A method and system for repairing charging pile faults

Abstract

This invention provides a method and system for repairing charging pile faults in the field of charging pile operation and maintenance technology. The method includes: Step S1, collecting a large amount of historical operating status data to construct a dataset; Step S2, creating a charging pile fault diagnosis model; Step S3, dividing the dataset into a training set, a validation set, and a test set to sequentially train, validate, and test the charging pile fault diagnosis model; Step S4, compressing the tested charging pile fault diagnosis model and deploying it to the charging pile; Step S5, collecting real-time operating status data during charging pile operation, preprocessing the real-time operating status data, and inputting it into the charging pile fault diagnosis model to obtain fault diagnosis results; Step S6, automatically performing fault repair operations based on the fault diagnosis results and recording the fault repair results. The advantage of this invention is that it greatly improves the intelligence level and overall efficiency of charging pile fault repair.

Description

Technical Field

[0001] This invention relates to the field of charging pile operation and maintenance technology, and in particular to a charging pile fault repair method and system. Background Technology

[0002] With the rapid development of the global electric vehicle industry, charging piles, as a key infrastructure for electric vehicle energy replenishment, directly impact the operational efficiency of the entire transportation system through their operational reliability and stability. Charging piles are complex devices integrating power electronics technology, embedded systems, communication protocols, and energy storage management functions. Often deployed outdoors or under high-load conditions, they endure multiple challenges such as voltage fluctuations, temperature changes, mechanical wear, and cybersecurity threats, making them prone to hardware failures (e.g., power module damage, connector aging) or software failures (e.g., firmware anomalies, data errors), leading to charging interruptions, safety incidents, or service quality degradation. Statistics show that under harsh operating conditions, the average failure rate of charging piles can reach 5%–10%, severely restricting user experience and increasing operation and maintenance costs. Therefore, developing efficient and intelligent fault repair methods to achieve rapid self-healing, minimize downtime, and reduce manual intervention has become an urgent industry need.

[0003] Currently, fault repair of charging piles largely relies on traditional methods, such as threshold-based monitoring mechanisms, periodic maintenance, or remote manual diagnosis. While these methods have some effectiveness, they suffer from significant shortcomings in terms of real-time performance, adaptability, and resource utilization efficiency. Especially with the increasing intelligence of charging piles, their internal storage systems (such as Flash memory), which store firmware, configuration parameters, and operational logs, have become high-risk areas for faults. Flash memory, limited by its physical characteristics (such as limited write cycles and bad block accumulation), is prone to data corruption or partitioning errors. Existing repair mechanisms are typically rather crude, such as complete erasure or component replacement, lacking refined dynamic management capabilities. Furthermore, fault prediction often employs rule-based methods or traditional neural networks, which struggle to effectively capture temporal anomalies (such as instantaneous overcurrent or communication delays), easily leading to false alarms or missed alarms. Repair strategies also generally lack flexibility, failing to dynamically adjust according to real-time operating conditions, resulting in low repair efficiency.

[0004] In existing technologies, charging pile fault repair mainly revolves around hardware redundancy, software monitoring, and cloud platform collaboration. For example, some solutions employ dual backup power supplies or modular designs, achieving fault isolation through hardware switching. However, these methods are costly and struggle to handle soft faults. Other solutions rely on sensor data (such as voltage and temperature) to set fixed thresholds, triggering alarms or system restarts when parameters exceed these limits. However, threshold settings often depend on empirical judgment, which can easily lead to malfunctions in dynamic operating scenarios (such as peak charging times). In recent years, the Internet of Things (IoT) and big data analytics have been introduced into this field. For instance, charging pile operation data is aggregated through cloud platforms, and machine learning models such as support vector machines and convolutional neural networks are used for fault prediction. However, these models are mostly trained on static data, making them difficult to adapt to real-time event streams and consuming significant computational resources, which is unfavorable for deployment in resource-constrained embedded environments.

[0005] In storage management, Flash partitioning technology is widely used for firmware upgrades and data storage in charging piles. Typical implementations divide Flash into boot sectors, application sectors, and data sectors, using copy-on-write or log-based structures to reduce write / erase wear. However, existing partitioning strategies are mostly based on static rules, such as periodic defragmentation or bad block mapping, lacking adaptive capabilities. When partition errors occur (such as bit flips or bad block propagation), the system often performs global recovery or shutdown repair, causing service interruptions. For example, some commercial charging pile solutions use the FAT file system to manage Flash storage, but FAT systems are prone to storage fragmentation, and the repair process requires manual intervention, with an average repair time of several hours. Furthermore, Flash durability faces severe challenges in the high-frequency write scenarios of charging piles; while existing wear leveling algorithms can extend lifespan, they cannot achieve predictive repair of potential faults.

[0006] In the field of intelligent diagnostics, neural network technology has been used for fault detection. Traditional artificial neural networks (such as BP networks) and deep learning models can train classifiers using historical data to identify common fault modes such as overvoltage and short circuits. However, they rely on batch data processing, making it difficult to achieve low-latency response and consuming a lot of power.

[0007] At the repair mechanism level, adaptive repair technology aims to achieve fault tolerance by dynamically adjusting system configuration. For example, in storage systems, fault points can be avoided through bad block remapping or data path adjustment. However, existing methods mostly rely on preset rules (such as threshold-triggered mapping table updates) and lack learning and evolution capabilities. When faced with complex hardware and software interaction faults in charging piles, fixed strategies often fail to achieve accurate repair and may even trigger secondary faults.

[0008] In summary, existing charging pile fault repair technologies mainly suffer from the following three problems: First, the rigid Flash partition management strategy leads to low repair efficiency. The charging pile's Flash memory stores critical data such as user authentication information and charging records; improper partition management can cause data loss or system crashes. For example, the accumulation of bad blocks can compromise firmware integrity, while existing solutions often employ passive response mechanisms, such as formatting the entire system after detecting an error, resulting in irreversible data loss and lengthy repair times. At a deeper level, the programming / erasing cycle of Flash memory requires predictive partition management, but existing methods fail to effectively integrate with fault prediction models, making proactive maintenance difficult.

[0009] Second, the application of neural networks in fault diagnosis is still insufficient, which limits the accuracy of diagnosis. Charging pile faults often manifest as time-series related events, such as starting current oscillations and communication timeouts, exhibiting sparse and bursty characteristics. However, current neural network research is mostly focused on image and speech recognition, with insufficient optimization for its application in industrial real-time systems. Furthermore, the interaction mechanism with the storage system has not been fully explored, making it impossible to utilize information such as error logs generated by Flash partitions to optimize network weights, resulting in a disconnect between the diagnostic model and the actual operating environment.

