Intelligent digital operation and maintenance management system and method for charging pile

Abstract

The invention discloses an intelligent digital operation and maintenance management system and method for a charging pile, and relates to the technical field of new energy infrastructure operation and maintenance, and the system comprises an information obtaining module which determines a delay information circulation channel, obtains real-time discussion content, processes a discussion text through employing a word segmentation technology, and carries out the operation and maintenance of the charging pile. If the confidence coefficient of the semantic feature vector is higher than a preset threshold value, the semantic feature vector is judged to be a specific question type through state space definition, classified question labels are obtained, the historical matching module obtains features associated with a historical question library, historical question records are matched, and the matching precision is evaluated through reward function design. Determining a historical solution most similar to the current problem; according to the intelligent digital operation and maintenance management system and method for the charging pile, the accuracy of problem classification and solution matching is continuously improved, and intelligentization and high efficiency of operation and maintenance management of the charging pile are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of new energy infrastructure operation and maintenance, and particularly to an intelligent digital operation and maintenance management system and method for charging piles. Background Art

[0002] The intelligent digital operation and maintenance management of the charging pile industry is a key area for promoting the efficient operation of new energy infrastructure. Its importance lies in ensuring the stability of the charging network and the continuous optimization of the user experience. With the popularization of electric vehicles, the demand for the intelligence of operation and maintenance management systems has become increasingly prominent. Especially in terms of improving operation and maintenance efficiency and reducing fault response time, intelligent collaboration and decision support have become the core driving forces.

[0003] However, the current operation and maintenance management systems have significant limitations in collaboration and problem-solving. Most existing solutions rely on manual communication or simple work order systems, lacking real-time team collaboration mechanisms, resulting in lagged information transmission and low problem handling efficiency. In addition, the repeated exploration of known problems is widespread. The operation and maintenance team often wastes time due to the lack of an automated knowledge matching mechanism and cannot quickly reuse historical solutions. In this context, the technical implementation of real-time collaboration and problem identification has become the main challenge. First, the construction of a real-time collaboration platform needs to address the instant communication needs across regions and multiple roles to ensure the efficient flow of information in complex operation and maintenance scenarios. Second, the intelligent identification of problem types faces technical difficulties. The system needs to accurately analyze the discussion content and associate it with historical problem patterns to avoid misjudgment or missed judgment. Finally, the realization of automatically matching relevant prototypes and generating actionable guidance requires the system to not only extract problem features but also present complex knowledge to operation and maintenance personnel in a concise and intuitive manner. These technical factors have not been resolved, resulting in the operation and maintenance team often falling into an inefficient trial-and-error cycle when dealing with faults and being difficult to quickly locate and solve problems. Summary of the Invention

[0004] The purpose of the present invention is to provide an intelligent digital operation and maintenance management system and method for charging piles, which can achieve efficient instant communication, and at the same time, by intelligently identifying the types of discussion problems, automatically matching relevant prototypes and generating actionable guidance, so as to avoid repeating the exploration of the solution paths of known problems.

[0005] To achieve the above purpose, the present invention provides the following technical solution: An intelligent digital operation and maintenance management system for charging piles, comprising: An information acquisition module, which determines a delayed information transfer channel from the charging pile operation and maintenance management system, acquires real-time discussion content, and processes the discussion text using word segmentation technology to obtain a semantic feature vector; A problem identification module, if the confidence level of the semantic feature vector is higher than a preset threshold, determines it as a specific problem type through state space definition, and obtains a classified problem label; The historical matching module obtains the associated features with the historical problem library, matches the historical problem records, designs an evaluation of the matching accuracy through a reward function, and determines the historical solution most similar to the current problem; The step extraction module extracts the key operation steps, maps and associates the operation steps with the scenario labels, sorts the execution order of the emergency scenarios, and obtains a structured knowledge representation; The guidance generation module generates actionable guidance, uses templated text generation technology to convert the operation steps into executable instructions for the operation and maintenance personnel, detects and identifies logical contradictions between the instructions, and determines the final guidance text; The adaptation module integrates the complex scenario adaptation module, triggers an external data adjustment instruction content adapted to the fault code parsed according to the device status and the environmental parameters, and obtains an adapted guidance solution; The feedback processing module obtains the feedback data after execution, stores the feedback and interaction records, and adjusts and optimizes the action selection and model parameter update; The rule update module obtains the optimized classification weights and matching thresholds, uses a convergence evaluation to model the stability, and updates the dynamic rule library through the integration of the feedback data to obtain the updated scenario label mapping rules.

[0006] Preferably, the information acquisition module determines the delayed information transfer channel from the charging pile operation and maintenance management system, obtains the real-time discussion content, and processes the discussion text using word segmentation technology to obtain semantic feature vectors including: Extract cross-regional communication data from the charging pile operation and maintenance log, use distributed message queue technology to process high-concurrency requests, generate a structured instant messaging dataset. If the data volume exceeds the preset threshold, distribute the data to multiple queue nodes through a partitioning strategy to obtain a structured instant messaging dataset. For the structured instant messaging dataset, establish a real-time collaboration platform using the WebSocket protocol to achieve cross-regional multi-role two-way communication, determine a low-latency information transfer channel, monitor the channel status through the real-time collaboration platform. If it is detected that the latency exceeds the preset threshold, adjust the load balancing of the communication nodes to obtain a stable low-latency channel. Obtain the real-time discussion text from the low-latency information transfer channel, apply word segmentation technology to process the text to generate a word segmentation sequence of the text, use a pre-trained word embedding model to convert the word segmentation sequence into a word vector sequence to obtain a preliminary semantic representation of the text. Obtain temperature and humidity environmental parameter data through a sensor interface, perform normalization on the environmental parameters using standardization processing to generate an environmental feature vector. If the dimension of the environmental feature vector does not match the dimension of the semantic representation, adjust the dimension through linear transformation to obtain an aligned environmental feature vector. For the preliminary semantic representation of the text and the aligned environmental feature vector, use an attention mechanism to fuse the information of both to generate a comprehensive semantic feature vector, process the comprehensive semantic feature vector through a multi-layer perceptron to obtain the final semantic feature vector of the problem description. Extract key semantic units from the final semantic feature vector, use a clustering algorithm to group the semantic units to generate a semantic classification result of the problem description, and generate an optimized scheduling instruction for cross-regional collaboration based on the semantic classification result to determine the priority ranking of operation and maintenance tasks.

[0007] Preferably, for the problem identification module, if the confidence of the semantic feature vector is higher than the preset threshold, it is judged as a specific problem type through state space definition, and the classified problem labels obtained include: Extract the semantic feature vector from the problem description, perform feature encoding using a pre-trained deep learning network to obtain an initial semantic representation, perform classification processing on the initial semantic representation through the pre-trained deep learning network to obtain a classification confidence distribution. If the classification confidence distribution is higher than the preset threshold, use a state space model to verify the problem type to determine a preliminary problem type label. According to the preliminary problem type label, obtain an association rule from a preset knowledge base to get a label correction basis, adjust the preliminary problem type label through the label correction basis to generate a corrected problem type label, use the corrected problem type label, combined with the semantic feature vector, to construct a label semantic consistency check to judge the final problem type label, and extract the classification result from the final problem type label to generate a problem type classification output.

[0008] Preferably, the historical matching module obtains the associated features related to the historical question library, matches the historical question records, designs an evaluation function through a reward function to evaluate the matching accuracy, and determines the historical solution most similar to the current question, including: Obtain the classification label from the input question tags, generate the associated features through a feature extraction method to obtain the feature vector representation, compare the feature vector with the feature vectors recorded in the historical question library through a vector similarity calculation method to obtain the similarity score. If the similarity score is greater than the preset threshold, obtain the corresponding matching record from the historical question library to determine the candidate historical solution. Adopt a reward function, combine the classification label and the context information of the matching record, calculate the evaluation score of each candidate solution to obtain the accuracy ranking, and according to the accuracy ranking, obtain the record with the highest score from the candidate solutions to determine the most similar historical solution. By analyzing the differences between the feature vector of the most similar solution and the current question tags, adjust the feature weights to obtain the optimized solution representation, and extract the key parameters from the optimized solution representation to generate the final solution output.

[0009] Preferably, the step extraction module extracts the key operation steps, maps and associates the operation steps with the scenario labels, and sorts the execution order of the emergency scenarios to obtain the structured knowledge representation, including: Obtain the records from the historical solution database, adopt text parsing technology to extract the key operation steps to obtain the operation step set. If the operation step set contains duplicate items, generate a unique operation step list through deduplication processing to obtain the refined step set. Adopt knowledge graph technology to perform semantic matching on the refined step set and the preset scenario labels to generate the mapping relationship between the steps and the labels. According to the mapping relationship, obtain the scenario label weights corresponding to each operation step, determine whether the weights exceed the preset threshold to obtain the high-correlation label set. For the high-correlation label set, adopt a priority sorting algorithm to sort the emergency scenarios according to the instruction priority to generate the scenario priority sequence. Through structured encoding technology, associate the scenario priority sequence with the operation step set to generate the structured knowledge representation. If there are missing fields in the structured knowledge representation, fill in the missing parts through semantic completion technology to obtain the complete knowledge representation.

