A charging device intelligent operation and maintenance management method and device and medium

Abstract

The application discloses a kind of charging equipment intelligent operation and maintenance management method, device and medium.Collect equipment multimodal operation data, build causal graph by causal structure learning, solve causal residual and abnormal score, combine bayesian change point detection dynamic threshold, realize adaptive anomaly detection;After abnormality exceeds, by three intelligent agents collaborative diagnosis.Rely on physical information neural network to build lightweight digital twin, by bayesian online learning minute level self-calibration;Diagnosis result is input offline reinforcement learning to generate operation and maintenance strategy, and land after twin simulation verification.Collect actual effect index to correct model, extract fault-reason-scheme triple to expand operation and maintenance knowledge graph.The application forms operation and maintenance whole-process closed-loop optimization, with adaptive detection, cross-device collaborative diagnosis, lightweight twin capacity, can reduce operation and maintenance cost, improve equipment availability.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for charging infrastructure, and in particular to a method, device and medium for intelligent operation and maintenance management of charging equipment. Background Technology

[0002] With the rapid development of the new energy vehicle industry, the scale of charging infrastructure continues to expand, and the operation and maintenance management of charging equipment faces increasingly severe challenges. Charging equipment is exposed to complex outdoor environments for extended periods, enduring the corrosive effects of high temperatures, high humidity, salt spray, dust, and other environmental factors. Coupled with factors such as the aging of electrical components and power module losses, the failure rate remains high, seriously affecting the user charging experience and operator revenue.

[0003] In the field of operation and maintenance management of charging equipment, a series of research works have been carried out on existing technologies.

[0004] In existing charging equipment operation and maintenance management solutions, one type of technical solution collects environmental data inside the charging pile and within the station by deploying sensor modules, and sets up a multi-level alarm prediction mechanism to predict the time of failure based on the rate of change of environmental quantities. However, this type of solution relies on fixed threshold rules for fault prediction, lacks self-learning and adaptive capabilities, requires manual adjustment of thresholds for different environments and different equipment, and the alarm mechanism is essentially a linear extrapolation prediction, which is difficult to cope with complex nonlinear fault evolution processes.

[0005] Another approach involves using artificial intelligence models for fault prediction and operation and maintenance scheduling. Equipment operating parameters are input into the fault prediction model, which is then combined with an operation and maintenance resource database to determine scheduling resources. Alternatively, natural language processing can be used to process operation and maintenance text to match historical solutions. However, the training of AI models in this approach relies on high-quality historical fault data. In real-world scenarios, fault samples are sparse, positive and negative samples are imbalanced, limiting the model's generalization ability. Furthermore, there is a lack of a unified modeling framework for multi-source heterogeneous data.

[0006] Another type of technical solution utilizes digital twin technology to construct virtual models of charging equipment, enabling virtual simulation and fault prediction. However, this type of solution suffers from high computational overhead in digital twin modeling, requires significant edge computing power, and is difficult to deploy on a large scale on existing charging pile hardware; the online calibration cycle of the model is also long, resulting in response lag in scenarios with rapid fault evolution.

[0007] In addition, existing technologies have proposed multi-source heterogeneous data anomaly detection schemes, health assessment and reinforcement learning decision-making schemes, and multi-dimensional feature extraction and fusion schemes. However, the above schemes generally have the following shortcomings: anomaly detection models are mostly trained offline, making it difficult to adapt to data distribution drift caused by equipment aging and environmental changes; the degree of multi-source data fusion is low, failing to fully utilize the causal relationships and complementary characteristics between different data sources; operation and maintenance decisions lack a closed-loop optimization mechanism, with fragmented fault diagnosis, work order dispatch, and maintenance feedback; and the collaborative operation and maintenance capabilities across devices and sites are lacking, making it impossible to achieve collaborative diagnosis and resource scheduling of regional charging equipment groups. Summary of the Invention

[0008] The present invention aims to solve at least one of the technical problems existing in the prior art, and provides a method, device and medium for intelligent operation and maintenance management of charging equipment.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a method for intelligent operation and maintenance management of charging equipment, comprising the following steps: Multimodal operating data of the charging equipment is collected by a multimodal sensor array deployed locally on the charging equipment; Based on the multimodal operating data, a causal structure graph of the monitoring parameters is constructed using a causal structure learning algorithm. Based on the causal structure graph, the causal residuals and anomaly scores of the real-time data are calculated. The statistical characteristics of the data distribution are monitored online using a Bayesian change point detection algorithm. When distribution drift is detected, the anomaly threshold is automatically updated to achieve adaptive anomaly detection. When the anomaly score exceeds the anomaly threshold, an anomaly event is triggered. The anomaly event and its corresponding causal path chain are used as the detection result and handed over to the three-level intelligent agent system to perform collaborative diagnosis to determine the fault type and root cause and output the collaborative diagnosis result. A lightweight digital twin model is constructed and maintained using a physical information neural network. A minute-level model self-calibration is achieved through a Bayesian online learning method. In response to receiving the collaborative diagnosis results, the model is input into an offline reinforcement learning model to generate candidate operation and maintenance strategies. The candidate operation and maintenance strategies are then simulated and verified in the lightweight digital twin model, and the optimal strategy is selected for execution. The actual performance metrics of the operation and maintenance are collected, the policy deviation between the actual performance metrics and the expected performance metrics predicted by the simulation is calculated, and the policy deviation is fed back to the offline reinforcement learning model to trigger policy fine-tuning. At the same time, the fault-cause-solution triplet is extracted from this operation and maintenance and added to the operation and maintenance knowledge graph.