[0010] Third, the level of intelligence in adaptive repair is insufficient. Repair should be a dynamic decision-making process, requiring flexible adjustments to strategies based on fault type, system load, and environmental conditions. However, existing repair mechanisms are mostly based on static rules, such as simply switching to a backup block, failing to consider the correlation between faults (e.g., concurrent errors in multiple partitions). In complex charging pile systems, faults can propagate in a chain, for example, a Flash error can cause a communication module to fail; fixed repair paths are insufficient to handle such coupled faults. The lack of adaptive capability is also reflected in the absence of a learning mechanism in the repair strategy; the system cannot iteratively optimize based on historical repair results, leading to a decline in repair effectiveness as the environment changes.

[0011] The root cause of these problems lies in the fragmented and uncoordinated nature of existing technologies, particularly in modules such as Flash partition management, fault diagnosis, and repair strategies. For example, partition error information is not fed back to the diagnostic model in real time, and repair decisions do not fully consider the actual state of the storage system. This fragmentation not only reduces overall system efficiency but also increases system complexity and maintenance costs.

[0012] Therefore, how to provide a method and system for repairing charging pile faults, and improve the intelligence level and overall efficiency of charging pile fault repair, has become an urgent technical problem to be solved. Summary of the Invention

[0013] The technical problem to be solved by the present invention is to provide a method and system for repairing charging pile faults, so as to improve the intelligence level and overall efficiency of charging pile fault repair.

[0014] In a first aspect, the present invention provides a method for repairing charging pile faults, comprising the following steps: Step S1: The server collects a large amount of historical operating status data from different charging piles, including at least Flash partition status information, electrical parameter timing events, and system load indicators. The server preprocesses the historical operating status data to construct a dataset. Step S2: The server creates a charging pile fault diagnosis model based on the feature extraction layer, temporal alignment layer, modality fusion layer, feature enhancement layer, and diagnostic prediction layer, and sets the loss function of the charging pile fault diagnosis model. Step S3: The server divides the dataset into a training set, a validation set, and a test set to train, validate, and test the charging pile fault diagnosis model in sequence. Step S4: After compressing the tested charging pile fault diagnosis model, the server deploys it to the charging pile. Step S5: During operation, the charging pile collects real-time operating status data. After preprocessing the real-time operating status data, it is input into the deployed charging pile fault diagnosis model to obtain fault diagnosis results that include at least the diagnosis conclusion, fault details, repair suggestions, associated fault risk warning, predictive maintenance suggestions, and confidence level. Step S6: The charging pile automatically performs fault repair operations based on the fault diagnosis results and records the fault repair results; Step S7: The charging pile uses a federated learning mechanism to iteratively optimize the charging pile fault diagnosis model based on the fault repair results and real-time operating status data.

[0015] Furthermore, step S1 specifically involves the server collecting a large amount of historical operating status data from different charging piles, including at least Flash partition status information, electrical parameter timing events, and system load indicators. The Flash partition status information includes at least partition integrity verification data, storage medium physical status data, and partition usage and performance data; the partition integrity verification data includes at least CRC checksum and file system metadata status; the storage medium physical status data includes at least bad block information, wear leveling statistics, and data retention capability indication; the partition usage and performance data includes at least available space ratio and fragmentation level, and read / write operation latency. The electrical parameter timing events include at least core power circuit parameters, charging control and communication events, and protection device action events; the core power circuit parameters include at least voltage and current waveforms and power module status; the charging control and communication events include at least charging handshake protocol sequences and network messages. The system load metrics include at least computing resource utilization, system process and service status, communication stack load, and task scheduling information; the computing resource utilization includes at least CPU load and memory usage; the system process and service status includes at least the heartbeat of critical daemons and the watchdog timer status; the communication stack load includes at least the number of network connections, network bandwidth utilization and packet throughput, and TCP / IP stack error statistics. The historical operational status data are preprocessed, including at least data cleaning, data integration and alignment, and feature engineering. The preprocessed historical operational status data are annotated with at least diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, and predictive maintenance suggestions. A dataset is constructed based on the annotated historical operational status data. The data cleaning includes at least missing value handling and outlier handling; the data integration and alignment includes at least timestamp alignment and data association; the feature engineering includes at least feature extraction and feature transformation.

[0016] Furthermore, in step S2, the feature extraction layer is constructed based on the Flash feature extraction module, the electrical parameter feature extraction module, and the system load feature extraction module; The Flash feature extraction module is used to extract first local features from Flash partition state information through a first 1D convolutional neural network, and to integrate global time information into the first local features through a Transformer encoder to obtain a Flash feature sequence. The electrical parameter feature extraction module is used to extract second local features from electrical parameter time-series events through a second 1D convolutional neural network, and to capture first time dependencies from the second local features through a long short-term memory network to obtain an electrical parameter feature sequence. The system load feature extraction module is used to extract third local features from system load indicators through a third 1D convolutional neural network, and to capture second time dependencies from the third local features through a gated recurrent unit to obtain a system load feature sequence. The timing alignment layer is used to perform time alignment operations on the Flash feature sequence, electrical parameter feature sequence and system load feature sequence through a dynamic time warping algorithm, and uses deformable convolution to adjust the sequence length and phase. The modal fusion layer is used to extract intermodal correlation features from the Flash feature sequence, electrical parameter feature sequence and system load feature sequence output by the timing alignment layer through a cross-attention mechanism. The gating fusion unit dynamically adjusts the weights of each modal feature according to the correlation features to obtain a fused feature sequence. The feature enhancement layer is constructed based on a self-supervised contrastive learning module and a knowledge graph enhancement module. The self-supervised contrastive learning module is used to construct positive and negative sample pairs through a contrastive loss function, and to perform enhancement operations on the fused feature sequence to obtain a first-level enhanced feature sequence. The knowledge graph enhancement module is used to interact the preset charging pile component relationship knowledge graph with the first-level enhanced feature sequence through a graph neural network, and to integrate semantic information using an attention mechanism to obtain the second-level enhanced feature sequence. The diagnostic prediction layer is constructed based on a multi-task diagnostic prediction module and an uncertainty estimation module; The multi-task diagnosis and prediction module is used to aggregate each secondary enhanced feature sequence into a context vector through an attention pooling layer, and output diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings and predictive maintenance suggestions through multiple fully connected layers. The uncertainty estimation module is used to evaluate the confidence level of diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, and predictive maintenance suggestions using Monte Carlo dropout technology, and outputs fault diagnosis results carrying diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, predictive maintenance suggestions, and confidence levels; The diagnostic conclusion includes the fault status indicator and the fault type; The fault details include at least the faulty component, the fault severity level, and the fault characteristic values; The repair recommendations include at least the recommended repair actions and repair strategy parameters.