[0010] Preferably, the guidance generation module generates the actionable guidance, adopts the templated text generation technology to convert the operation steps into executable instructions for the operation and maintenance personnel, and detects and identifies the logical contradictions between the instructions to determine the final guidance text, including: Obtain structured knowledge, extract operation and maintenance related rules and data from a preset knowledge base to obtain a knowledge representation. Adopt templating technology to generate initial operation and maintenance instructions for the rules and data in the knowledge representation. Determine the set of executable instructions. Through conflict detection, analyze the logical relationships in the set of executable instructions. If there are contradictions, mark the relevant instructions to obtain a contradiction identifier. According to the contradiction identifier, perform rule analysis, extract the conflict points from the marked instructions, and determine the logical contradiction set. For the logical contradiction set, adopt text transformation technology to adjust the expression or order of the contradictory instructions to obtain an optimized instruction set. Through templating technology, integrate the optimized instruction set to generate the final text, and determine the operation and maintenance guidance text. If there are still contradictions in the optimized instruction set, repeat conflict detection and rule analysis to obtain the final contradiction-free guidance text.

[0011] Preferably, the adaptation module integrates a complex scenario adaptation module, and triggers an external data adjustment instruction content adapted to the fault code parsed according to the device state and environmental parameters. The obtained adapted guidance solution includes: Obtain the device state data in a complex scenario, analyze the signal characteristics of the trigger fault code, determine the initial fault state. Through the analysis process, extract the key variables from the environmental parameters to obtain the environmental impact factors associated with the fault code. If the environmental impact factor exceeds the preset threshold, adjust the instruction content according to the external data to generate a temporary optimized instruction. For the temporary optimized instruction, adopt an adaptation adjustment mechanism to integrate the fault code and environmental parameters to obtain a fine-tuning instruction set. Record the revision log of the fine-tuning instruction set, store the adjustment details, generate a dynamic version of the optimized solution. According to the dynamic version of the optimized solution, obtain the real-time feedback of the device state, judge the instruction execution effect. If the execution effect does not meet the preset standard, re-parse the fault code through the trigger mechanism to obtain an updated optimized solution.

[0012] Preferably, the feedback processing module obtains the feedback data after execution, stores the feedback and interaction records, and adjusts the optimization action selection and model parameter update, including: Obtain the feedback data from the interaction record, judge the data integrity through a preset threshold to obtain a filtered feedback data set. Adopt an experience replay mechanism to store the filtered feedback data set and action records to generate a structured replay buffer. Through an exploration strategy, analyze the action records in the replay buffer, adjust the action selection probability to obtain an optimized action distribution. Update the model parameters according to the optimized action distribution, and adopt a gradient descent algorithm for iterative training to generate an updated classification network. If the confidence level of the problem type output by the classification network is lower than the preset threshold, obtain similar problem records from the historical solutions, judge the matching parameters, and combine the matching parameters with the output of the classification network to adjust the weights of the historical solutions to obtain an accurate problem type classification result. According to the accurate problem type classification result, generate the corresponding solution parameters and store them in the replay buffer to optimize subsequent interactions.

[0013] Preferably, the feedback processing module obtains the feedback data after execution, stores the feedback and interaction records, and adjusts and optimizes the action selection and model parameter update, including: Analyze the input data through the problem classification network to obtain the classification weight and matching threshold. If the classification weight is lower than the preset threshold, adjust the network parameters to obtain the optimized classification weight. Extract features from the historical solutions, combine with the optimized classification weight to determine the matching threshold. If the deviation between the matching threshold and the features of the historical solutions exceeds the range, adjust the threshold through weighted average to obtain a stable matching threshold. Use the convergence evaluation method to process the optimized classification weight and the stable matching threshold to judge the model stability. If the convergence evaluation result shows fluctuations, adjust the parameters through iterative optimization to obtain stable model parameters. Obtain the feedback data, analyze its relevance to the dynamic rule library, update the content of the rule library, fuse the feedback data through the data integration method to generate an updated dynamic rule library. According to the updated dynamic rule library, extract the scenario label features to generate preliminary mapping rules. If the matching degree between the preliminary mapping rules and the scenario labels is insufficient, optimize through rule adjustment to obtain optimized mapping rules. Process the scenario labels through the optimized mapping rules to generate the final scenario label mapping result. Use consistency verification to verify the mapping result to obtain the final scenario label mapping rules. Extract the key parameters from the final scenario label mapping rules to update the initial configuration of the problem classification network, and adjust the network structure through parameter backpropagation to obtain an optimized classification network model.

[0014] A method for intelligent digital operation and maintenance management of charging piles, using the described intelligent digital operation and maintenance management system for charging piles, the method includes: Determine the delay information transfer channel from the charging pile operation and maintenance management system, obtain the real-time discussion content, and use the word segmentation technology to process the discussion text to obtain the semantic feature vector; If the confidence of the semantic feature vector is higher than the preset threshold, determine it as a specific problem type through state space definition to obtain the classified problem label; Obtain the associated features with the historical problem library, match the historical problem records, design a reward function to evaluate the matching accuracy, and determine the historical solution most similar to the current problem; Extract the key operation steps, map and associate the operation steps with the scenario labels, sort the execution order of the emergency scenarios to obtain the structured knowledge representation; Generate actionable guidance, use the templated text generation technology to convert the operation steps into executable instructions for the operation and maintenance personnel, detect and identify logical contradictions between the instructions to determine the final guidance text; Integrate the complex scenario adaptation module, trigger the external data adjustment instruction content adapted to the fault code parsed according to the device status and the environmental parameters to obtain the adapted guidance plan; Obtain the feedback data after execution, store the feedback and interaction records, and adjust and optimize the action selection and model parameter update; Obtain the optimized classification weights and matching thresholds, evaluate the model stability using a convergence evaluation, and update the dynamic rule library through the integration of feedback data to obtain the updated scenario label mapping rules.

[0015] As can be seen from the above technical solutions, the present invention has the following beneficial effects: The intelligent digital operation and maintenance management system and method for charging piles process high-concurrency requests through a distributed message queue, implement cross-regional multi-role two-way communication using WebSocket, and construct a low-latency information flow channel. For real-time discussion content, the present invention uses word segmentation technology and environmental parameter fusion to generate a semantic feature vector of the problem description, and classifies the problem through a deep learning network. Subsequently, the present invention matches historical solutions, extracts key operation steps, generates a structured knowledge representation, and converts it into executable instructions. The present invention also integrates a complex scenario adaptation module to dynamically adjust the instruction content according to the device status and environmental parameters. Through the experience replay mechanism and exploration strategy optimization, the present invention continuously improves the accuracy of problem classification and solution matching, realizing the intelligence and efficiency of charging pile operation and maintenance management. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a connection diagram of the system modules of the present invention; Figure 2 It is a flowchart of the method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] The technical solutions of the present invention will be clearly and completely described below with reference to the drawings in the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0018] Such as Figure 1As shown in the figure, the present invention provides a technical solution: an intelligent digital operation and maintenance management system for charging piles, including an information acquisition module, which determines a delayed information flow channel from the charging pile operation and maintenance management system, obtains real-time discussion content, processes the discussion text using word segmentation technology, and obtains a semantic feature vector; a problem recognition module, if the confidence level of the semantic feature vector is higher than a preset threshold, determines it as a specific problem type through state space definition, and obtains a classified problem label; a historical matching module, obtains associated features with the historical problem library, matches historical problem records, evaluates the matching accuracy through reward function design, and determines the historical solution most similar to the current problem; a step extraction module, extracts key operation steps, maps and associates the operation steps with scene labels, sorts the execution order of emergency scenes, and obtains a structured knowledge representation; a guidance generation module, generates actionable guidance, uses templated text generation technology to convert the operation steps into executable instructions for operation and maintenance personnel, detects and identifies logical contradictions between instructions, and determines the final guidance text; an adaptation module, integrates a complex scene adaptation module, triggers an external data adjustment instruction content adapted to the fault code and environmental parameters parsed according to the device state, and obtains an adapted guidance solution; a feedback processing module, obtains the feedback data after execution, stores the feedback and interaction records, adjusts and optimizes the action selection and model parameter update; a rule update module, obtains the optimized classification weight and matching threshold, uses a convergence evaluation model to evaluate the stability, and updates the dynamic rule library through feedback data integration, and obtains an updated scene label mapping rule.