[0010] In some possible embodiments, the multimodal sensor array includes electrical parameter sensors, mechanical parameter sensors, and environmental parameter sensors. The multimodal operating data includes electrical parameters, mechanical parameters, environmental parameters, and user behavior data acquired during the charging interaction process through the main control unit and edge computing unit of the charging device. Based on the multimodal operating data, a causal structure graph of the monitoring parameters is constructed using a causal structure learning algorithm. Based on the causal structure graph, the causal residuals and anomaly scores of the real-time data are calculated. A Bayesian change point detection algorithm is used to monitor the statistical characteristics of the data distribution online. When distribution drift is detected, the anomaly threshold is automatically updated to achieve adaptive anomaly detection. Specifically, this includes: using a PC algorithm or a greedy equivalence search algorithm, based on historical multimodal operating data within a sliding time window, automatically learning the causal dependencies between each monitoring parameter, and constructing a causal structure graph in the form of a directed acyclic graph. Where the node set V represents the monitoring parameters, and the directed edge set E represents the causal relationship; for real-time collected multidimensional data Based on the set of parent nodes of each node in the causal structure graph, the expected predicted value of each monitoring parameter is calculated. Then calculate the causal residuals. ; Calculate the anomaly score based on the causal residuals. ,in An adaptive weight is dynamically adjusted based on historical anomaly detection accuracy; a Bayesian change point detection algorithm is used to estimate the moment of abrupt changes in the data distribution online, and the anomaly threshold is automatically recalculated when distribution drift is detected. When the abnormal score Exceeding the aforementioned abnormal threshold When an abnormal event is triggered, the causal path chain corresponding to the abnormal event is output.

[0011] In some possible embodiments, the three-level intelligent agent system includes an edge agent deployed locally on the charging equipment, a station-level coordination agent deployed at the charging station, and a global knowledge agent deployed in the cloud. When the anomaly score exceeds the anomaly threshold, an anomaly event is triggered. The anomaly event and its corresponding causal path chain are used as detection results and handed over to the three-level intelligent agent system for collaborative diagnosis to determine the fault type and root cause, and output the collaborative diagnosis results. Specifically, this includes: the edge agent using a personalized diagnostic model trained on local data to perform a preliminary diagnosis of the anomaly event, outputting the diagnostic results and their confidence levels; when the confidence level is lower than the preset threshold, the edge agent sends a notification to the station-level coordination agent. A collaborative diagnostic request is initiated; after receiving the collaborative diagnostic request, the station-level coordinating agent retrieves the concurrent operating data of other charging devices within the same station and performs spatial consistency verification: if multiple devices exhibit similar abnormal patterns simultaneously, it is determined to be a station-level common-cause fault; if only a single device is abnormal, it is determined to be a device-level fault; the station-level coordinating agent feeds back the collaborative diagnostic results to the edge agent, and the edge agent updates its local diagnostic confidence based on the collaborative diagnostic results; the global knowledge agent receives diagnostic records from each station-level coordinating agent, uses a federated averaging algorithm to aggregate model parameters of devices of the same type to generate a shared base model, and distributes the shared base model to each edge agent for local fine-tuning.

[0012] In some possible embodiments, the construction and maintenance of a lightweight digital twin model using a physical information neural network, and the achievement of minute-level model self-calibration through a Bayesian online learning method, specifically includes: embedding the physical equations of the charging device as a regularization term of the loss function into the neural network training process; using the physical information neural network as a simulation engine to construct a lightweight digital twin model, which maintains physical consistency with a small amount of training data; dividing the lightweight digital twin model into three levels: component-level twin sub-models, equipment-level twin models, and site-level twin models; using a Bayesian online learning method to incrementally update the posterior distribution of the parameters of the lightweight digital twin model after each round of data collection; and monitoring the prediction residuals of the lightweight digital twin model in real time, automatically triggering local retraining when the prediction residuals continuously increase beyond a preset threshold, updating only the parameters of the affected component-level twin sub-models.

[0013] In some possible embodiments, the response to receiving the collaborative diagnostic result, inputting it into an offline reinforcement learning model to generate candidate operation and maintenance strategies, and then performing simulation verification of the candidate operation and maintenance strategies in the lightweight digital twin model before selecting the optimal strategy for execution, specifically includes: inputting the collaborative diagnostic result into a pre-trained offline reinforcement learning model to generate multiple candidate operation and maintenance strategies, wherein the offline reinforcement learning model is based on a Markov decision process framework and trained using historical operation and maintenance data, and its state space includes equipment health, fault type, spare parts inventory, and personnel location, and its action space includes dispatch priority, maintenance plan selection, and spare parts scheduling plan; performing multiple rounds of simulation verification on the multiple candidate operation and maintenance strategies in the lightweight digital twin model, evaluating the fault repair time, operation and maintenance cost, and equipment availability indicators of each candidate operation and maintenance strategy under different scenarios, and selecting the comprehensively optimal candidate operation and maintenance strategy for execution.

[0014] In some possible embodiments, the actual performance indicators of the collected maintenance and operation tasks are collected, the strategy deviation between the actual performance indicators and the expected performance indicators predicted by the simulation is calculated, and the strategy deviation is fed back to the offline reinforcement learning model to trigger strategy fine-tuning. Simultaneously, fault-cause-solution triples are extracted from the current maintenance and operation tasks and added to the maintenance and operation knowledge graph. Specifically, this includes: after the maintenance and operation task is completed, automatically collecting the actual repair time, whether the repair was successful on the first attempt, and user satisfaction indicators as the actual performance indicators; comparing the actual performance indicators with the expected performance indicators predicted by the lightweight digital twin model simulation to calculate the strategy deviation; feeding the strategy deviation as a correction term for the reward signal back to the offline reinforcement learning model to trigger model parameter fine-tuning; automatically extracting the fault type, root cause analysis, and repair plan from the current maintenance and operation task, constructing the fault-cause-solution triples, and adding them to the maintenance and operation knowledge graph after manual confirmation.

[0015] Secondly, the present invention provides a digital intelligent operation and maintenance management device for charging equipment, comprising: A multimodal data acquisition module, deployed at the charging equipment end, is used to acquire multimodal operating data of the charging equipment through a multimodal sensor array; An edge computing module, deployed locally on the charging device, communicates with the multimodal data acquisition module to run a causal inference engine and an edge agent, enabling adaptive anomaly detection and local preliminary diagnosis. A station-level collaboration module is deployed at the charging station and communicates with the edge computing module. It is used to run a station-level coordination agent and a lightweight digital twin engine to realize collaborative diagnosis of multiple devices within the station and digital twin simulation. The cloud-based collaborative platform, deployed in the cloud, communicates with the station-level collaborative module and is used to run the global knowledge intelligent body, the offline reinforcement learning training engine, and the operation and maintenance knowledge graph management module to realize global model aggregation, operation and maintenance strategy optimization, and knowledge accumulation. The communication module is connected to the edge computing module, the station-level collaboration module and the cloud collaboration platform respectively, and supports adaptive switching of multiple protocols such as MQTT, HTTP / 2 and Modbus-TCP.