[0017] Furthermore, step S3 specifically includes: The server uses k-fold cross-validation to divide the dataset into training, validation, and test sets in a ratio of 7:1.5:1.5. The charging pile fault diagnosis model is trained using the training set until the loss value of the loss function is less than a preset loss threshold. The F1-Score, precision, and recall are calculated using the validation set to validate the trained charging pile fault diagnosis model. If the validation fails, the training set is expanded and training continues. If the validation passes, AUC-ROC, AUC-PR, and accuracy are calculated using the test set to test the validated charging pile fault diagnosis model. If the test fails, the training set is expanded and training continues. If the test passes, training ends.

[0018] Furthermore, step S4 specifically includes: The server sequentially compresses the tested charging pile fault diagnosis model through knowledge distillation, dynamic pruning, and quantization techniques. After fine-tuning and performance verification of the compressed charging pile fault diagnosis model, it is deployed to the charging pile. Step S5 specifically involves: During operation, the charging pile collects real-time operating status data. Through multi-threading technology, the real-time operating status data is preprocessed, including at least data cleaning, data integration and alignment. The preprocessed real-time operating status data and historical operating status data of the local machine in a preset period are input into the deployed charging pile fault diagnosis model for inference. The output includes at least the diagnosis conclusion, fault details, repair suggestions, associated fault risk warning, predictive maintenance suggestions and confidence level of the fault diagnosis result. Step S6 specifically involves: The charging pile analyzes the fault diagnosis results. When the diagnosis conclusion is a fault, it determines whether the repair suggestion has the execution authority and execution capability. If so, the fault repair operation is automatically executed based on the repair suggestion, and the fault repair result is recorded. If not, a fault notification is generated based on the fault diagnosis results, and the fault notification is pushed to the management terminal in real time. Step S7 specifically involves: The charging pile uses a federated learning mechanism to construct an incremental dataset based on the fault repair results and real-time operating status data at preset iteration intervals. The local charging pile fault diagnosis model is trained using the incremental dataset to obtain local model parameters. The local model parameters are then uploaded to the server for aggregation to obtain global model parameters fed back from the server. The local charging pile fault diagnosis model is updated based on the global model parameters to iteratively optimize the charging pile fault diagnosis model.

[0019] Secondly, the present invention provides a charging pile fault repair system, comprising the following modules: The dataset construction module is used by the server to collect a large amount of historical operating status data from different charging piles, including at least Flash partition status information, electrical parameter timing events and system load indicators, and to preprocess the historical operating status data to construct the dataset. The charging pile fault diagnosis model creation module is used by the server to create a charging pile fault diagnosis model based on a feature extraction layer, a temporal alignment layer, a modal fusion layer, a feature enhancement layer, and a diagnosis prediction layer, and to set the loss function of the charging pile fault diagnosis model. The charging pile fault diagnosis model training module is used by the server to divide the dataset into a training set, a validation set, and a test set, so as to train, validate, and test the charging pile fault diagnosis model in sequence. The charging pile fault diagnosis model deployment module is used by the server to compress the tested charging pile fault diagnosis model and deploy it to the charging pile. The fault diagnosis module is used to collect real-time operating status data during the operation of the charging pile. After preprocessing the real-time operating status data, it is input into the deployed charging pile fault diagnosis model to obtain fault diagnosis results that include at least the diagnosis conclusion, fault details, repair suggestions, associated fault risk warning, predictive maintenance suggestions, and confidence level. The fault repair module is used to automatically perform fault repair operations based on the fault diagnosis results of the charging pile and record the fault repair results; The charging pile fault diagnosis model iteration module is used to iteratively optimize the charging pile fault diagnosis model based on the fault repair results and real-time operating status data through a federated learning mechanism.

[0020] Furthermore, the dataset construction module is specifically used for: the server to collect a large amount of historical operating status data from different charging piles, including at least Flash partition status information, electrical parameter timing events, and system load indicators; The Flash partition status information includes at least partition integrity verification data, storage medium physical status data, and partition usage and performance data; the partition integrity verification data includes at least CRC checksum and file system metadata status; the storage medium physical status data includes at least bad block information, wear leveling statistics, and data retention capability indication; the partition usage and performance data includes at least available space ratio and fragmentation level, and read / write operation latency. The electrical parameter timing events include at least core power circuit parameters, charging control and communication events, and protection device action events; the core power circuit parameters include at least voltage and current waveforms and power module status; the charging control and communication events include at least charging handshake protocol sequences and network messages. The system load metrics include at least computing resource utilization, system process and service status, communication stack load, and task scheduling information; the computing resource utilization includes at least CPU load and memory usage; the system process and service status includes at least the heartbeat of critical daemons and the watchdog timer status; the communication stack load includes at least the number of network connections, network bandwidth utilization and packet throughput, and TCP / IP stack error statistics. The historical operational status data are preprocessed, including at least data cleaning, data integration and alignment, and feature engineering. The preprocessed historical operational status data are annotated with at least diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, and predictive maintenance suggestions. A dataset is constructed based on the annotated historical operational status data. The data cleaning includes at least missing value handling and outlier handling; the data integration and alignment includes at least timestamp alignment and data association; the feature engineering includes at least feature extraction and feature transformation.