[0019] The system first accesses the real-time data channel of the charging pile operation and maintenance management system through the information acquisition module. For the discussion information of the operation and maintenance group, it uses the word segmentation algorithm in natural language processing to extract semantic features and generate a problem feature vector that can be used for judgment. Then, the problem recognition module classifies the problem based on the confidence level of the feature vector and the state space mapping rule, and obtains a clear problem label. The system calls the historical matching module, extracts the features of the corresponding label from the historical problem library and compares them with the current problem, uses the reward function to evaluate the matching accuracy, and finally selects the historical solution closest to the current problem. Subsequently, the step extraction module extracts the key operation process from this solution, associates it with the pre-marked scene label, and automatically sorts it according to the urgency of the scene, thus forming a structured task execution sequence. This structured step passes through the guidance generation module, combines templated technology to generate operation and maintenance instructions in the form of natural language. The system synchronously performs logical consistency detection, eliminates contradictory steps, and forms a final guidance text with high consistency. On this basis, the adaptation module automatically calibrates the operation instructions according to the current device fault code and environmental parameters to achieve scene adaptation. Finally, the feedback processing module stores and analyzes the data after execution, and hands it over to the rule update module for dynamic adjustment of the classification weight and threshold, so that the system has the ability of continuous learning and optimization.

[0020] The present invention can significantly improve the intelligence level of the charging pile operation and maintenance process, realize closed-loop management from fault identification, solution matching, operation guidance, environmental adaptation to feedback optimization, and solve key problems in existing systems such as low problem identification efficiency, vague guidance operation, and lack of dynamic learning. In terms of performance, it improves the efficiency and accuracy of problem handling, reduces labor costs and equipment failure rates; in terms of scalability, the system supports operation and maintenance adaptation in multiple scenarios and multi-parameter environments, improving the stability and robustness of the system; at the same time, the rule update mechanism ensures the continuous evolution of the system's decision-making ability, and has good long-term operation and maintenance value.

[0021] For example, taking a certain brand of electric vehicle charging station in City A as an example, the station frequently has current overload alarms during the high temperature period in summer. The operation and maintenance personnel submitted a discussion on the management platform about "abnormal tripping of pile No. 13 at the station due to high temperature". The system information acquisition module captured keywords such as "high temperature" and "tripping" and judged it as a "temperature control abnormality" type of problem through semantic recognition. Then, it matched similar events and corresponding solutions from the same period last year from the historical problem library, including measures such as "reducing charging current" and "enhancing ventilation". The system automatically extracts the operation steps "adjusting the module output current to less than 20A" and "starting fan mode 2", etc., and converts them into clear operation instructions. After scene adaptation, because the ambient temperature on that day was 38°C, the system automatically recommends "extending the cooling time by 3 minutes" and outputs a complete guidance text. After the operation and maintenance personnel follow the instructions, the system records the feedback and promotes the update of the rule library to optimize the response strategy for similar scenarios in the future.

[0022] An information acquisition module determines a delayed information transfer channel from a charging pile operation and maintenance management system, obtains real-time discussion content, processes the discussion text using word segmentation technology to obtain semantic feature vectors, including extracting cross-regional communication data from charging pile operation and maintenance logs, using distributed message queue technology to process high-concurrent requests, generating a structured instant messaging dataset. If the data volume exceeds a preset threshold, the data is distributed to multiple queue nodes through a partitioning strategy to obtain a structured instant messaging dataset. For the structured instant messaging dataset, a real-time collaboration platform is established using the WebSocket protocol to achieve cross-regional multi-role two-way communication, determine a low-latency information transfer channel, monitor the channel status through the real-time collaboration platform. If it is detected that the latency exceeds the preset threshold, the load balancing of communication nodes is adjusted to obtain a stable low-latency channel. Real-time discussion text is obtained from the low-latency information transfer channel, the text is processed using word segmentation technology to generate a word segmentation sequence of the text, and a pre-trained word embedding model is used to convert the word segmentation sequence into a word vector sequence to obtain a preliminary semantic representation of the text. Environmental parameter data such as temperature and humidity is obtained through a sensor interface, and the environmental parameters are normalized using standardization processing to generate an environmental feature vector. If the dimension of the environmental feature vector does not match the dimension of the semantic representation, the dimension is adjusted through linear transformation to obtain an aligned environmental feature vector. For the preliminary semantic representation of the text and the aligned environmental feature vector, the attention mechanism is used to fuse the information of both to generate a comprehensive semantic feature vector, and the comprehensive semantic feature vector is processed through a multi-layer perceptron to obtain the final semantic feature vector of the problem description. Key semantic units are extracted from the final semantic feature vector, and a clustering algorithm is used to group the semantic units to generate a semantic classification result of the problem description. According to the semantic classification result, an optimized scheduling instruction for cross-regional collaboration is generated, and the priority ranking of operation and maintenance tasks is determined.

[0023] The working process of the present invention first extracts communication data between multiple regions from the operation and maintenance system logs of charging piles. These data sources include device alarm records, maintenance feedback, monitoring information upload, etc. When processing this communication data, in order to handle high-concurrency scenarios, the system introduces distributed message queue technology. When new data arrives, the system first reads the keywords contained in this piece of communication data, such as "high temperature", "tripping", etc. Then, the system processes these keywords through a hash function to obtain an integer value. This integer value will be used to perform a modulo operation with the total number of channels in the current system to determine which queue node this piece of data should be sent to for processing. For example, if the system has a total of 10 message processing nodes deployed, the keyword with a hash result of 27 will be sent to the processing node numbered 7. This process ensures that all data is reasonably distributed and prevents a certain node from being delayed due to excessive tasks. The system will monitor the processing time of all communication records. It records the difference between the sending time and the receiving time of each message to calculate the delay time of the communication channel. If it is detected that the delay time of a certain message exceeds the system-set threshold, such as 100 milliseconds, the system will calculate the request load situation of each current node. This load refers to the total number of messages currently being processed by a node divided by the processing capacity of the node per unit time (such as the number of messages that can be processed per second). The system will preferentially reassign tasks to the node with the smallest load value to ensure that the overall communication delay is controlled within an acceptable range. After the communication channel is stable, the system obtains the real-time discussion text content from it. These texts will first be processed by a Chinese word segmentation tool to obtain a word segmentation sequence. For example, a text of "site failure, unable to charge" may be segmented into words such as "site", "failure", "unable", "charge", etc. Each word will be converted into a high-dimensional vector through a pre-trained word embedding model (such as Word2Vec or BERT's Token Embedding). Assuming the output dimension of the model is 300 dimensions, then each word will be converted into a vector containing 300 floating-point numbers. At the same time, the system obtains the environmental parameters of the current site through connected environmental sensors, such as the temperature is 34°C and the humidity is 82%. The system inputs these values into a normalization function for normalization processing. Assuming the average temperature is 25°C and the standard deviation is 5°C, then the normalized temperature is (34 - 25) / 5 = 1.8. Similarly, the humidity is normalized in the same way. The normalized environmental vector will ultimately be expressed in a fixed format, such as [1.8, 1.4]. If the dimension of this environmental vector is not 300 (inconsistent with the word vector dimension), the system will perform dimension adjustment on it through a linear transformation module. The specific method is to multiply the original two-dimensional vector by a weight matrix with 2 rows and 300 columns and add a bias term to finally obtain a 300-dimensional vector, achieving alignment with the word vector in format. The system then inputs the word vector sequence and the environmental vector into an attention fusion module.In this module, the system calculates the degree of association between each word vector and the environment vector. The result will be used as a weight to participate in the weighted summation process of the word vectors, and finally generate a comprehensive vector representing the common semantics of the full text content and environmental factors. This comprehensive vector is fed into a multi-layer neural network. This network contains at least two layers, and each layer consists of a set of linear transformations, non-linear activation functions (such as ReLU), and possible regularization operations. The system inputs the vector into the network layer by layer and outputs a final semantic representation result. The system then extracts the semantic core part from this semantic vector, such as keywords "high temperature", "tripping", etc., and forms a set of semantic units. The system then calls a clustering algorithm (such as K-means) to automatically group according to the semantic features of these units, and classify problems with similar meanings into the same type. According to the clustering results, the system prioritizes the tasks. For example, if a category of problems belongs to the high-risk equipment failure category, the system assigns a higher priority to it and pushes the processing task to the scheduling module. At the same time, combined with the availability of maintenance personnel in each place, an optimal cross-regional collaborative task allocation plan is generated to achieve efficient and intelligent joint response between multiple places.

[0024] For example, in an electric bus system operating in a coastal city, due to the sudden increase in humidity in spring, the communication modules of some charging piles are frequently disconnected. The maintenance personnel reported "multiple charging piles in the South Fourth Road bus stop in the North Area are disconnected" in the platform. The system captured the text content in real time and extracted semantic features. After combining with the humidity data (higher than 90%) provided by the sensor, the semantic weight of the environmental interference factor was strengthened through the attention mechanism. Finally, the model identified it as a problem of "environment-induced communication interruption", automatically dispatched network engineers and environmental detection personnel to the scene, and ranked the task priority as the highest. The instruction clearly required checking the tightness of the communication module and updating the moisture-proof cover sealing strip, which greatly improved the response efficiency and repair accuracy.