[0016] In some possible embodiments, the edge computing module includes: a causal inference engine, used to construct a causal structure graph of monitoring parameters based on the multimodal operating data using a causal structure learning algorithm, calculate the causal residuals and anomaly scores of real-time data based on the causal structure graph, and use a Bayesian change point detection algorithm to monitor the statistical characteristics of data distribution online, automatically updating the anomaly threshold when distribution drift is detected to achieve adaptive anomaly detection; an edge agent, used to perform preliminary diagnosis of abnormal events using a personalized diagnostic model trained on local data, and initiate a collaborative diagnosis request to the station-level collaborative module when the diagnostic confidence is lower than a preset threshold; and a lightweight model inference module, used to run a lightweight neural network model after model pruning and knowledge distillation compression to achieve real-time inference at the edge.

[0017] In some possible embodiments, the cloud-based collaborative platform includes: a federated learning engine, used to aggregate model parameters uploaded by the station-level collaborative module using a federated averaging algorithm, generate a shared base model, and distribute it to the edge computing module; an offline reinforcement learning engine, used to train an offline reinforcement learning model using historical operation and maintenance data, and generate candidate operation and maintenance strategies; and an operation and maintenance knowledge graph management module, used to store and manage fault-cause-solution triples, and support intelligent question answering based on graph retrieval.

[0018] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it can realize the intelligent operation and maintenance management method for charging equipment as described above.

[0019] The beneficial effects of the intelligent operation and maintenance management method, device, and medium for charging equipment according to embodiments of the present invention are as follows: This invention constructs a causal structure graph of monitoring parameters using a causal structure learning algorithm and calculates anomaly scores based on causal residuals, replacing the alarm method that relies on fixed thresholds in existing technologies. When data distribution drifts due to equipment aging or environmental changes, the Bayesian change point detection algorithm can identify distribution abrupt changes online and automatically update the anomaly threshold, thereby reducing false alarms and missed alarms caused by fixed thresholds being unsuitable for data distribution changes. Simultaneously, when an anomaly event is triggered, a causal path chain is output synchronously, allowing maintenance personnel to intuitively understand the evolution of the fault. Collaborative diagnosis is achieved through a three-level intelligent agent system. The edge agent uses a personalized diagnostic model trained on local data for preliminary diagnosis; the station-level coordination agent distinguishes between station-level common-cause faults and equipment-level faults through spatial consistency verification; and the global knowledge agent achieves cross-site knowledge sharing through federated learning, improving diagnostic accuracy without leaking original data. When the diagnostic confidence of the edge agent is insufficient, the station-level coordination agent retrieves concurrent operating data from other equipment within the same station for horizontal comparison, effectively distinguishing between individual equipment faults and station-level common-cause faults. A lightweight digital twin model is constructed using a physical information neural network as a simulation engine, integrating the physical information of the charging equipment. The equation, used as a regularization term in the loss function, is embedded in the neural network training process, enabling the model to maintain physical consistency even with limited training data. Layered deployment and maintenance are achieved by dividing the digital twin model into three levels: component-level, device-level, and site-level. A Bayesian online learning method is used to incrementally update model parameters after each round of data collection, shortening the model's self-calibration cycle to minutes. When the prediction residual continuously increases, only local retraining of the affected component-level twin sub-models is triggered, avoiding the computational overhead of full retraining. Operational decisions are formalized as Markov decision processes, utilizing offline reinforcement learning models. This invention learns optimal strategies from historical operation and maintenance data, avoiding potential security risks associated with online exploration. The generated candidate operation and maintenance strategies are validated through multiple rounds of simulation in a lightweight digital twin model before execution. After execution, actual performance metrics are collected and compared with the predicted performance metrics from the simulation. Strategy deviation is calculated and fed back to the offline reinforcement learning model to trigger strategy fine-tuning, forming a complete closed loop from strategy generation to execution feedback and model update. Simultaneously, fault-cause-solution triples are automatically extracted from each operation and maintenance session and added to the operation and maintenance knowledge graph, enabling continuous accumulation and reuse of operation and maintenance experience. Through the comprehensive application of these technologies, this invention can effectively reduce the operation and maintenance costs of charging equipment and improve equipment availability. Attached Figure Description

[0020] Figure 1 This is a flowchart of the intelligent operation and maintenance management method for charging equipment according to the present invention; Figure 2 This is a schematic diagram of the structure of the intelligent operation and maintenance management device for charging equipment according to the present invention. Detailed Implementation

[0021] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0023] According to an embodiment of the present invention, an embodiment of a digital intelligent operation and maintenance management method for charging equipment is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0024] Figure 1 This is a flowchart of the intelligent operation and maintenance management method for charging equipment according to the present invention. Figure 1As shown, a digital intelligent operation and maintenance management method for charging equipment S100 includes the following steps: Step S101: Collecting multimodal operation data of the charging equipment through a multimodal sensor array deployed locally on the charging equipment; Step S102: Based on the multimodal operation data, constructing a causal structure graph of monitoring parameters using a causal structure learning algorithm, calculating the causal residuals and anomaly scores of real-time data based on the causal structure graph, and using a Bayesian change point detection algorithm to monitor the statistical characteristics of data distribution online. When distribution drift is detected, the anomaly threshold is automatically updated to achieve adaptive anomaly detection; Step S103: When the anomaly score exceeds the anomaly threshold, an anomaly event is triggered. The anomaly event and its corresponding causal path chain are used as the detection result and handed over to a three-level intelligent agent system for collaborative diagnosis. To determine the fault type and root cause and output collaborative diagnostic results; Step S104: Construct and maintain a lightweight digital twin model using a physical information neural network, achieve minute-level model self-calibration through Bayesian online learning method, and in response to receiving collaborative diagnostic results, input them into an offline reinforcement learning model to generate candidate operation and maintenance strategies. After simulating and verifying the candidate operation and maintenance strategies in the lightweight digital twin model, select the optimal strategy for execution; Step S105: Collect actual performance indicators of operation and maintenance execution, calculate the strategy deviation between the actual performance indicators and the expected performance indicators predicted by simulation, and feed the strategy deviation back to the offline reinforcement learning model to trigger strategy fine-tuning. At the same time, extract the fault-cause-solution triplet from this operation and maintenance and add it to the operation and maintenance knowledge graph.