[0021] Furthermore, in the charging pile fault diagnosis model creation module, the feature extraction layer is constructed based on the Flash feature extraction module, the electrical parameter feature extraction module, and the system load feature extraction module; The Flash feature extraction module is used to extract first local features from Flash partition state information through a first 1D convolutional neural network, and to integrate global time information into the first local features through a Transformer encoder to obtain a Flash feature sequence. The electrical parameter feature extraction module is used to extract second local features from electrical parameter time-series events through a second 1D convolutional neural network, and to capture first time dependencies from the second local features through a long short-term memory network to obtain an electrical parameter feature sequence. The system load feature extraction module is used to extract third local features from system load indicators through a third 1D convolutional neural network, and to capture second time dependencies from the third local features through a gated recurrent unit to obtain a system load feature sequence. The timing alignment layer is used to perform time alignment operations on the Flash feature sequence, electrical parameter feature sequence and system load feature sequence through a dynamic time warping algorithm, and uses deformable convolution to adjust the sequence length and phase. The modal fusion layer is used to extract intermodal correlation features from the Flash feature sequence, electrical parameter feature sequence and system load feature sequence output by the timing alignment layer through a cross-attention mechanism. The gating fusion unit dynamically adjusts the weights of each modal feature according to the correlation features to obtain a fused feature sequence. The feature enhancement layer is constructed based on a self-supervised contrastive learning module and a knowledge graph enhancement module. The self-supervised contrastive learning module is used to construct positive and negative sample pairs through a contrastive loss function, and to perform enhancement operations on the fused feature sequence to obtain a first-level enhanced feature sequence. The knowledge graph enhancement module is used to interact the preset charging pile component relationship knowledge graph with the first-level enhanced feature sequence through a graph neural network, and to integrate semantic information using an attention mechanism to obtain the second-level enhanced feature sequence. The diagnostic prediction layer is constructed based on a multi-task diagnostic prediction module and an uncertainty estimation module; The multi-task diagnosis and prediction module is used to aggregate each secondary enhanced feature sequence into a context vector through an attention pooling layer, and output diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings and predictive maintenance suggestions through multiple fully connected layers. The uncertainty estimation module is used to evaluate the confidence level of diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, and predictive maintenance suggestions using Monte Carlo dropout technology, and outputs fault diagnosis results carrying diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, predictive maintenance suggestions, and confidence levels; The diagnostic conclusion includes the fault status indicator and the fault type; The fault details include at least the faulty component, the fault severity level, and the fault characteristic values; The repair recommendations include at least the recommended repair actions and repair strategy parameters.

[0022] Furthermore, the charging pile fault diagnosis model training module is specifically used for: The server uses k-fold cross-validation to divide the dataset into training, validation, and test sets in a ratio of 7:1.5:1.5. The charging pile fault diagnosis model is trained using the training set until the loss value of the loss function is less than a preset loss threshold. The F1-Score, precision, and recall are calculated using the validation set to validate the trained charging pile fault diagnosis model. If the validation fails, the training set is expanded and training continues. If the validation passes, AUC-ROC, AUC-PR, and accuracy are calculated using the test set to test the validated charging pile fault diagnosis model. If the test fails, the training set is expanded and training continues. If the test passes, training ends.

[0023] Furthermore, the charging pile fault diagnosis model deployment module is specifically used for: The server sequentially compresses the tested charging pile fault diagnosis model through knowledge distillation, dynamic pruning, and quantization techniques. After fine-tuning and performance verification of the compressed charging pile fault diagnosis model, it is deployed to the charging pile. The fault diagnosis module is specifically used for: During operation, the charging pile collects real-time operating status data. Through multi-threading technology, the real-time operating status data is preprocessed, including at least data cleaning, data integration and alignment. The preprocessed real-time operating status data and historical operating status data of the local machine in a preset period are input into the deployed charging pile fault diagnosis model for inference. The output includes at least the diagnosis conclusion, fault details, repair suggestions, associated fault risk warning, predictive maintenance suggestions and confidence level of the fault diagnosis result. The fault repair module is specifically used for: The charging pile analyzes the fault diagnosis results. When the diagnosis conclusion is a fault, it determines whether the repair suggestion has the execution authority and execution capability. If so, the fault repair operation is automatically executed based on the repair suggestion, and the fault repair result is recorded. If not, a fault notification is generated based on the fault diagnosis results, and the fault notification is pushed to the management terminal in real time. The charging pile fault diagnosis model iteration module is specifically used for: The charging pile uses a federated learning mechanism to construct an incremental dataset based on the fault repair results and real-time operating status data at preset iteration intervals. The local charging pile fault diagnosis model is trained using the incremental dataset to obtain local model parameters. The local model parameters are then uploaded to the server for aggregation to obtain global model parameters fed back from the server. The local charging pile fault diagnosis model is updated based on the global model parameters to iteratively optimize the charging pile fault diagnosis model.

[0024] The advantages of this invention are: 1. A server collects a large amount of historical operational status data from different charging piles, including at least Flash partition status information, electrical parameter timing events, and system load indicators. This historical operational status data is preprocessed to construct a dataset. Then, the server creates a charging pile fault diagnosis model based on a feature extraction layer, timing alignment layer, modality fusion layer, feature enhancement layer, and diagnostic prediction layer. A loss function for the charging pile fault diagnosis model is set, and the dataset is divided into training, validation, and test sets for sequential training, validation, and testing. The tested charging pile fault diagnosis model is compressed and deployed to the charging piles. During operation, real-time operational status data is collected from the charging piles and input into the charging pile fault diagnosis model to obtain fault diagnosis results including at least diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, predictive maintenance suggestions, and confidence levels. The system automatically executes fault repair operations based on fault diagnosis results and records the repair results. Through a federated learning mechanism, the charging pile fault diagnosis model iteratively optimizes itself based on fault repair results and real-time operational status data. This involves integrating multimodal data such as Flash partition status information, electrical parameter timing events, and system load indicators into a dedicated charging pile fault diagnosis model, enabling accurate diagnosis and predictive early warning of complex faults. Furthermore, it automatically executes repair operations based on the fault diagnosis results and innovatively introduces a federated learning mechanism, allowing the charging pile fault diagnosis model to continuously self-optimize using local fault repair results. Ultimately, this forms an intelligent operation and maintenance system with accuracy, adaptability, and continuous evolution capabilities, fundamentally solving the problems of module fragmentation, rigid strategies, and lack of learning ability in traditional solutions. This significantly improves the intelligence level and overall efficiency of charging pile fault repair.

[0025] 2. By collecting multimodal historical operating status data, including Flash partition status information, electrical parameter timing events, and system load indicators, and performing preprocessing such as data cleaning, integration, alignment, and feature engineering, a high-quality dataset is constructed. This comprehensive and refined data collection and processing method ensures that the input information for fault diagnosis is rich and reliable, and can capture subtle anomalies in the operation of charging piles, thereby significantly improving the accuracy and robustness of fault diagnosis, reducing the risk of false alarms and missed alarms, and demonstrating the advantages of the data-driven approach.