[0025] For the problem recognition module, if the confidence of the semantic feature vector is higher than the preset threshold, it is judged as a specific problem type through the state space definition. The obtained classified problem labels include extracting the semantic feature vector from the problem description, using a pre-trained deep learning network for feature encoding to obtain an initial semantic representation, classifying the initial semantic representation through the pre-trained deep learning network to obtain a classification confidence distribution. If the classification confidence distribution is higher than the preset threshold, a state space model is used to verify the problem type to determine the preliminary problem type label. According to the preliminary problem type label, association rules are obtained from the preset knowledge base to get the label correction basis, and the preliminary problem type label is adjusted through the label correction basis to generate a corrected problem type label. Using the corrected problem type label, combined with the semantic feature vector, a label semantic consistency check is constructed to judge the final problem type label, and the classification result is extracted from the final problem type label to generate the problem type classification output.

[0026] The system extracts semantic feature vectors from the discussion text through a semantic encoder and inputs them into a pre-trained deep neural network for initial semantic encoding. Let the input be vector X, and this vector will be encoded through the following operations: The weight matrix of the first layer is multiplied by the input vector and added with a bias term, and after being processed by the activation function, a hidden representation is obtained, that is: the output is equal to the weight of the first layer × the input vector + the bias term). If the deep neural network adopts a three-layer structure, the final output is the initial semantic representation for subsequent classification. Subsequently, this representation is fed into a classifier module (such as a Softmax classification layer) to calculate the probability distribution of each problem category, in the form of: the probability of each category = the output value after being processed by the exponential function / the sum of the exponential values of the outputs of all categories. For example, if the system has three types of problems (communication anomaly, current anomaly, control logic failure), and its output is [0.2, 0.7, 0.1], then the confidence level of the communication anomaly is 0.2. The system sets a confidence threshold (such as 0.6). If the confidence level of a certain category exceeds this value, the classification result is determined to be credible. Then it enters the state space model verification. The state space S is composed of the set of problem states recognized by the current system, and the transition probability between each state is defined as: state transition probability = the number of times state A transitions to state B in historical samples / the total number of times state A appears. If the transition probability of the current classification label to the previous state is higher than a preset threshold (such as 0.4), then it is confirmed as a preliminary problem label. Subsequently, the system searches the knowledge base for correction rules related to this problem type according to the preliminary problem label. For example: If the "communication anomaly" problem usually co-occurs with "network module offline" or "IP conflict", the system will determine whether these semantic features or environmental evidences exist. If so, the confidence level of this label is strengthened. The label correction operation is specifically determined by the linear weighted calculation of the rule matching score and the current label confidence level: corrected confidence level = original confidence level × (1 - correction coefficient) + matching strength × correction coefficient. Among them, the correction coefficient is set empirically (such as 0.3), and the matching strength is quantified according to the rule hit degree (such as 0.8), and finally the corrected label is generated. Finally, the system performs a consistency check on the corrected label and the original semantic feature vector, and determines its logical consistency by calculating the semantic similarity (such as cosine similarity). If the similarity is higher than the set threshold (such as 0.7), then this label is confirmed as the final problem type label and the classification result is output.

[0027] For example, at a charging pile site in a certain area, the system receives a discussion text "The device network is disconnected and it is often impossible to remotely restart". The system inputs its semantic features into the model and initially classifies it as "communication anomaly". The confidence level of this classification reaches 0.72, exceeding the threshold of 0.6, so it is considered valid. The state space model detects that the previous problem was "server maintenance", and the historical probability of transitioning from "server maintenance" to "communication anomaly" is 0.45, which is higher than the transition threshold. The knowledge base finds that "network disconnection" often co-occurs with "IP conflict". The text mentions "impossible to remotely restart", and this symptom matches the co-occurrence rule. The system gives an additional confidence correction and finally determines it as "communication module disconnection" through semantic consistency verification, and outputs this classification result for subsequent module processing.

[0028] The historical matching module obtains the associated features related to the historical problem library, matches the historical problem records, evaluates the matching accuracy through the design of the reward function, and determines the historical solution most similar to the current problem. This includes obtaining the classification label from the input problem label, generating the associated features through the feature extraction method to obtain the feature vector representation, comparing the feature vector with the feature vectors of the records in the historical problem library through the vector similarity calculation method to obtain the similarity score. If the similarity score is greater than the preset threshold, the corresponding matching record is obtained from the historical problem library to determine the candidate historical solution. The reward function is used to calculate the evaluation score of each candidate solution in combination with the classification label and the context information of the matching record to obtain the accuracy ranking. According to the accuracy ranking, the record with the highest score is obtained from the candidate solutions to determine the most similar historical solution. By analyzing the differences between the feature vector of the most similar solution and the current problem label, the feature weights are adjusted to obtain the optimized solution representation, and the key parameters are extracted from the optimized solution representation to generate the final solution output.

[0029] The system extracts the classification tags of the current problem and invokes the feature extraction algorithm according to the problem type indicated by the tags. The algorithm may include keyword extraction, context encoding, TF-IDF vectorization, sentence vector embedding, etc., to convert the problem semantics into a vector form with a fixed dimension, such as a semantic feature vector containing 300 numerical values. Subsequently, the system compares the feature vector of the current problem with the feature vector of each record in the historical problem library, and calculates the similarity between the two using cosine similarity. The specific calculation method is as follows: multiply the two vectors to get the numerator part, then calculate the product of the norms of the two vectors respectively as the denominator, and finally take the ratio of the two. The closer the ratio result is to 1, the higher the similarity. For example, if the current vector is vector A and the historical record vector is vector B, the similarity is the cosine value of the angle between A and B. The system sets a similarity threshold, such as 0.75. If the similarity of a certain historical record is higher than this value, it is considered a candidate record. All candidate records are incorporated into the next step of processing, namely, the reward function evaluation. The reward function comprehensively considers whether the classification tags of the current problem match the semantic context of the historical record. The form of the reward function is: matching score = similarity score × semantic overlap degree × tag consistency weight. Among them: the similarity score is the vector cosine similarity calculated in the previous step; the semantic overlap degree is obtained by the ratio of the number of overlapping keywords in the text to the total number of keywords; the tag consistency weight is an empirically set value, such as 1 for complete consistency and 0.7 for partial consistency. The system sorts all candidate solutions in descending order according to the matching score, and selects the record with the highest score as the most similar historical solution. Thereafter, the system further analyzes the feature differences between the feature vector of the optimal record and the current problem. The system applies amplification or attenuation factors to the dimensions with large weights in the difference vector to adjust the weight distribution of the feature vector of the current problem, so as to form an optimized feature representation closer to the historical successful solution. Finally, the system extracts the parameters strongly related to the actual execution operations, such as "fault type", "processing order", "operation instruction number", etc., from the optimized feature vector, and outputs a specific execution plan for the operation and maintenance personnel.

[0030] For example, at a bus charging station in the northern part of a certain city, the system receives the problem label "severe voltage fluctuation". After vector modeling, the similarity between its feature vector and the record of "interruption caused by frequent fluctuation of charging current" in the historical database reaches 0.82, exceeding the system-set threshold of 0.75. The candidate records include 5 solutions such as "poor ground wire contact" and "current induction interference". After calculating through the reward function, the historical record number HX107 has the highest matching score because the semantic content mentions "fluctuation" and "instantaneous current increase" many times. The system further compares the parts of the processing sequence parameters extracted from HX107 that do not match the current environmental parameters, and adjusts the "power-off protection delay" value in the parameters to 3 seconds recommended by the current environment before outputting the final solution. The finally generated execution plan includes specific operation instructions such as checking the looseness of the grounding cable, increasing the power-off delay, and resetting the power management module.

[0031] The step extraction module extracts key operation steps, maps and associates the operation steps with scenario labels, sorts the execution order of emergency scenarios, and obtains a structured knowledge representation, including obtaining records from the historical solution database, using text parsing technology to extract key operation steps to obtain an operation step set. If the operation step set contains duplicate items, a unique operation step list is generated through deduplication to obtain a refined step set. Using knowledge graph technology, semantic matching is performed on the refined step set and preset scenario labels to generate a mapping relationship between steps and labels. According to the mapping relationship, the scenario label weight corresponding to each operation step is obtained, and it is judged whether the weight exceeds the preset threshold to obtain a set of highly relevant labels. For the set of highly relevant labels, a priority sorting algorithm is used to sort the emergency scenarios according to the instruction priority to generate a scenario priority sequence. Through structured encoding technology, the scenario priority sequence is associated with the operation step set to generate a structured knowledge representation. If there are missing fields in the structured knowledge representation, the missing part is filled through semantic completion technology to obtain a complete knowledge representation.