[0025] In some embodiments, the multimodal sensor array includes electrical parameter sensors, mechanical parameter sensors, and environmental parameter sensors. Multimodal operating data includes electrical parameters, mechanical parameters, environmental parameters, and user behavior data acquired during the charging interaction process through the main control unit and edge computing unit of the charging device. Based on the multimodal operating data, a causal structure graph of the monitoring parameters is constructed using a causal structure learning algorithm. Based on the causal structure graph, the causal residuals and anomaly scores of the real-time data are calculated. A Bayesian change point detection algorithm is used to monitor the statistical characteristics of the data distribution online. When distribution drift is detected, the anomaly threshold is automatically updated to achieve adaptive anomaly detection. Specifically, this includes: using a PC algorithm or a greedy equivalence search algorithm, based on historical multimodal operating data within a sliding time window, automatically learning the causal dependencies between each monitoring parameter, and constructing a causal structure graph in the form of a directed acyclic graph. Where the node set V represents the monitoring parameters, and the directed edge set E represents the causal relationship; for real-time collected multidimensional data Based on the set of parent nodes of each node in the causal structure graph, the expected value of the prediction for each monitoring parameter is calculated. Then calculate the causal residuals. Anomaly scores are calculated based on causal residuals. ,in An adaptive weight is dynamically adjusted based on historical anomaly detection accuracy; a Bayesian change point detection algorithm is used to estimate the moment of abrupt changes in the data distribution online, and the anomaly threshold is automatically recalculated when distribution drift is detected. When abnormal scoring Exceeding the abnormal threshold When an abnormal event is triggered, the causal path chain corresponding to the abnormal event is output.

[0026] Specifically, the Peter-Clark PC algorithm is a causal structure learning algorithm based on conditional independence testing, discovering causal relationships between variables by progressively removing irrelevant edges. The Greedy Equivalence Search (GES) algorithm is a causal structure learning algorithm based on score search, optimizing the score function by greedily adding and removing edges. A Directed Acyclic Graph (DAG) is a graph structure without directed cycles, used to represent the causal direction between variables. The parent node set refers to all upstream nodes in the causal structure graph that directly point to a given node. The predicted expected value is the predicted value calculated based on the variables in the parent node set using methods such as regression or neural networks. The causal residual is the difference between the actual observed value and the predicted expected value, reflecting the degree of deviation from the normal causal relationship. Adaptive weights are coefficients dynamically adjusted based on the accuracy of each monitoring parameter in historical anomaly detection; parameters with higher accuracy have larger weights. The Bayesian change point detection algorithm is an online change point detection method based on Bayesian statistical inference, identifying abrupt changes in data distribution by calculating the posterior probability of each time point being a change point. Distribution drift refers to a significant change in the statistical properties of the data generation process, such as changes in the mean, variance, or higher-order moments. A causal path chain refers to the directed path from the root cause of an anomaly to its final manifestation, used to explain the occurrence and propagation of anomalies.

[0027] In some embodiments, the three-level intelligent agent system includes an edge agent deployed locally on the charging equipment, a station-level coordination agent deployed at the charging station, and a global knowledge agent deployed in the cloud. When the anomaly score exceeds the anomaly threshold, an anomaly event is triggered. The anomaly event and its corresponding causal path chain are used as the detection result and handed over to the three-level intelligent agent system for collaborative diagnosis to determine the fault type and root cause, and output the collaborative diagnosis result. Specifically, this includes: the edge agent using a personalized diagnostic model trained on local data to perform a preliminary diagnosis of the anomaly event and outputting the diagnosis result and its confidence level; when the confidence level is lower than a preset threshold, the edge agent initiates a collaborative diagnosis with the station-level coordination agent. The station-level coordinating agent receives the collaborative diagnostic request and retrieves the concurrent operating data of other charging devices within the same station, performing spatial consistency checks: if multiple devices exhibit similar abnormal patterns simultaneously, it is determined to be a station-level common-cause fault; if only a single device is abnormal, it is determined to be a device-level fault. The station-level coordinating agent feeds back the collaborative diagnostic results to the edge agent, which updates its local diagnostic confidence based on the collaborative diagnostic results. The global knowledge agent receives diagnostic records from each station-level coordinating agent, uses a federated averaging algorithm to aggregate the model parameters of devices of the same type to generate a shared base model, and distributes the shared base model to each edge agent for local fine-tuning.

[0028] Specifically, edge agents are lightweight AI agents deployed locally on charging equipment, responsible for real-time data processing and local diagnostic inference. Personalized diagnostic models refer to fault classification or anomaly detection models independently trained using historical operating data from local devices, capable of capturing the individual characteristics of the device. Confidence level refers to the probability estimate of the diagnostic model's output, reflecting the reliability of the diagnostic results. Preset thresholds are lower confidence limits set based on operational experience or cross-validation; below this threshold, station-level information is needed for auxiliary judgment. Station-level coordination agents are AI agents deployed at the charging station, responsible for coordinating the diagnostic process of multiple charging devices within the station. Simultaneous operating data refers to the operating parameters of other charging devices within the station during the same time period, used for horizontal comparative analysis. Spatial consistency verification is a cross-validation method based on spatial dimensions, comparing the operating modes of multiple devices within the same station to determine whether a fault is an individual device problem or a station-level common problem. Station-level common-cause faults refer to similar faults occurring simultaneously on multiple devices due to station-level factors (such as grid fluctuations or environmental changes). Equipment-level faults refer to faults caused solely by the fault of a single device. Federated Averaging (FedAvg) is a distributed machine learning method where each node trains its model locally and then uploads only the model parameters to a central server. The central server performs a weighted average of all uploaded parameters to generate a global model, which is then distributed to all nodes, thus achieving collaborative learning without sharing the original data. The shared base model refers to a general diagnostic model obtained through federated learning, reflecting the common characteristics of similar devices. Local fine-tuning refers to the edge agent receiving the shared base model and using limited local data to iteratively update the model parameters to adapt them to the individual characteristics of the local device.