[0026] 3. The charging pile fault diagnosis model integrates a feature extraction layer, a temporal alignment layer, a modality fusion layer, a feature enhancement layer, and a diagnostic prediction layer. It also integrates a variety of advanced technologies such as 1D convolutional neural networks, Transformer, LSTM, GRU, dynamic time warping, cross-attention mechanism, self-supervised contrastive learning, and knowledge graph. This multi-technology fusion architecture can effectively process temporal data and multimodal information, extract deep features, and enhance the model's generalization ability. The diagnostic results not only include fault conclusions but also provide details, suggestions, and confidence levels, thus improving the comprehensiveness and interpretability of fault diagnosis.

[0027] 4. Automatic fault repair operations based on fault diagnosis results have been implemented. The charging pile can analyze the fault diagnosis results, determine the feasibility of the repair suggestions, and automatically execute the repair or generate a notification. This automated process reduces the need for manual intervention, improves fault response speed and processing efficiency, and provides data support for subsequent optimization by recording fault repair results. This demonstrates the practicality of intelligent operation and maintenance and helps to reduce the downtime and maintenance costs of charging piles.

[0028] 5. By adopting a federated learning mechanism, charging piles can iteratively optimize the model locally based on real-time data and repair results. Only local model parameters are uploaded to the server for aggregation, rather than the original data. This method effectively protects user privacy and data security, while realizing distributed continuous improvement of the model, avoiding the risks brought by data centralization, and ensuring that the model can adapt to the personalized operating environment of different charging piles, thus improving the scalability and long-term applicability of the technology.

[0029] 6. The model is compressed through knowledge distillation, dynamic pruning, and quantization techniques, and fine-tuned and verified before deployment to ensure that the fault diagnosis model runs efficiently on resource-constrained charging pile equipment. This compression and optimization strategy balances model performance and computational efficiency, enabling complex AI models to achieve real-time inference on edge devices, improving the practicality and promotion value of the technology, and making it suitable for the low latency requirements of large-scale charging pile networks.

[0030] 7. The diagnostic prediction layer adopts a multi-task diagnostic prediction module and an uncertainty estimation module, which can simultaneously output diagnostic conclusions, fault details, repair suggestions, risk warnings and maintenance suggestions, and assess confidence level. This multi-task and uncertainty assessment design enhances the reliability and credibility of diagnostic results, provides decision support for users, reduces the possibility of blind operation, and is particularly suitable for safety-critical charging pile scenarios, improving the overall system robustness.

[0031] 8. By regularly updating the model based on incremental data through a federated learning mechanism, the charging pile fault diagnosis model is continuously and adaptively optimized. This iterative learning capability enables the model to improve itself over time and with changes in the operating environment, adapting to new fault modes and avoiding the problem of model obsolescence. It reflects the dynamic evolutionary characteristics of artificial intelligence technology in industrial applications and has long-term technical value.

[0032] 9. By comprehensively collecting multimodal operating data and constructing an advanced fault diagnosis model, high-precision and automated charging pile fault diagnosis and repair are achieved. Its advantages lie in the integration of multi-source feature extraction, temporal alignment and knowledge enhancement technologies to ensure diagnostic reliability, combined with model compression to achieve efficient edge deployment, and continuous optimization of the model through a federated learning mechanism while protecting data privacy. Ultimately, a highly efficient operation and maintenance system integrating intelligent early warning, autonomous repair and adaptive learning is formed, which significantly improves the availability and maintenance efficiency of charging piles. Attached Figure Description

[0033] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0034] Figure 1 This is a flowchart of a charging pile fault repair method according to the present invention.

[0035] Figure 2 This is a schematic diagram of the structure of a charging pile fault repair system according to the present invention.

[0036] Figure 3 This is a schematic diagram of the architecture of the charging pile fault diagnosis model of the present invention. Detailed Implementation

[0037] The technical solution in this application embodiment has the following general idea: multimodal data such as Flash partition status information, electrical parameter timing events, and system load indicators are fused and input into a dedicated charging pile fault diagnosis model, realizing accurate diagnosis and predictive early warning of complex faults; then, repair operations are automatically executed based on the fault diagnosis results, and a federated learning mechanism is innovatively introduced, enabling the charging pile fault diagnosis model to continuously optimize itself using local fault repair results, ultimately forming an intelligent operation and maintenance system with accuracy, adaptability, and continuous evolution capabilities, so as to improve the intelligence level and comprehensive efficiency of charging pile fault repair.

[0038] Please refer to Figures 1 to 3 As shown, a preferred embodiment of a charging pile fault repair method of the present invention includes the following steps: Step S1: The server collects a large amount of historical operating status data from different charging piles, including at least Flash partition status information, electrical parameter timing events, and system load indicators. The server preprocesses the historical operating status data to construct a dataset. Step S2: The server creates a charging pile fault diagnosis model based on the feature extraction layer, temporal alignment layer, modality fusion layer, feature enhancement layer, and diagnostic prediction layer, and sets the loss function of the charging pile fault diagnosis model. Step S3: The server divides the dataset into a training set, a validation set, and a test set to train, validate, and test the charging pile fault diagnosis model in sequence. Step S4: After compressing the tested charging pile fault diagnosis model, the server deploys it to the charging pile. Step S5: During operation, the charging pile collects real-time operating status data. After preprocessing the real-time operating status data, it is input into the deployed charging pile fault diagnosis model to obtain fault diagnosis results that include at least the diagnosis conclusion, fault details, repair suggestions, associated fault risk warning, predictive maintenance suggestions, and confidence level. Step S6: The charging pile automatically performs fault repair operations based on the fault diagnosis results and records the fault repair results; In step S7, the charging pile uses a federated learning mechanism to iteratively optimize the charging pile fault diagnosis model based on the fault repair results and real-time operating status data.

[0039] By constructing a closed-loop system integrating multimodal data perception, intelligent diagnosis, automatic execution, and continuous evolution capabilities, the intelligence level and overall efficiency of charging pile fault repair have been significantly improved. It collects and deeply integrates Flash partition health status, electrical parameter timing events, and system load indicators, utilizing an advanced multimodal timing fusion model for accurate diagnosis and predictive maintenance. Based on the fault diagnosis results, it dynamically generates repair strategies and executes them automatically. Finally, through a federated learning mechanism, the repair results are fed back to the model for continuous iterative optimization. This achieves a leap from passive response to proactive prediction, from static rules to dynamic learning, and from modular fragmentation to collaborative autonomy, effectively improving repair efficiency, system reliability, and operational economy.