[0032] This module first retrieves records from the historical solution database that match the current problem tags. The system uses natural language processing techniques (such as dependency syntax analysis and action recognition) to parse the text, identify verb phrases with operation intentions and parameter conditions, such as "disconnect the power supply", "restart the module", etc., and form a set of operation steps. If there are steps in the set that are semantically repetitive or slightly different in expression, such as "turn off the power supply" and "power-off operation", the system uses a text similarity calculation method (such as cosine similarity of sentence vectors) to determine whether they are semantically repetitive. If the similarity is greater than the threshold (such as 0.85), only one item is retained to obtain a refined set of steps. The system then calls the knowledge graph engine to embed each operation step into the semantic vector space and match it with the preset set of scenario tags. The matching method uses the following rules: calculate the semantic relevance between each operation step and the scenario tag, and obtain a score through vector cosine similarity; according to the matching score and the weight model, obtain the matching weight of the operation step with each scenario tag; for example, the matching degree of an operation step with the "overload power-off" tag is 0.9, and with the "insufficient power" tag is 0.3. If the system sets the matching threshold to 0.7, only the scenario tags with a matching degree greater than 0.7 are regarded as highly relevant tags. For the set of highly relevant tags selected, the system uses a priority sorting algorithm to sort them. This algorithm comprehensively considers the following indicators: the danger level of the scenario corresponding to the tag; the frequency of occurrence of the step in the historical solutions; the operation time and execution cost. Finally, a scenario priority sequence is obtained. For example: "power-off protection" > "short-circuit detection" > "communication reset". Then, the system binds the above scenario sequence and the corresponding operation steps one by one through a structured encoding method to form a structured knowledge representation in the form of triples: [operation step, scenario tag, priority weight]. If an operation step lacks execution parameters or the tags are incomplete, the system calls the semantic completion model to infer the missing fields based on the existing semantic context to ensure the integrity and executability of the structured representation. For example, taking the operation and maintenance record of a charging station "High-temperature protection trigger detected, it is recommended to disconnect the main power supply and check the module cooling system" as an example, the system identifies "disconnect the main power supply" and "check the module cooling system" as key operation steps. After deduplication, the unique expression is retained. The matching degree of the step "disconnect the main power supply" with the "temperature control protection scenario" in the knowledge graph is 0.92, and the matching degree of "check the module cooling system" with the "equipment maintenance scenario" is 0.88, both higher than the threshold of 0.7. The sorting logic determines that the former is more urgent with a priority of 1, and the latter is 2. Finally, the system outputs the structured triples: [disconnect the main power supply, temperature control protection, 1], [check the module cooling system, equipment maintenance, 2]. At the same time, the "module ID" field is completed, and a complete operation task package is output for the front-end platform to call.

[0033] The guidance generation module generates actionable guidance. It uses templated text generation technology to convert operation steps into executable instructions for operation and maintenance personnel, detects and identifies logical contradictions between instructions, and determines that the final guidance text includes obtaining structured knowledge, extracting operation and maintenance-related rules and data from a preset knowledge base to obtain a knowledge representation, using templated technology to generate initial operation and maintenance instructions for the rules and data in the knowledge representation, determining a set of executable instructions, analyzing the logical relationships in the set of executable instructions through conflict detection. If there are contradictions, relevant instructions are marked to obtain a contradiction identifier. According to the contradiction identifier, rule analysis is performed, conflict points are extracted from the marked instructions to determine a logical contradiction set. For the logical contradiction set, text conversion technology is used to adjust the expression or order of the contradictory instructions to obtain an optimized instruction set. Through templated technology, the optimized instruction set is integrated to generate the final text, and the operation and maintenance guidance text is determined. If there are still contradictions in the optimized instruction set, conflict detection and rule analysis are repeated to obtain the final contradiction-free guidance text.

[0034] The system first obtains the sorted operation steps and their corresponding execution parameters from the structured knowledge representation, such as operation actions, device numbers, execution time windows, etc. These contents are input into the knowledge template engine, and the system will call the rule entries stored in the knowledge base, such as operation and maintenance specifications like "Module disassembly is prohibited before power-off" and "Hot plugging is not allowed", and combine them with the operation steps to form specific knowledge expressions. Subsequently, the system uses the templatized generation method to convert the knowledge expressions into initial operation and maintenance instructions in natural language form. For example, the template item is: "Please perform the [action] operation on the [device], time requirement: [time]." Through variable substitution technology, an initial instruction list is generated in batches to form an "executable instruction set". The system then calls the conflict detection module, which uses Boolean logic rules to analyze whether there are logical inconsistencies between instructions. For example, if one instruction requires "Turn off the main power" and another requires "Detect the voltage of the power module", there is a contradiction in the execution conditions between the two. The detection algorithm is based on the following logic: If the result variable in instruction A conflicts with the precondition of instruction B and the execution times overlap, then record this contradiction; the logical conflict score for each pair of instructions is calculated by the following method: Logical contradiction score = Condition conflict intensity × Time overlap coefficient. Among them: The condition conflict intensity is scored according to the degree of semantic opposition, set to 1 for complete opposition and 0.5 for weak conflict; the time overlap coefficient is the ratio of the intersection length of the execution time windows of the two instructions to the total window length. If the score is greater than the set threshold (such as 0.6), then this pair of instructions is marked as contradictory. The system forms a "logical contradiction set" for all contradictory instructions and calls the rule analysis module to deeply analyze the context logic and sequential relationship of these instructions. Subsequently, text conversion technology is used to modify the contradictory instructions, including: semantic layer adjustment, such as changing "Power off immediately" to "Power off after detection is completed"; sequential adjustment, such as delaying the "Power off" operation until after the "Detect module"; instruction reconstruction, such as combining two commands into "Perform power off after detecting the module". After optimization, the system re-summarizes the instruction set and calls the template engine again to output the final integrated text. If there are still logical contradictions, the next round of conflict detection and adjustment is started until a consistent, clear, and conflict-free final operation and maintenance guidance text is generated.

[0035] For example, when an alarm of "High-voltage module anomaly" appears at an electric vehicle charging station, the system generates the following operation steps based on structured knowledge: "Disconnect the main power supply", "Check the high-voltage connection line", and "Measure the voltage of the high-voltage module". The initially generated operation and maintenance instructions are: Please immediately disconnect the main power supply; Please detect the module voltage; Please check the high-voltage connection line. The system identifies a logical conflict between the operations of "power off" and "voltage detection" during conflict detection because voltage measurement cannot be performed after power off. The system adjusts the execution order according to the rules and optimizes it to: Please detect the module voltage; Please check the high-voltage connection line; Please disconnect the main power supply after the detection is completed. Finally, an operation and maintenance guidance text with clear structure and no conflicting instructions is output to ensure the continuity of operation and the safety of execution.

[0036] The adaptation module integrates a complex scenario adaptation module, triggers external data adjustment instructions adapted to the fault codes and environmental parameters parsed according to the device status, and obtains the adapted guidance plan, including obtaining device status data in complex scenarios, parsing the signal characteristics of the triggered fault codes, determining the initial fault status, extracting key variables from the environmental parameters through the parsing process, obtaining the environmental impact factors associated with the fault codes. If the environmental impact factor exceeds the preset threshold, the instruction content is adjusted according to the external data adjustment instructions to generate temporary optimization instructions. For the temporary optimization instructions, an adaptation adjustment mechanism is adopted to integrate the fault codes and environmental parameters to obtain a fine-tuning instruction set, record the revision log of the fine-tuning instruction set, store the adjustment details, generate a dynamic version of the optimization plan, obtain the real-time feedback of the device status according to the dynamic version of the optimization plan, judge the execution effect of the instructions. If the execution effect does not meet the preset standard, the fault codes are re-parsed through the trigger mechanism to obtain an updated optimization plan.