[0029] In some embodiments, a lightweight digital twin model is constructed and maintained using a physical information neural network, and minute-level model self-calibration is achieved through a Bayesian online learning method. Specifically, this includes: embedding the physical equations of the charging device as the regularization term of the loss function into the neural network training process; using the physical information neural network as a simulation engine to construct a lightweight digital twin model, which maintains physical consistency with a small amount of training data; dividing the lightweight digital twin model into three levels: component-level twin sub-models, device-level twin models, and site-level twin models; using a Bayesian online learning method to incrementally update the posterior distribution of parameters of the lightweight digital twin model after each round of data collection; and monitoring the prediction residuals of the lightweight digital twin model in real time. When the prediction residuals continue to increase beyond a preset threshold, local retraining is automatically triggered, updating only the parameters of the affected component-level twin sub-models.

[0030] Specifically, the Physics-Informed Neural Network (PINN) is a modeling method that embeds physical laws into the neural network loss function as a regularization term, enabling the model to still follow physical laws even with sparse data. Physical equations include, but are not limited to, mathematical expressions describing the operating mechanism of charging equipment, such as heat conduction equations, circuit equations, and mechanical equations. The regularization term in the loss function refers to adding the residuals of the physical equations as a penalty term to the standard data fitting loss, guiding the model to learn solutions that conform to physical laws. Component-level twin models are digital twin models of individual key components in the charging equipment (such as power modules, contactors, and cooling fans). They have small parameter scales and are suitable for edge deployment. Equipment-level twin models are whole-machine digital twin models composed of multiple component-level twin models, running on station-level edge servers. Station-level twin models are overall digital twin models of the charging station aggregated from multiple equipment-level twin models, running in the cloud. Bayesian online learning is an incremental learning framework based on Bayesian statistics. After acquiring a new batch of data, it uses the posterior distribution of the current parameters as the prior distribution for the next round, updating the parameter distribution by combining it with the likelihood function of the new data, thus achieving continuous learning. The posterior distribution refers to the probability distribution of the model parameters given the observed data. The prediction residual refers to the difference between the output value of the digital twin model and the actual sensor measurement value. Local retraining refers to retraining only the component-level twin models with detected prediction residual anomalies, without affecting the model parameters of other normal components, thereby reducing computational overhead.

[0031] In some embodiments, in response to receiving collaborative diagnostic results, the results are input into an offline reinforcement learning model to generate candidate operation and maintenance strategies. The optimal strategy is then selected for execution after simulation verification of the candidate operation and maintenance strategies in a lightweight digital twin model. Specifically, this includes: inputting the collaborative diagnostic results into a pre-trained offline reinforcement learning model to generate multiple candidate operation and maintenance strategies. The offline reinforcement learning model is based on a Markov decision process framework and trained using historical operation and maintenance data. Its state space includes equipment health, fault type, spare parts inventory, and personnel location; its action space includes dispatch priority, maintenance plan selection, and spare parts scheduling plan. Multiple rounds of simulation verification are performed on the multiple candidate operation and maintenance strategies in the lightweight digital twin model to evaluate the fault repair time, operation and maintenance cost, and equipment availability indicators of each candidate operation and maintenance strategy under different scenarios. The optimal candidate operation and maintenance strategy is then selected for execution.

[0032] Specifically, a Markov Decision Process (MDP) is a mathematical framework for sequential decision problems, consisting of five elements: state space, action space, state transition probabilities, reward function, and discount factor. The state space refers to the set of all relevant factors influencing operational decisions, including equipment health (e.g., remaining useful life, health score), fault type (e.g., electrical fault, mechanical fault, communication fault), spare parts inventory (e.g., available quantity of spare parts for each model), and personnel location (e.g., real-time geographical location of each operational staff member). The action space refers to the set of all action options that the operational decision-maker can execute, including dispatch priority (e.g., urgent, high, medium, low), maintenance plan selection (e.g., on-site maintenance, remote restart, spare parts replacement), and spare parts scheduling plan (e.g., from which warehouse to allocate, and by whom to deliver). Offline reinforcement learning is a method that uses only historical static datasets for policy learning, without interacting with the environment during training, thus avoiding the safety risks that online exploration may bring. The optimal mapping relationship refers to the mapping function from state to action that maximizes the expected value of long-term cumulative reward under that mapping. Candidate maintenance strategies refer to multiple possible action sequences output by the offline reinforcement learning model for the current state. Multi-round simulation verification refers to repeatedly simulating each candidate strategy in a digital twin environment to evaluate its average performance under the influence of random factors. Fault repair time refers to the total time from the generation of a maintenance work order to the complete repair of the fault. Maintenance costs include labor costs, spare parts costs, transportation costs, and downtime losses. Equipment availability refers to the proportion of time that equipment is in a serviceable state out of the total time.

[0033] In some embodiments, actual performance metrics of the operation and maintenance (O&M) execution are collected, the policy deviation between the actual performance metrics and the expected performance metrics predicted by the simulation is calculated, and the policy deviation is fed back to the offline reinforcement learning model to trigger policy fine-tuning. Simultaneously, fault-cause-solution triples are extracted from the current O&M operation and added to the O&M knowledge graph. Specifically, this includes: after the O&M task is completed, the actual repair time, whether the repair was successful on the first attempt, and user satisfaction metrics are automatically collected as actual performance metrics; the actual performance metrics are compared with the expected performance metrics predicted by the lightweight digital twin model simulation to calculate the policy deviation; the policy deviation is fed back to the offline reinforcement learning model as a correction term for the reward signal to trigger model parameter fine-tuning; fault type, root cause analysis, and repair plan are automatically extracted from the current O&M task to construct fault-cause-solution triples, which are then added to the O&M knowledge graph after manual confirmation.

[0034] Specifically, actual repair time refers to the actual time elapsed from the arrival of maintenance personnel on-site to the complete resolution of the fault. A successful repair on the first attempt means that maintenance personnel complete the fault repair on a single visit, without needing to return. User satisfaction metrics refer to the user evaluation scores for the charging service experience obtained through user feedback or APP ratings. Expected effect metrics refer to the predicted values ​​of the digital twin model for the execution effect of candidate strategies during strategy simulation verification. Policy bias refers to the measure of the difference between the actual effect metrics and the expected effect metrics, which can be an absolute difference, a relative difference, or a weighted comprehensive difference. The reward signal correction term refers to adding the policy bias as a penalty or reward to the reward function of the offline reinforcement learning model, causing the model to tend to choose strategies that are closer to the expected actual effect in subsequent decisions. Model parameter fine-tuning refers to using the policy bias signal to perform a small number of iterative updates to the parameters of the offline reinforcement learning model, shifting the model's policy distribution towards a direction that yields better actual results. The fault-cause-solution triple is a structured knowledge representation where the fault type describes the phenomenon and classification of the fault, the root cause analysis describes the fundamental cause of the fault, and the repair solution describes the specific operational steps and required spare parts to resolve the fault. An operations and maintenance knowledge graph is a domain knowledge base organized with a graph structure. Nodes represent entities (such as fault type, equipment model, repair solution, spare parts information), and edges represent semantic relationships between entities (such as causal relationships, accompanying relationships, and solution relationships), supporting intelligent question answering and reasoning based on graph retrieval.