[0040] Specifically, step S1 involves the server collecting a large amount of historical operating status data from different charging piles, including at least Flash partition status information, electrical parameter timing events, and system load indicators. The Flash partition status information includes at least partition integrity verification data, storage medium physical status data, and partition usage and performance data. The partition integrity verification data includes at least CRC checksums (used for integrity verification of critical static data such as the boot sector and application firmware area), and file system metadata status (such as the integrity of inodes and the consistency of superblocks). The storage medium physical status data includes at least bad block information (including the number of marked bad blocks, the growth rate of bad blocks, and the physical location of bad block distribution (e.g., whether they are concentrated in critical data areas, which is the direct basis for determining whether "bad blocks in the data area exceed the threshold"), wear leveling statistics (the number of erases for each physical block; average wear level and maximum wear level), and data retention capability indication (indirectly determining whether data has decayed due to prolonged lack of refresh by reading the voltage level or bit error rate of the storage unit). The partition usage and performance data includes at least the available space ratio and fragmentation level (insufficient space or severe fragmentation will reduce write performance and exacerbate wear), and read / write operation latency (an abnormal increase in read / write speed may indicate storage medium performance degradation or controller failure). The electrical parameter timing events include at least core power circuit parameters, charging control and communication events, and protection device action events (action timing of fuses, contactors, and relays: recording their opening and closing times and performing correlation analysis with electrical parameter abnormal events to distinguish between real faults and false alarms); the core power circuit parameters include at least voltage and current waveforms (instantaneous overvoltage / undervoltage, overcurrent; voltage drop / surge, current spike; waveform distortion), and power module status (such as IGBT drive signals, switching frequency, radiator temperature, etc., used to diagnose "power module IGBT signal abnormalities"); the charging control and communication events include at least charging handshake protocol sequences (message timing and response time for communication with the vehicle BMS; faults such as "communication timeout with BMS" need to be judged by analyzing such event sequences), and network messages (timing, packet loss rate, and error frames of key control commands and status feedback messages). The system load metrics include at least the following: computing resource utilization, system process and service status, communication stack load, and task scheduling information (readiness delay and scheduling delay of real-time tasks; excessive delay may indicate high system load, affecting the real-time performance of critical control loops). The computing resource utilization includes at least CPU load (not only overall utilization, but also CPU usage of each task / process to identify "abnormal exit of critical processes" or resource preemption issues), and memory usage (total memory usage, available memory, and memory fragmentation; memory leaks are a common cause of system instability). The system process and service status includes at least the heartbeat of critical daemons (whether network services, payment services, monitoring agents, etc., are running normally) and the watchdog timer status ("watchdog timeout reset" is the final manifestation of excessive system load or deadlock). The communication stack load includes at least the number of network connections (the number of concurrent connections with the backend server, payment gateway, other stubs, etc.), network bandwidth utilization and packet throughput, and TCP / IP stack error statistics (such as retransmission rate, connection errors, etc., used to diagnose the root cause of "interrupted heartbeat connection with the server"). The historical operational status data are preprocessed, including at least data cleaning, data integration and alignment, and feature engineering. The preprocessed historical operational status data are annotated with at least diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, and predictive maintenance suggestions. A dataset is constructed based on the annotated historical operational status data. The data cleaning includes at least missing value handling (i.e., checking for missing values ​​such as null values, NaN (non-numeric), or specifically labeled missing values ​​in the dataset, and taking measures such as deletion or interpolation), and outlier handling (using statistical methods, business rule methods, or DBSCAN clustering to identify outliers, and taking measures such as correction, deletion, or retention and labeling). The data integration and alignment includes at least timestamp alignment (unifying all different types of data streams to the same time base; for data with different collection frequencies (such as millisecond-level voltage waveforms and second-level CPU loads), resampling is required), and data association (associating the collected status data with historical fault event records, which is the key to building a supervised learning dataset). The feature engineering includes at least feature extraction (e.g., extracting the bad block growth rate from "bad block information"; extracting the average erase / write count and maximum erase / write count from "wear leveling statistics"; calculating insufficient space warning from "available space ratio" (e.g., 1 if available space < 10%, otherwise 0); extracting the mean, standard deviation, peak value, valley value, and ripple coefficient from voltage and current waveforms). The system performs various functions, including: 1) 2) Fast Fourier Transform (FFT) to extract the main frequency component and harmonic components, which are highly effective for assessing the health status of power devices; 3) Encoding of the "charging handshake protocol sequence" and "protection device action events" to extract features such as protocol interaction timeout counts and protection action frequency; 4) Extracting sliding window statistics from CPU and memory data, such as the average load, peak load, and trend (slope) over the past 5 minutes; 5) Extracting the error rate from "TCP / IP stack error statistics"; and 6) Extracting the number of restarts from "watchdog timer status" and performing feature transformations (normalization / standardization). Normalization: Transform features of different dimensions and orders of magnitude (such as voltage values ​​of several hundred volts, CPU load of 0-100%) to the same scale. Common methods include Min-Max normalization (to the [0,1] interval) and Z-Score standardization (mean of 0, variance of 1), which can accelerate model convergence and improve stability; Discretization: Segment continuous values, such as dividing the number of erases in "wear balance statistics" into "low", "medium" and "high" risk levels; One-hot encoding: Convert categorical variables (such as "fault type") into binary vectors to facilitate model understanding.

[0041] In step S2, the feature extraction layer is constructed based on the Flash feature extraction module, the electrical parameter feature extraction module, and the system load feature extraction module; The Flash feature extraction module is used to extract first local features from Flash partition state information through a first 1D convolutional neural network, and to integrate global temporal information into the first local features through a Transformer encoder to obtain a Flash feature sequence. The first 1D convolutional neural network is used to capture local temporal patterns of Flash data, such as short-term changes in partition integrity verification data. The Transformer encoder is used to capture long-term dependencies, such as trends in storage medium physical state data. The advantages include the ability to effectively handle the temporal dynamics and multi-scale features of Flash data.

[0042] The electrical parameter feature extraction module is used to extract second local features from electrical parameter time-series events through a second 1D convolutional neural network, and to capture first time dependencies from the second local features through a long short-term memory network (LSTM) to obtain an electrical parameter feature sequence; the second 1D convolutional neural network is used to extract high-frequency local features such as voltage and current waveforms; the LSTM is used to model long-term events in the sequence, such as the time context of protection device action events; the advantages include strong modeling capability for electrical parameter time-series events and adaptability to variable-length sequences.