[0037] This module first receives the device status data of the charging pile, which includes voltage, current, temperature, humidity, module load rate, etc. The system uses a signal detection algorithm to identify abnormal fluctuation characteristics in the data, such as instantaneous surges and drops in current, abnormal temperature gradient changes, etc., and determines the initial fault state by comparing the feature pattern matching with the built-in fault code analysis rules. Subsequently, key variables related to the current fault type, such as ambient temperature, humidity, wind speed, etc., are extracted from the environmental parameter set, and each parameter is evaluated through the impact factor calculation function. The impact factor scoring function is defined as follows: Environmental impact score = current value minus reference benchmark value, divided by the benchmark tolerance range. If the score result is greater than 1, it means that the current environmental variable exceeds the normal tolerance range, that is, it is judged as a "key environmental impact factor". When the environmental impact factor is abnormal, the system adjusts the initial operation suggestion in combination with external data sources (such as weather API, device historical response data). For example, in a high temperature scenario, the "restart module immediately" instruction is adjusted to "restart module after cooling for 10 minutes" to generate a "temporary optimization instruction". The instruction will enter the "adaptation adjustment mechanism", which generates a fine-tuning instruction set by integrating the following two dimensions of information: the semantic vector of the fault code, which matches the processing method in the historical fault instance; the normalized vector of the environmental variable, which participates in the decision weight adjustment as a constraint factor. The decision logic of the final fine-tuning instruction is expressed as: instruction strength = fault code historical resolution rate × weight coefficient + environmental adaptability × correction factor. Among them, the weight coefficient and correction factor are system experience parameters (such as 0.6 and 0.4), which are used to balance the adaptation contribution of historical experience and the current environment. After the system generates the fine-tuning instruction set, it records the revision source, environmental influencing factors, adjustment reasons and other information of each instruction, writes them into the revision log, and forms a "dynamic version" of the optimization plan. Subsequently, the system enters the instruction effect evaluation process: real-time collection of device status feedback after the execution of the instruction, such as whether the power is restored, whether the voltage is stable, whether the alarm is lifted, etc. If the set indicators (such as "voltage stabilization time is less than 5 seconds" and "module temperature drops by more than 3°C") are not achieved, the system automatically triggers a new round of fault code analysis mechanism, and generates a new optimization plan in combination with the latest environmental parameters to form an adaptive adjustment closed loop.

[0038] For example, taking the charging piles in a highway service area as an example, there was a sudden drop in module current during the noon period, accompanied by a temperature control alarm. The system identified the fault code as "module over-temperature automatic protection", and collected the ambient temperature as high as 39°C and the humidity as 78%. The impact factor score shows that the ambient temperature exceeds the standard threshold by more than 20%. The system adjusts the initial instruction "restart the module immediately" to "wait for the temperature to drop before starting the fan and delay starting the module for 5 minutes", and writes the adjustment to the revision log. After execution, real-time feedback data shows that the module temperature dropped by 4.2°C and the current returned to the stable range. The system records the version number, adjustment source and effect evaluation of this round of optimization plan for subsequent reference.

[0039] The feedback processing module obtains the feedback data after execution, stores the feedback and interaction records, and adjusts and optimizes the action selection and model parameter update. This includes obtaining the feedback data from the interaction records, judging the data integrity through a preset threshold to obtain a filtered feedback data set, using the experience replay mechanism to store the filtered feedback data set and action records, generating a structured replay buffer, analyzing the action records in the replay buffer through an exploration strategy, adjusting the action selection probability to obtain an optimized action distribution, updating the model parameters according to the optimized action distribution, and using the gradient descent algorithm for iterative training to generate an updated classification network. If the confidence level of the problem type output by the classification network is lower than the preset threshold, similar problem records are obtained from the historical solutions, the matching parameters are judged, and the weights of the historical solutions are adjusted by combining the matching parameters and the output of the classification network to obtain an accurate problem type classification result. According to the accurate problem type classification result, the corresponding solution parameters are generated and stored in the replay buffer to optimize subsequent interactions.

[0040] After the user finishes executing the guidance instruction, the feedback processing module collects the corresponding execution feedback information and operation records. The information includes: operation response time, whether the fault is resolved, changes in device status parameters, etc. First, the system filters the feedback data according to the set integrity threshold (such as at least three types of keyword fields are required), and retains the valid data to form a "filtered feedback data set". This data set corresponds one-to-one with the operation steps executed by the operation and maintenance personnel and is jointly stored in the "experience replay buffer". The form of each record is: state s, action a, feedback r, next state s'. Subsequently, the system performs action analysis through the ε-greedy strategy, that is: the system explores new operation actions with a certain probability (ε value, such as 0.1), and selects the current optimal action with the remaining probability. The action selection probability Pi(t) is adjusted according to the feedback reward function, and its update method is: P i (t + 1)=P i (t)+α×[r - P i (t)]. Where: P i(t) is the selection probability of action i at time t; r is the feedback score corresponding to the action in the current replay record; α is the learning rate (such as 0.05), representing the speed of action probability adjustment. The action distribution calculated by this strategy is used to optimize the policy network or the policy layer of the classification model. The system takes the new optimized action distribution as the target and uses the gradient descent algorithm to adjust the parameters of the current classification network. The loss function is the cross-entropy error between the predicted classification output and the true feedback, and the model weights are updated through the backpropagation of the error in each round. If after the latest round of training, the output confidence of the network for a certain type of problem is lower than the threshold (such as 0.5), it indicates that the model has not fully learned the characteristics of this type of problem. At this time, the system obtains case records similar to the current problem feature vector from the historical solution database, and compares and comprehensively scores the matching parameters (such as feature vector similarity, environmental variable matching degree, etc.). The system adjusts the weight of the historical solution according to the following method: Solution weight = classification network confidence × β + matching parameter score × (1 - β). Where: β is the confidence ratio adjustment parameter (such as 0.6), used to balance the model output and historical experience; the matching parameter score comes from the weighted average of semantic similarity and scene matching degree. The finally output new classification result is considered to be more accurate, and the system records the corresponding solution parameters and rewrites them into the experience replay buffer for the next round of interactive optimization.

[0041] For example, at a charging station in a high-humidity area in the south, after the operation and maintenance personnel performed the operation "restart the voltage module" recommended by the system, the feedback was "restart failed, the module has no response". The system recorded the "operation failed" status and the current environmental parameters (humidity 92%, temperature 36°C), and determined that the feedback was valid. This feedback was written into the experience replay buffer. By reviewing other similar scenario operation records in the replay, the system found that the success rate of the "power off first and then warm start" solution in similar scenarios was 78%, while the confidence of the current model in this solution was only 0.46. Therefore, the system introduced the historical record HX115 with a matching similarity of 0.83, and after adjustment, re-output the recommended solution: "Disconnect the main power supply and then restart the module after a 5-second delay". The system used this solution as the new recommended operation and recorded its adjustment process for subsequent learning.

[0042] Feedback processing module, which obtains the feedback data after execution, stores the feedback and interaction records, and adjusts and optimizes the action selection and model parameter update. This includes analyzing the input data through a problem classification network to obtain classification weights and matching thresholds. If the classification weight is lower than the preset threshold, the network parameters are adjusted to obtain optimized classification weights. Features are extracted from historical solutions, combined with the optimized classification weights to determine the matching threshold. If the deviation between the matching threshold and the features of the historical solution exceeds the range, the threshold is adjusted by weighted average to obtain a stable matching threshold. A convergence evaluation method is used to process the optimized classification weights and the stable matching threshold to judge the model stability. If the convergence evaluation result shows fluctuations, the parameters are adjusted through iterative optimization to obtain stable model parameters. The feedback data is obtained, its relevance to the dynamic rule base is analyzed, and the rule base content is updated. The feedback data is fused through a data integration method to generate an updated dynamic rule base. According to the updated dynamic rule base, scene label features are extracted to generate preliminary mapping rules. If the matching degree between the preliminary mapping rules and the scene labels is insufficient, the rules are adjusted and optimized to obtain optimized mapping rules. The scene labels are processed through the optimized mapping rules to generate the final scene label mapping result. The mapping result is verified by consistency check to obtain the final scene label mapping rules. Key parameters are extracted from the final scene label mapping rules to update the initial configuration of the problem classification network, and the network structure is adjusted through parameter backpropagation to obtain an optimized classification network model.

[0043] After receiving the operation and maintenance problem text or feature vector, the system first performs forward propagation through the classification network, outputs the probability distribution of each problem type, and extracts the highest weight value as the classification weight. If this weight value is lower than the set confidence threshold (such as 0.65), it is considered that the current network has insufficient confidence in classifying this problem. The system calls the network adjustment function, calculates the gradient based on the current classification error, and updates the neural network parameters (such as the weight matrix and bias term) to obtain a more reliable optimized classification weight. This process uses the standard backpropagation and gradient descent algorithm, and the update rule is: weight update = current weight - learning rate × current gradient. At the same time, the system extracts the feature distribution associated with the current problem label from the historical solution library, compares the Euclidean distance or cosine similarity between the current input feature vector and the historical feature vector, calculates its matching degree, and derives the current matching threshold. If the deviation of this matching threshold from the historical average is greater than the set range (such as 10%), the system adjusts it by weighted average: stable matching threshold = current threshold × α + historical mean × (1 - α). Here, α is the balance coefficient, and the recommended value is 0.4. To ensure the stability of the model parameters, the system calls the convergence evaluation function after each round of update. This function counts the change rate of the classification weight in the recent N times (such as N = 5). If the fluctuation is greater than the set standard deviation (such as σ > 0.1), it triggers another optimization iteration. This continues until the model output fluctuation converges. In the feedback processing link, the system structurally compares the collected execution feedback and interaction data to analyze whether it matches the rule patterns in the current dynamic rule library. If they are inconsistent, the system will execute the rule update algorithm, merge the historical rules based on the new data, and generate an updated dynamic rule library. According to the rule library, the system constructs a preliminary "scenario label - operation suggestion" mapping relationship, compares the semantic features of the known labels, and calculates the semantic matching degree (such as based on vector cosine similarity). If the matching degree is insufficient (such as less than 0.7), the system will adjust the rule content or add auxiliary conditions to generate a more accurate optimized mapping rule. Finally, the optimized mapping rule is used to regenerate the scenario label results, and the system performs a consistency check on it (such as rule application closed-loop detection) to ensure that there are no contradictions in terms of semantics, structure, and logic. This final mapping structure will extract key parameters, such as labels, context conditions, priority factors, etc., and send them back to the classification network to update the model structure (such as adjusting the number of hidden layer nodes, activation function configuration, etc.), thereby realizing the dynamic evolution of the classification model structure.