[0035] The intelligent operation and maintenance management method for charging equipment according to embodiments of the present invention has the following advantages: This invention constructs a causal structure graph of monitoring parameters using a causal structure learning algorithm and calculates anomaly scores based on causal residuals, replacing the alarm method that relies on fixed thresholds in existing technologies. When data distribution drifts due to equipment aging or environmental changes, the Bayesian change point detection algorithm can identify distribution abrupt changes online and automatically update the anomaly threshold, thereby reducing false alarms and missed alarms caused by fixed thresholds being unsuitable for data distribution changes. Simultaneously, when an anomaly event is triggered, a causal path chain is output synchronously, allowing maintenance personnel to intuitively understand the evolution of the fault. Collaborative diagnosis is achieved through a three-level intelligent agent system. The edge agent uses a personalized diagnostic model trained on local data for preliminary diagnosis; the station-level coordination agent distinguishes between station-level common-cause faults and equipment-level faults through spatial consistency verification; and the global knowledge agent achieves cross-site knowledge sharing through federated learning, improving diagnostic accuracy without leaking original data. When the diagnostic confidence of the edge agent is insufficient, the station-level coordination agent retrieves concurrent operating data from other equipment within the same station for horizontal comparison, effectively distinguishing between individual equipment faults and station-level common-cause faults. A lightweight digital twin model is constructed using a physical information neural network as a simulation engine, integrating the physical information of the charging equipment. The equation, used as a regularization term in the loss function, is embedded in the neural network training process, enabling the model to maintain physical consistency even with limited training data. Layered deployment and maintenance are achieved by dividing the digital twin model into three levels: component-level, device-level, and site-level. A Bayesian online learning method is used to incrementally update model parameters after each round of data collection, shortening the model's self-calibration cycle to minutes. When the prediction residual continuously increases, only local retraining of the affected component-level twin sub-models is triggered, avoiding the computational overhead of full retraining. Operational decisions are formalized as Markov decision processes, utilizing offline reinforcement learning models. This invention learns optimal strategies from historical operation and maintenance data, avoiding potential security risks associated with online exploration. The generated candidate operation and maintenance strategies are validated through multiple rounds of simulation in a lightweight digital twin model before execution. After execution, actual performance metrics are collected and compared with the predicted performance metrics from the simulation. Strategy deviation is calculated and fed back to the offline reinforcement learning model to trigger strategy fine-tuning, forming a complete closed loop from strategy generation to execution feedback and model update. Simultaneously, fault-cause-solution triples are automatically extracted from each operation and maintenance session and added to the operation and maintenance knowledge graph, enabling continuous accumulation and reuse of operation and maintenance experience. Through the comprehensive application of these technologies, this invention can effectively reduce the operation and maintenance costs of charging equipment and improve equipment availability.

[0036] Figure 2 This is a schematic diagram of the intelligent operation and maintenance management device for charging equipment according to the present invention. Figure 2 As shown, a digital intelligent operation and maintenance management device for charging equipment includes: The multimodal data acquisition module 201 is deployed at the charging equipment end and is used to acquire multimodal operating data of the charging equipment through a multimodal sensor array; Edge computing module 202 is deployed locally on the charging device and communicates with the multimodal data acquisition module to run the causal inference engine and edge agent to achieve adaptive anomaly detection and local preliminary diagnosis. The station-level collaboration module 203 is deployed in the charging station and communicates with the edge computing module. It is used to run the station-level coordination agent and the lightweight digital twin engine to realize collaborative diagnosis of multiple devices in the station and digital twin simulation. The cloud-based collaborative platform 204 is deployed in the cloud and communicates with the station-level collaborative module. It is used to run the global knowledge intelligent body, the offline reinforcement learning training engine, and the operation and maintenance knowledge graph management module to realize global model aggregation, operation and maintenance strategy optimization, and knowledge accumulation. The communication module 205 communicates with the edge computing module, the station-level collaboration module, and the cloud collaboration platform, respectively, and supports adaptive switching of multiple protocols such as MQTT, HTTP / 2, and Modbus-TCP.

[0037] In some embodiments, the edge computing module includes: a causal inference engine, used to construct a causal structure graph of monitoring parameters based on multimodal running data using a causal structure learning algorithm, calculate the causal residuals and anomaly scores of real-time data based on the causal structure graph, and use a Bayesian change point detection algorithm to monitor the statistical characteristics of data distribution online, and automatically update the anomaly threshold when distribution drift is detected to achieve adaptive anomaly detection; an edge agent, used to perform preliminary diagnosis of abnormal events using a personalized diagnostic model trained on local data, and initiate a collaborative diagnosis request to the station-level collaborative module when the diagnostic confidence is lower than a preset threshold; and a lightweight model inference module, used to run a lightweight neural network model after model pruning and knowledge distillation compression to achieve real-time inference at the edge.

[0038] In some embodiments, the cloud-based collaborative platform includes: a federated learning engine, used to aggregate model parameters uploaded by station-level collaborative modules using a federated averaging algorithm, generate a shared base model, and distribute it to the edge computing module; an offline reinforcement learning engine, used to train an offline reinforcement learning model using historical operation and maintenance data, and generate candidate operation and maintenance strategies; and an operation and maintenance knowledge graph management module, used to store and manage fault-cause-solution triples, and support intelligent question answering based on graph retrieval.