[0043] The system load feature extraction module is used to extract third local features from system load indicators through a third 1D convolutional neural network, and to capture second time dependencies from the third local features through a gated recurrent unit (GRU) to obtain a system load feature sequence; the third 1D convolutional neural network is used to capture short-term fluctuations in indicators such as computing resource utilization; the GRU is used to efficiently learn the temporal evolution of system processes and service states; the advantages include high computational efficiency and suitability for processing real-time system load data.

[0044] The timing alignment layer is used to perform time alignment operations on Flash feature sequences, electrical parameter feature sequences, and system load feature sequences using a dynamic time warping algorithm, and uses deformable convolution to adjust sequence length and phase; advantages include the ability to handle modal data with different sampling rates and improve the accuracy of fusion.

[0045] The modal fusion layer is used to extract intermodal correlation features from the Flash feature sequence, electrical parameter feature sequence, and system load feature sequence output by the timing alignment layer through a cross-attention mechanism. The gating fusion unit dynamically adjusts the weights of each modal feature (Flash feature sequence, electrical parameter feature sequence, and system load feature sequence) based on the correlation features to obtain a fused feature sequence. The advantages include adaptive weight allocation and enhanced modal complementarity.

[0046] The feature enhancement layer is constructed based on a self-supervised contrastive learning module and a knowledge graph enhancement module. The self-supervised contrastive learning module is used to construct positive sample pairs (data with similar fault states) and negative sample pairs (data with different fault states) through a contrastive loss function, and to perform enhancement operations on the fused feature sequence to obtain a first-level enhanced feature sequence. The advantages include improving the robustness of features against noise and outliers without the need for additional annotation.

[0047] The knowledge graph enhancement module is used to interact with a preset knowledge graph of charging pile component relationships (such as fault propagation paths) and a first-level enhanced feature sequence through a graph neural network, and to integrate semantic information using an attention mechanism to obtain a second-level enhanced feature sequence. The advantages include injecting domain knowledge and improving the accuracy and interpretability of diagnosis.

[0048] The diagnostic prediction layer is constructed based on a multi-task diagnostic prediction module and an uncertainty estimation module; The multi-task diagnostic prediction module is used to aggregate the secondary enhanced feature sequences into a context vector through an attention pooling layer, and output diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, and predictive maintenance suggestions through multiple fully connected layers; the attention pooling layer is used to adaptively focus on key time steps; the fully connected layer is used to perform multi-task prediction; the advantages include shared feature representation, improved training efficiency and generalization ability.

[0049] The uncertainty estimation module is used to evaluate the confidence level of diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, and predictive maintenance suggestions using Monte Carlo dropout technology, and outputs fault diagnosis results carrying diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, predictive maintenance suggestions, and confidence levels; advantages include providing reliability indicators and enhancing the practicality of the model.

[0050] The diagnostic conclusions include fault status indicators (clearly determining the current system status, such as "normal", "warning", "fault") and fault types (Flash storage faults: such as "boot partition CRC check error", "data area bad block exceeds threshold", "file system index corruption"; power supply electrical faults: such as "DC side overvoltage protection", "AC input undervoltage", "power module IGBT overheating"; communication faults: such as "communication timeout with BMS", "heartbeat connection interruption with server"; software / system faults: such as "watchdog timeout reset", "critical process abnormal exit", "system overload"). The fault details include at least the faulty component (accurate to the smallest operable unit, such as: "Flash chip 2, block address 0x8000-0x8FFF", "current sensor of phase A charging module", "SSL layer of TCP / IP communication stack"), fault severity level (assessing the urgency of the fault's impact on the system, such as "fatal (system crash)", "serious (loss of function)", "moderate (performance degradation)", "warning (requires attention)"), and fault characteristic values ​​(providing specific values ​​that trigger diagnosis, such as: "peak current reaches 52A (threshold 50A)", "Flash block erase count is 98,752 times (warning threshold 100,000 times)", "CPU utilization rate is above 95% for 3 consecutive minutes"). The repair recommendations include at least recommended repair actions (providing specific executable instructions, such as: "Enable Flash redundant backup partition and perform master-slave switch"; "Remap bad blocks in the data area and attempt to repair the file system"; "Perform a soft reboot of the charging module and reduce the current charging power to 80%"; "Restart the network service process") and repair strategy parameters (providing parameterized guidance for the repair actions, such as: "After reboot, the charging current limit is 30A"; "Set the priority of data migration to high").

[0051] Associated fault risk warnings indicate secondary or associated faults that may be caused by the current fault; for example, when a "Flash boot sector error" is diagnosed, a warning should be issued that "the system may fail to start, causing all charging services to be interrupted"; when a "communication module fault" is diagnosed, a warning should be issued that "fault data cannot be reported, affecting remote monitoring".

[0052] Predictive maintenance recommendations provide forward-looking warnings; for example: "The current Flash wear rate is high, and the predicted remaining lifespan is 45 days. Preventive maintenance is recommended." or "The power module's heat dissipation efficiency has decreased, and overheating risk is expected in high-temperature environments."

[0053] Confidence level is a probability value given to the reliability of the fault diagnosis result, such as "diagnosis confidence level: 92%"; this helps to adopt a more conservative repair strategy (such as requesting manual intervention) when the confidence level is low.

[0054] The formula for the loss function is: ; Where L represents the loss value of the loss function; Indicates multitasking loss; λ represents the comparison loss; λ represents the hyperparameter of the weighted balancing comparison loss; K represents the total number of tasks, with a value of 5, representing the five prediction tasks: diagnostic conclusion, fault details, repair suggestions, associated fault risk warning, and predictive maintenance suggestions; k represents the task index. This represents the weight of the k-th task; represents the loss of the k-th task; c represents the category index; This represents the true label of the k-th task and the c-th category; B represents the probability distribution of the prediction for the c-th category of the k-th task; B represents the batch size; sim() represents the cosine similarity function; z i z represents the feature vector of the i-th sample; i+ represents the positive sample feature of the i-th sample; τ represents the temperature parameter, used to control the sharpness of the similarity distribution, with a value of 0.1; z j Let z represent the feature vector of the j-th sample. i For negative sample features.

[0055] Step S3 specifically involves: The server uses k-fold cross-validation to divide the dataset into training, validation, and test sets in a ratio of 7:1.5:1.5. The charging pile fault diagnosis model is trained using the training set until the loss value of the loss function is less than a preset loss threshold. The F1-Score, precision, and recall are calculated using the validation set to validate the trained charging pile fault diagnosis model. If the validation fails, the training set is expanded and training continues. If the validation passes, AUC-ROC, AUC-PR, and accuracy are calculated using the test set to test the validated charging pile fault diagnosis model. If the test fails, the training set is expanded and training continues. If the test passes, training ends.