[0044] For example, an operation platform receives an exception report: "The charging pile No. 9 at the site fails to start and the screen has no display". The system classification network outputs a weight of "start-up exception" as 0.58, which is lower than the threshold of 0.65, triggering model optimization. The historical solutions show that such problems are often caused by "damage to the power control module", and the feature similarity is 0.73, which is lower than the average of 0.81. The system adjusts the matching threshold to 0.76, and uses the feedback "return to normal after replacing the power module" as a new input to correct the matching rule. Subsequently, a new label "power control failure" is remapped and generated, and the optimization result is used to update the classification network structure. The recognition accuracy of subsequent similar events is increased to 0.84, and the stability of the classification output is significantly enhanced.

[0045] As Figure 2 shown, a method for intelligent digital operation and maintenance management of charging piles is also provided. Using the described intelligent digital operation and maintenance management system for charging piles, the method includes determining a delay information transfer channel from the charging pile operation and maintenance management system, obtaining real-time discussion content, processing the discussion text using word segmentation technology to obtain semantic feature vectors; if the confidence level of the semantic feature vectors is higher than a preset threshold, determining it as a specific problem type through state space definition to obtain a classified problem label; obtaining associated features with the historical problem library, matching historical problem records, evaluating the matching accuracy through reward function design, and determining the historical solution most similar to the current problem; extracting key operation steps, mapping and associating the operation steps with scenario labels, sorting the execution order of emergency scenarios to obtain a structured knowledge representation; generating actionable guidance, converting the operation steps into executable instructions for operation and maintenance personnel using templated text generation technology, detecting and identifying logical contradictions between the instructions to determine the final guidance text; integrating a complex scenario adaptation module, triggering an external data adjustment instruction content adapted to the fault codes parsed according to the device status and environmental parameters to obtain an adapted guidance plan; obtaining feedback data after execution, storing feedback and interaction records, adjusting and optimizing action selection and model parameter updates; obtaining optimized classification weights and matching thresholds, evaluating the model stability using a convergence evaluation model, and updating the dynamic rule library through feedback data integration to obtain an updated scenario label mapping rule.

[0046] The method first obtains real-time discussion text from the charging pile operation and maintenance management system, converts it into a semantic feature vector through word segmentation and semantic embedding models, and determines whether its confidence is higher than the set threshold; if the condition is met, the state space model is combined to determine the type of problem and generate a problem label. The system then searches for similar cases in the historical problem library based on the label and semantic vector, calculates the semantic similarity and label consistency score, and evaluates the matching accuracy through the reward function, and selects the most similar historical solution. The operation steps in the solution are extracted, and after semantic deduplication, they are matched with the scene label for knowledge graph matching, and a structured operation sequence is generated in combination with the scene priority. Through the templated text generation module, the system converts the operation steps into natural language instructions, performs logical consistency detection, and performs semantic reconstruction or sequence adjustment after contradictions are found to form a conflict-free final guidance text. At the same time, the adaptation module analyzes whether the current device status and environmental variables, such as temperature and humidity, have an impact on the operation. If so, the instruction content is dynamically adjusted to generate a "fine-tuning instruction set" that integrates device features and environmental factors. After the operation and maintenance personnel execute the instructions, the system collects feedback information and records data such as the operation success rate and response time, and updates the action selection probability and classification model parameters through reinforcement learning methods. If the confidence of the classification result is insufficient or the output fluctuates greatly, the system will introduce historical solutions to perform weighted parameter corrections and iteratively optimize the model until convergence. Finally, the system adjusts the scene label mapping rules based on the feedback data, obtains the final rule structure through consistency verification, and updates the initial configuration of the problem classification model to achieve continuous learning and dynamic optimization of the operation and maintenance system.

[0047] For example, in the operation center of an inland city, a discussion message: "Pile No. 8 trips frequently and restart is invalid" was captured by the system and a semantic vector was generated. Matching the historical case "The module was restarted without cooling after overload", the system recommended the solution of "cooling down for 10 minutes after power failure and then restarting". The templated instruction is: "Please wait for 10 minutes after power failure and then restart the power module". Since the current outdoor temperature is higher than 35℃, the system adaptation module automatically adjusts the cooling time to 15 minutes. After execution, the feedback shows that the trip is released. The system updates the rule base based on this record and adjusts the "module overload in high temperature scenario" mapping strategy to enhance the response efficiency of similar scenarios next time.

[0048] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent digital operation and maintenance management system for charging piles, characterized in that: include: The information acquisition module determines the delayed information flow channel from the charging pile operation and maintenance management system, obtains the real-time discussion content, and processes the discussion text using word segmentation technology to obtain a semantic feature vector; The question recognition module, if the confidence of the semantic feature vector is higher than the preset threshold, determines it as a specific question type through the state space definition and obtains the classified question label; The historical matching module obtains features associated with the historical problem library, matches historical problem records, evaluates matching accuracy through reward function design, and determines the historical solution that is most similar to the current problem; The step extraction module extracts key operation steps, associates the operation steps with the scenario label mapping, sorts the execution order of emergency scenarios, and obtains structured knowledge representation; The guidance generation module generates actionable guidance, uses templated text generation technology to convert operation steps into executable instructions for operation and maintenance personnel, detects and identifies logical contradictions between instructions, and determines the final guidance text; The adaptation module integrates complex scenario adaptation modules, triggers external data adjustment instruction content based on the fault code parsed by the equipment status and the environmental parameters, and obtains the guidance plan after adaptation; Feedback processing module, which obtains feedback data after execution, stores feedback and interaction records, and adjusts and optimizes action selection and model parameter updates; The rule updating module obtains the optimized classification weights and matching thresholds, uses convergence to evaluate the model stability, and updates the dynamic rule base through feedback data integration to obtain updated scene label mapping rules.

2. The intelligent digital operation and maintenance management system for charging piles according to claim 1, characterized in that: The information acquisition module determines the delayed information flow channel from the charging pile operation and maintenance management system, obtains the real-time discussion content, and processes the discussion text using word segmentation technology to obtain the semantic feature vector including: Extract cross-regional communication data from the charging pile operation and maintenance log, use distributed message queue technology to process high-concurrency requests, generate a structured instant messaging data set, and if the data volume exceeds the preset threshold, distribute the data to multiple queue nodes through the partitioning strategy to obtain a structured instant messaging data set. For the structured instant messaging data set, use the WebSocket protocol to establish a real-time collaboration platform to achieve cross-regional multi-role two-way communication, determine a low-latency information flow channel, and monitor the channel status through the real-time collaboration platform. If it is detected that the delay exceeds the preset threshold, adjust the load balancing of the communication node to obtain a stable low-latency channel. Obtain real-time discussion text from the low-latency information flow channel, apply word segmentation technology to process the text, generate a word segmentation sequence of the text, and use a pre-trained word embedding model to convert the word segmentation sequence into a word vector sequence. , get the preliminary semantic representation of the text, obtain the temperature and humidity environmental parameter data through the sensor interface, normalize the environmental parameters by standardization processing, and generate an environmental feature vector. If the dimension of the environmental feature vector does not match the dimension of the semantic representation, adjust the dimension by linear transformation to obtain an aligned environmental feature vector. For the preliminary semantic representation of the text and the aligned environmental feature vector, use the attention mechanism to fuse the information of the two to generate a comprehensive semantic feature vector. The comprehensive semantic feature vector is processed by a multi-layer perceptron to obtain the final semantic feature vector of the problem description. The key semantic units are extracted from the final semantic feature vector, and the semantic units are grouped by a clustering algorithm to generate a semantic classification result of the problem description. According to the semantic classification result, an optimized scheduling instruction for cross-regional collaboration is generated to determine the priority of operation and maintenance tasks.