[0039] The intelligent operation and maintenance management device for charging equipment according to embodiments of the present invention has the following advantages: This invention utilizes a multimodal data acquisition module to simultaneously collect electrical, mechanical, environmental, and user behavior data, providing a more comprehensive data foundation for anomaly detection. An edge computing module, deployed locally on the charging equipment, can complete adaptive anomaly detection and preliminary diagnosis within milliseconds, avoiding network latency caused by uploading all data to the cloud for processing and improving the real-time performance of fault response. A station-level collaboration module, acting as an intermediate layer, undertakes collaborative diagnosis of multiple devices within the station and station-level maintenance of the digital twin model, reducing the computational and communication bandwidth pressure on the cloud. The cloud-based collaboration platform, serving as the global decision-making center, achieves cross-site knowledge sharing through federated learning, generates operation and maintenance strategies through offline reinforcement learning, and continuously accumulates operation and maintenance experience through knowledge graphs. The communication module supports multi-protocol adaptive switching and disconnection resumption, ensuring the reliability and real-time performance of data transmission between the end-edge-cloud three-tier architecture. Through the collaborative work of these five modules, the device provided by this invention forms a complete operation and maintenance management system covering the end-edge-cloud three-tier architecture, effectively reducing the operation and maintenance costs of charging equipment and improving equipment availability.

[0040] In another aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, can implement the intelligent operation and maintenance management method for charging equipment as described above.

[0041] The computer-readable medium may be included in the apparatus, device, or system disclosed herein, or it may exist independently.

[0042] The computer-readable storage medium can be any tangible medium that contains or stores a program, and can be an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device. More specific examples include, but are not limited to, electrical connections having one or more wires, portable computer disks, hard disks, optical fibers, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0043] Specifically, in this invention, the computer-readable storage medium can be any medium capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0044] The computer-readable storage medium may also include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code, specific examples of which include, but are not limited to, electromagnetic signals, optical signals, or any suitable combination thereof.

[0045] When executed by a processor, this computer program can achieve all functions, including adaptive anomaly detection, three-level intelligent agent collaborative diagnosis, lightweight digital twin model construction and minute-level self-calibration, offline reinforcement learning strategy generation and simulation verification, operation and maintenance strategy deviation feedback and model fine-tuning, and fault-cause-solution triple extraction and knowledge graph updating. This software product, in the form of a computer-readable storage medium, can be sold and distributed independently of hardware devices. Users can load it into the edge computing node of a charging station or a cloud server to implement the intelligent operation and maintenance management method for charging equipment of this invention, which has broad application prospects and commercial value.

[0046] Parameter description: All parameter values ​​mentioned in this manual are not subjective assumptions, but rather a comprehensive result of the application scenario's security / efficiency requirements, industry standards and specifications, and industry practice experience thresholds. In actual applications, the parameters will be fine-tuned according to the relevant scenarios.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent operation and maintenance management of charging equipment, characterized in that, Includes the following steps: Multimodal operating data of the charging equipment is collected by a multimodal sensor array deployed locally on the charging equipment; Based on the multimodal operating data, a causal structure graph of the monitoring parameters is constructed using a causal structure learning algorithm. Based on the causal structure graph, the causal residuals and anomaly scores of the real-time data are calculated. The statistical characteristics of the data distribution are monitored online using a Bayesian change point detection algorithm. When distribution drift is detected, the anomaly threshold is automatically updated to achieve adaptive anomaly detection. When the anomaly score exceeds the anomaly threshold, an anomaly event is triggered. The anomaly event and its corresponding causal path chain are used as the detection result and handed over to the three-level intelligent agent system to perform collaborative diagnosis to determine the fault type and root cause and output the collaborative diagnosis result. A lightweight digital twin model is constructed and maintained using a physical information neural network. A minute-level model self-calibration is achieved through a Bayesian online learning method. In response to receiving the collaborative diagnosis results, the model is input into an offline reinforcement learning model to generate candidate operation and maintenance strategies. The candidate operation and maintenance strategies are then simulated and verified in the lightweight digital twin model, and the optimal strategy is selected for execution. The actual performance metrics of the operation and maintenance are collected, the policy deviation between the actual performance metrics and the expected performance metrics predicted by the simulation is calculated, and the policy deviation is fed back to the offline reinforcement learning model to trigger policy fine-tuning. At the same time, the fault-cause-solution triplet is extracted from this operation and maintenance and added to the operation and maintenance knowledge graph.

2. The intelligent operation and maintenance management method for charging equipment according to claim 1, characterized in that, The multimodal sensor array includes electrical parameter sensors, mechanical parameter sensors, and environmental parameter sensors. The multimodal operating data includes electrical parameters, mechanical parameters, environmental parameters, and user behavior data acquired during the charging interaction process through the main control unit and edge computing unit of the charging device. Based on the multimodal operating data, a causal structure graph of the monitoring parameters is constructed using a causal structure learning algorithm. Based on the causal structure graph, the causal residuals and anomaly scores of the real-time data are calculated. A Bayesian change point detection algorithm is used to monitor the statistical characteristics of the data distribution online. When distribution drift is detected, the anomaly threshold is automatically updated to achieve adaptive anomaly detection. Specifically, this includes: using a PC algorithm or a greedy equivalence search algorithm, based on historical multimodal operating data within a sliding time window, automatically learning the causal dependencies between each monitoring parameter, and constructing a causal structure graph in the form of a directed acyclic graph. Where the node set V represents the monitoring parameters, and the directed edge set E represents the causal relationship; for real-time collected multidimensional data Based on the set of parent nodes of each node in the causal structure graph, the expected predicted value of each monitoring parameter is calculated. Then calculate the causal residuals. ; Calculate the anomaly score based on the causal residuals. ,in An adaptive weight is dynamically adjusted based on historical anomaly detection accuracy; a Bayesian change point detection algorithm is used to estimate the moment of abrupt changes in the data distribution online, and the anomaly threshold is automatically recalculated when distribution drift is detected. When the abnormal score Exceeding the aforementioned abnormal threshold When an abnormal event is triggered, the causal path chain corresponding to the abnormal event is output.

