on the customer real-time repair call audio streams to obtain an adaptive gain audio stream are as follows: Performing short-time Fourier transform on the customer real-time repair call audio streams to obtain an audio spectrum graph; Performing abnormal frequency mutation detection on the audio spectrum graph to mark environmental noise points; Calculating the frequency mutation amplitude of the environmental noise points; Performing high-frequency filtering and denoising on the customer real-time repair call audio streams based on the frequency mutation amplitude to obtain a high-frequency filtered and denoised audio stream; Calculate the user speech speed of the adaptive denoising audio stream, evaluate the semantic clarity, obtain multi-band voiceprint feature recognition, and obtain a multi-band voiceprint vector; Based on the multi-band voiceprint vector, dynamic voiceprint adaptive gain is performed to obtain an adaptive gain audio stream. 5.The multi-source data based power service management method of claim 1, wherein, The specific steps of step S3 are: Decompose the real-time audio intelligent fault work order and the global fault instance graph according to the fault demand to obtain real-time fault demand information of each work order; According to the real-time fault demand information, multiple sub-task sequences are generated by multi-task decomposition; According to the multiple sub-task sequences, the processing scheduling type and resource utilization are quantitatively calculated to obtain the fault processing resource portrait of each sub-task; Based on the fault processing resource portrait, the power supply service resource dynamic configuration regulation is performed to construct a service resource regulation engine. 6.The multi-source data based power service management method of claim 1, wherein, The specific steps of step S4 are: Based on the service resource regulation engine, the global fault work order is processed in real time, and the real-time state information of the fault work order is collected; According to the real-time state information of the fault work order, the fault service progress is identified to obtain the real-time service progress of each work order; The real-time fault demand information is calculated to obtain the work order service priority; According to the work order service priority, the service time limit is analyzed to generate the service time limit of each work order; Based on the service time limit, the real-time service progress is identified for the time limit exceeded work order, and the time limit exceeded work order is warned to generate a service time limit warning strategy. 7.The multi-source data based power service management method of claim 1, wherein, The specific steps of step S5 are: According to the service resource regulation engine, the fault space position and fault type of all work orders are calculated; The fault space position is divided into adjacent regions to obtain multiple fault work order regions; Current meteorological monitoring data is collected according to the multiple fault work order regions; According to the fault type and the current meteorological monitoring data, the regional meteorological-fault correlation analysis is performed on the multiple fault work order regions to obtain the regional meteorological-fault correlation characteristics; According to the fault type, the fault type frequency is counted, and the regional fault frequency distribution is analyzed according to the multiple fault work order regions to generate a regional fault frequency distribution graph; Based on the regional meteorological-fault correlation characteristics, the regional fault frequency distribution graph is used to dynamically predict the equipment fault risk to obtain multiple fault risk prediction points. 8.The multi-source data based power service management method of claim 1, wherein, The specific steps of step S6 are: According to the multiple fault risk prediction points, the preventive work order is intelligently created to obtain the fault prediction prevention work order; Based on the real-time audio intelligent fault work order, the global fault instance graph, and the fault prediction prevention work order, the service area fault information is mapped to construct a power supply network digital twin model; Based on the service time limit warning strategy and the service resource regulation engine, the power supply network digital twin model is used to plan the fault operation time sequence and path to construct an intelligent power supply service management model.

9. A multi-source data based power supply service management system, characterized by, The power supply service management method based on multi-source data is used to execute the method of claim 1, comprising: A fault instance module is used to collect current fault repair work orders based on a power supply service platform, and to perform deep work order text semantic analysis and standard fault instance modeling to construct a global fault instance graph; An audio analysis module is configured to identify a customer real-time repair call audio stream, perform speech analysis on each time window, and fill in an intelligent fault work order, so as to obtain a real-time audio intelligent fault work order; A resource regulation module is configured to decompose fault demands of the real-time audio intelligent fault work order and a global fault instance graph, dynamically configure and regulate power supply service resources, and construct a service resource regulation engine; A time limit early warning module is configured to perform real-time processing of global fault work orders based on the service resource regulation engine, perform overdue work order early warning, and generate a service time limit early warning strategy; A fault risk prediction module is configured to analyze regional fault frequency distribution and dynamically predict device fault risks based on the service resource regulation engine, so as to obtain a plurality of fault risk prediction points; An intelligent service management module is configured to intelligently create a preventive work order based on the plurality of fault risk prediction points, perform fault operation timing and path planning based on the service time limit early warning strategy, and construct an intelligent power supply service management model.

Images