3. The intelligent digital operation and maintenance management system for charging piles according to claim 1 is characterized by: The problem identification module, if the confidence of the semantic feature vector is higher than a preset threshold, determines it as a specific problem type through the state space definition, and obtains the classified problem label including: A semantic feature vector is extracted from the problem description, and a pre-trained deep learning network is used for feature encoding to obtain an initial semantic representation. The initial semantic representation is classified through the pre-trained deep learning network to obtain the classification confidence distribution. If the classification confidence distribution is higher than a preset threshold, the state space model is used to verify the problem type and determine the preliminary problem type label. According to the preliminary problem type label, association rules are obtained from a preset knowledge base to obtain a label correction basis. The preliminary problem type label is adjusted according to the label correction basis to generate a corrected problem type label. The corrected problem type label is used in combination with the semantic feature vector to construct a label semantic consistency check to determine the final problem type label. The classification result is extracted from the final problem type label to generate a problem type classification output.

4. The intelligent digital operation and maintenance management system for charging piles according to claim 1, characterized in that: The history matching module obtains features associated with the history problem library, matches the history problem records, evaluates the matching accuracy through reward function design, and determines the most similar history solution to the current problem, including: The classification label is obtained from the input question label, and the associated features are generated through the feature extraction method to obtain the feature vector representation. The feature vector is compared with the feature vector recorded in the historical question library through the vector similarity calculation method to obtain the similarity score. If the similarity score is greater than the preset threshold, the corresponding matching record is obtained from the historical question library to determine the candidate historical solutions. The reward function is used to combine the classification label and the context information of the matching record to calculate the evaluation score of each candidate solution to obtain the precision ranking. According to the precision ranking, the record with the highest score is obtained from the candidate solutions to determine the most similar historical solution. By analyzing the difference between the feature vector of the most similar solution and the current question label, the feature weight is adjusted to obtain the optimized solution representation. The key parameters are extracted from the optimized solution representation to generate the final solution output.

5. The intelligent digital operation and maintenance management system for charging piles according to claim 1 is characterized by: The step extraction module extracts key operation steps, associates the operation steps with scenario label mapping, sorts the execution order of emergency scenarios, and obtains structured knowledge representation including: Records are obtained from the historical solution database, and key operation steps are extracted using text parsing technology to obtain an operation step set. If the operation step set contains duplicates, a unique operation step list is generated through deduplication processing to obtain a simplified step set. Knowledge graph technology is used to perform semantic matching between the simplified step set and the preset scenario labels to generate a mapping relationship between steps and labels. Based on the mapping relationship, the scenario label weight corresponding to each operation step is obtained to determine whether the weight exceeds the preset threshold to obtain a high-correlation label set. For the high-correlation label set, a priority sorting algorithm is used to sort emergency scenarios according to the instruction priority to generate a scenario priority sequence. Through structured coding technology, the scenario priority sequence is associated with the operation step set to generate a structured knowledge representation. If there are missing fields in the structured knowledge representation, the missing parts are filled through semantic completion technology to obtain a complete knowledge representation.

6. The intelligent digital operation and maintenance management system for charging piles according to claim 1, characterized in that: The guidance generation module generates actionable guidance, uses templated text generation technology to convert operation steps into executable instructions for operation and maintenance personnel, detects and identifies logical contradictions between instructions, and determines the final guidance text, including: Acquire structured knowledge, extract operation and maintenance related rules and data from the preset knowledge base to obtain knowledge representation, use template technology to generate initial operation and maintenance instructions for the rules and data in the knowledge representation, determine the executable instruction set, and analyze the logical relationship in the executable instruction set through conflict detection. If there is a contradiction, mark the relevant instructions to obtain the contradiction mark. According to the contradiction mark, perform rule analysis, extract conflict points from the marked instructions, and determine the logical contradiction set. For the logical contradiction set, use text conversion technology to adjust the expression or order of the contradictory instructions to obtain the optimized instruction set. Through template technology, integrate the optimized instruction set, generate the final text, and determine the operation and maintenance guidance text. If there are still contradictions in the optimized instruction set, repeat the conflict detection and rule analysis to obtain the final guidance text without contradictions.

7. The intelligent digital operation and maintenance management system for charging piles according to claim 1, characterized in that: The adaptation module integrates a complex scenario adaptation module, triggers the external data adjustment instruction content adapted according to the fault code and environmental parameters analyzed by the device status, and obtains the guidance scheme after adaptation, including: Acquire equipment status data in complex scenarios, analyze the signal characteristics that trigger fault codes, determine the initial fault state, extract key variables from environmental parameters through the analysis process, and obtain environmental impact factors associated with the fault code. If the environmental impact factor exceeds the preset threshold, adjust the instruction content according to external data and generate temporary optimization instructions. For temporary optimization instructions, adopt an adaptive adjustment mechanism to integrate fault codes and environmental parameters to obtain a fine-tuning instruction set, record a revision log of the fine-tuning instruction set, store adjustment details, and generate a dynamic version of the optimization plan. According to the dynamic version of the optimization plan, obtain real-time feedback on the equipment status and judge the execution effect of the instruction. If the execution effect does not meet the preset standard, re-analyze the fault code through the trigger mechanism to obtain an updated optimization plan.

8. The intelligent digital operation and maintenance management system for charging piles according to claim 1, characterized in that: The feedback processing module obtains feedback data after execution, stores feedback and interaction records, and adjusts and optimizes action selection and model parameter update, including: Get feedback data from the interaction records, judge the data integrity by the preset threshold, get the filtered feedback data set, use the experience replay mechanism to store the filtered feedback data set and action records, generate a structured replay buffer, analyze the action records in the replay buffer by exploration strategy, adjust the action selection probability, get the optimized action distribution, update the model parameters according to the optimized action distribution, use the gradient descent algorithm to iteratively train, and generate an updated classification network. If the confidence of the problem type output by the classification network is lower than the preset threshold, get similar problem records from historical solutions, judge the matching parameters, combine the matching parameters with the classification network output, adjust the weights of the historical solutions, and get accurate problem type classification results. According to the accurate problem type classification results, generate corresponding solution parameters and store them in the replay buffer to optimize subsequent interactions.

9. The intelligent digital operation and maintenance management system for charging piles according to claim 1, characterized in that: The feedback processing module obtains feedback data after execution, stores feedback and interaction records, and adjusts and optimizes action selection and model parameter update, including: Analyze the input data through the problem classification network to obtain the classification weight and matching threshold. If the classification weight is lower than the preset threshold, adjust the network parameters to obtain the optimized classification weight, extract features from the historical solution, and determine the matching threshold in combination with the optimized classification weight. If the matching threshold and the feature deviation of the historical solution are out of range, adjust the threshold by weighted average to obtain a stable matching threshold. Use the convergence evaluation method to process the optimized classification weight and stable matching threshold to determine the model stability. If the convergence evaluation result shows fluctuations, adjust the parameters by iterative optimization to obtain stable model parameters, obtain feedback data, and analyze its relationship with the dynamic rule base. Connectivity, update the rule base content, fuse the feedback data through data integration methods, generate an updated dynamic rule base, extract scene label features according to the updated dynamic rule base, generate preliminary mapping rules, if the preliminary mapping rules do not match the scene labels enough, obtain optimized mapping rules through rule adjustment and optimization, process scene labels through the optimized mapping rules, generate the final scene label mapping results, verify the mapping results through consistency check, obtain the final scene label mapping rules, extract key parameters from the final scene label mapping rules, update the initial configuration of the problem classification network, adjust the network structure through parameter feedback, and obtain the optimized classification network model.

10. A charging pile intelligent digital operation and maintenance management method, using a charging pile intelligent digital operation and maintenance management system according to any one of claims 1 to 9, characterized in that: The method comprises: From the charging pile operation and maintenance management system, determine the delayed information flow channel, obtain the real-time discussion content, use word segmentation technology to process the discussion text, and obtain the semantic feature vector; If the confidence of the semantic feature vector is higher than the preset threshold, it is judged as a specific question type through the state space definition and the classified question label is obtained; Obtain features associated with the historical problem database, match historical problem records, evaluate matching accuracy through reward function design, and determine the historical solution most similar to the current problem; Extract key operation steps, associate the operation steps with scenario label mapping, sort the execution order of emergency scenarios, and obtain structured knowledge representation; Generate actionable guidance, use templated text generation technology to convert operation steps into executable instructions for operation and maintenance personnel, detect and identify logical contradictions between instructions, and determine the final guidance text; Integrate complex scenario adaptation modules to trigger external data adjustment instruction content based on the fault code analyzed by the equipment status and the adaptation of environmental parameters, and obtain guidance solutions after adaptation; Obtain feedback data after execution, store feedback and interaction records, and adjust and optimize action selection and model parameter updates; The optimized classification weights and matching thresholds are obtained, the convergence is used to evaluate the model stability, the dynamic rule base is updated through feedback data integration, and the updated scene label mapping rules are obtained.

Images