3. The intelligent operation and maintenance management method for charging equipment according to claim 1, characterized in that, The three-tiered intelligent agent system includes edge agents deployed locally on the charging equipment, station-level coordination agents deployed at the charging station, and global knowledge agents deployed in the cloud. When the anomaly score exceeds the anomaly threshold, an anomaly event is triggered. The anomaly event and its corresponding causal path chain are used as detection results and handed over to the three-tiered intelligent agent system for collaborative diagnosis to determine the fault type and root cause, and output collaborative diagnosis results. Specifically, the edge agent uses a personalized diagnostic model trained on local data to perform a preliminary diagnosis of the anomaly event and outputs the diagnostic results and their confidence levels. When the confidence level is lower than the preset threshold, the edge agent initiates collaborative diagnosis to the station-level coordination agent. The request is processed as follows: After receiving the collaborative diagnosis request, the station-level coordinating agent retrieves the concurrent operating data of other charging devices within the same station and performs spatial consistency verification: if multiple devices exhibit similar abnormal patterns simultaneously, it is determined to be a station-level common-cause fault; if only a single device is abnormal, it is determined to be a device-level fault. The station-level coordinating agent feeds back the collaborative diagnosis results to the edge agent, which updates its local diagnostic confidence based on the collaborative diagnosis results. The global knowledge agent receives diagnostic records from each station-level coordinating agent, uses a federated averaging algorithm to aggregate model parameters of devices of the same type to generate a shared base model, and distributes the shared base model to each edge agent for local fine-tuning.

4. The intelligent operation and maintenance management method for charging equipment according to claim 1, characterized in that, The method of constructing and maintaining a lightweight digital twin model using a physical information neural network and achieving minute-level model self-calibration through a Bayesian online learning method specifically includes: embedding the physical equations of the charging device as the regularization term of the loss function into the neural network training process; using the physical information neural network as a simulation engine to construct a lightweight digital twin model, which maintains physical consistency with a small amount of training data; dividing the lightweight digital twin model into three levels: component-level twin sub-models, equipment-level twin models, and site-level twin models; using a Bayesian online learning method to incrementally update the posterior distribution of the parameters of the lightweight digital twin model after each round of data collection; and monitoring the prediction residuals of the lightweight digital twin model in real time. When the prediction residuals continue to increase beyond a preset threshold, local retraining is automatically triggered, updating only the parameters of the affected component-level twin sub-models.

5. The intelligent operation and maintenance management method for charging equipment according to claim 1, characterized in that, The response to receiving the collaborative diagnostic result involves inputting it into an offline reinforcement learning model to generate candidate operation and maintenance strategies. The optimal strategy is then selected for execution after simulation verification of the candidate strategies within the lightweight digital twin model. Specifically, this includes: inputting the collaborative diagnostic result into a pre-trained offline reinforcement learning model to generate multiple candidate operation and maintenance strategies. The offline reinforcement learning model is based on a Markov decision process framework and trained using historical operation and maintenance data. Its state space includes equipment health, fault type, spare parts inventory, and personnel location; its action space includes dispatch priority, maintenance plan selection, and spare parts scheduling plan. Multiple rounds of simulation verification are performed on the multiple candidate operation and maintenance strategies within the lightweight digital twin model to evaluate the fault repair time, operation and maintenance cost, and equipment availability indicators of each candidate strategy under different scenarios. The optimal candidate operation and maintenance strategy is then selected for execution.

6. The intelligent operation and maintenance management method for charging equipment according to claim 1, characterized in that, The process involves collecting actual performance metrics of the collected maintenance tasks, calculating the policy deviation between these actual performance metrics and the predicted performance metrics from the simulation, and feeding this policy deviation back to the offline reinforcement learning model to trigger policy fine-tuning. Simultaneously, fault-cause-solution triples are extracted from the current maintenance task and added to the maintenance knowledge graph. Specifically, this includes: automatically collecting actual repair time, success rate of repair on the first attempt, and user satisfaction metrics as actual performance metrics after the maintenance task is completed; comparing these actual performance metrics with the predicted performance metrics from the lightweight digital twin model's simulation to calculate the policy deviation; feeding this policy deviation back to the offline reinforcement learning model as a correction term for the reward signal to trigger model parameter fine-tuning; and automatically extracting fault type, root cause analysis, and repair plan from the current maintenance task to construct the fault-cause-solution triples, which are then added to the maintenance knowledge graph after manual confirmation.

7. A digital intelligent operation and maintenance management device for charging equipment, characterized in that, include: A multimodal data acquisition module, deployed at the charging equipment end, is used to acquire multimodal operating data of the charging equipment through a multimodal sensor array; An edge computing module, deployed locally on the charging device, communicates with the multimodal data acquisition module to run a causal inference engine and an edge agent, enabling adaptive anomaly detection and local preliminary diagnosis. A station-level collaboration module is deployed at the charging station and communicates with the edge computing module. It is used to run a station-level coordination agent and a lightweight digital twin engine to realize collaborative diagnosis of multiple devices within the station and digital twin simulation. The cloud-based collaborative platform, deployed in the cloud, communicates with the station-level collaborative module and is used to run the global knowledge intelligent body, the offline reinforcement learning training engine, and the operation and maintenance knowledge graph management module to realize global model aggregation, operation and maintenance strategy optimization, and knowledge accumulation. The communication module is connected to the edge computing module, the station-level collaboration module and the cloud collaboration platform respectively, and supports adaptive switching of multiple protocols such as MQTT, HTTP / 2 and Modbus-TCP.

8. The intelligent operation and maintenance management device for charging equipment according to claim 7, characterized in that, The edge computing module includes: a causal inference engine, used to construct a causal structure graph of monitoring parameters based on the multimodal operating data using a causal structure learning algorithm, calculate the causal residuals and anomaly scores of real-time data based on the causal structure graph, and use a Bayesian change point detection algorithm to monitor the statistical characteristics of data distribution online. When distribution drift is detected, the anomaly threshold is automatically updated to achieve adaptive anomaly detection; an edge agent, used to perform preliminary diagnosis of abnormal events using a personalized diagnostic model trained on local data, and to initiate a collaborative diagnosis request to the station-level collaborative module when the diagnostic confidence is lower than a preset threshold; and a lightweight model inference module, used to run a lightweight neural network model after model pruning and knowledge distillation compression to achieve real-time inference at the edge.

9. The intelligent operation and maintenance management device for charging equipment according to claim 7, characterized in that, The cloud-based collaborative platform includes: a federated learning engine, used to aggregate model parameters uploaded by the station-level collaborative module using a federated averaging algorithm, generate a shared base model, and distribute it to the edge computing module; an offline reinforcement learning engine, used to train an offline reinforcement learning model using historical operation and maintenance data, and generate candidate operation and maintenance strategies; and an operation and maintenance knowledge graph management module, used to store and manage fault-cause-solution triples, and support intelligent question answering based on graph retrieval.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the intelligent operation and maintenance management method for charging equipment as described in any one of claims 1 to 6.

Images