[0056] Step S4 specifically involves: The server sequentially compresses the tested charging pile fault diagnosis model through knowledge distillation, dynamic pruning, and quantization techniques. After fine-tuning and performance verification of the compressed charging pile fault diagnosis model, it is deployed to the charging pile. Step S5 specifically involves: During operation, the charging pile collects real-time operating status data. Through multi-threading technology, the real-time operating status data is preprocessed, including at least data cleaning, data integration and alignment. The preprocessed real-time operating status data and historical operating status data of the local machine in a preset period are input into the deployed charging pile fault diagnosis model for inference. The output includes at least the diagnosis conclusion, fault details, repair suggestions, associated fault risk warning, predictive maintenance suggestions and confidence level of the fault diagnosis result. Step S6 specifically involves: The charging pile analyzes the fault diagnosis results. When the diagnosis conclusion is a fault, it determines whether the repair suggestion has the execution authority and execution capability. If so, the fault repair operation is automatically executed based on the repair suggestion, and the fault repair result is recorded. If not, a fault notification is generated based on the fault diagnosis results, and the fault notification is pushed to the management terminal in real time. The fault repair results should include at least a summary of the repair execution, a detailed description of the repair operation process, post-repair verification data, resource consumption and impact assessment, correlation analysis and suggested revisions; The repair execution summary is a high-level overview of the repair actions and their final status, equivalent to a quick report, including: Original fault identifier: a unique ID or timestamp associated with the original "fault diagnosis result" that triggered this repair; Repair action trigger time: the precise time point at which the charging pile decided to execute the automatic repair; Executed repair operation: which one or more of the "repair suggestions" were specifically executed, such as: "performed a soft reboot of the charging module", "completed the Flash primary / backup partition switch"; Final repair status: the final result of this repair operation, usually an enumerated value.

[0057] The detailed repair process log records detailed data and event sequences during the repair process for in-depth analysis and problem diagnosis, including: Automated operation sequence logs: complete command-line logs or API call records of the execution of repair instructions, which are crucial for repairing failed debugging attempts; Key parameter snapshots: parameter settings of relevant systems during the execution of repair operations, such as: process status before reboot, number of bad blocks before switching partitions, adjusted power limit values, etc.; Manual intervention records (if any): if manual intervention is required or integrated during the repair process (e.g., administrator confirmation), the personnel involved, time, and operation details must be recorded.

[0058] Post-repair verification data records whether the repair is truly effective and needs to be verified with data. This part is the bridge connecting "action" and "effect", including: Post-repair system status re-inspection results: After the repair operation is completed, the charging pile will run a simplified diagnosis again, record the real-time status of key indicators, and compare it with that before the repair; Short-term stability monitoring summary: Within a short period of time after the repair (such as 5 minutes), the system continues to monitor relevant parameters and record whether there are any abnormalities again or whether it tends to stabilize.

[0059] Resource consumption and impact assessment records the impact of the repair operation itself on the system. It is used to evaluate the cost and optimization direction of the repair strategy, including: Repair time: The total time from the start of the repair to the completion of verification, which is important for evaluating the recovery time objective (RTO); System resource overhead: The CPU, memory, network bandwidth, etc. consumed in performing the repair operation (such as data migration, process restart); Business impact: Did the charging service get affected during the repair? For example: "The repair caused a 2-minute charging interruption" or "The repair was completed in the background with no business impact".

[0060] Correlation analysis and suggestion correction involve providing feedback and revising the initial diagnosis and suggestions based on the actual effects of the repair, demonstrating the system's learning ability. This includes: Diagnostic confidence verification: comparing the "confidence" of the "repair result" with the "diagnostic conclusion." If a high-confidence diagnosis fails to be repaired, it indicates that the model may have blind spots, which is an important negative feedback signal; Repair suggestion effectiveness rating: rating the effectiveness of the implemented "repair suggestions" (e.g., effective, moderately effective, ineffective). If the repair fails, new, practice-validated "repair suggestions" may be generated; Correlation risk warning verification: checking whether the previously predicted "correlation fault risk warning" has occurred to verify the accuracy of the risk prediction model.

[0061] Step S7 specifically involves: The charging pile uses a federated learning mechanism to construct an incremental dataset based on the fault repair results and real-time operating status data at preset iteration intervals. The local charging pile fault diagnosis model is trained using the incremental dataset to obtain local model parameters. The local model parameters are then uploaded to the server for aggregation to obtain global model parameters fed back from the server. The local charging pile fault diagnosis model is updated based on the global model parameters to iteratively optimize the charging pile fault diagnosis model.

[0062] A preferred embodiment of the charging pile fault repair system of the present invention includes the following modules: The dataset construction module is used by the server to collect a large amount of historical operating status data from different charging piles, including at least Flash partition status information, electrical parameter timing events and system load indicators, and to preprocess the historical operating status data to construct the dataset. The charging pile fault diagnosis model creation module is used by the server to create a charging pile fault diagnosis model based on a feature extraction layer, a temporal alignment layer, a modal fusion layer, a feature enhancement layer, and a diagnosis prediction layer, and to set the loss function of the charging pile fault diagnosis model. The charging pile fault diagnosis model training module is used by the server to divide the dataset into a training set, a validation set, and a test set, so as to train, validate, and test the charging pile fault diagnosis model in sequence. The charging pile fault diagnosis model deployment module is used by the server to compress the tested charging pile fault diagnosis model and deploy it to the charging pile. The fault diagnosis module is used to collect real-time operating status data during the operation of the charging pile. After preprocessing the real-time operating status data, it is input into the deployed charging pile fault diagnosis model to obtain fault diagnosis results that include at least the diagnosis conclusion, fault details, repair suggestions, associated fault risk warning, predictive maintenance suggestions, and confidence level. The fault repair module is used to automatically perform fault repair operations based on the fault diagnosis results of the charging pile and record the fault repair results; The charging pile fault diagnosis model iteration module is used to iteratively optimize the charging pile fault diagnosis model based on the fault repair results and real-time operating status data through a federated learning mechanism.

[0063] By constructing a closed-loop system integrating multimodal data perception, intelligent diagnosis, automatic execution, and continuous evolution capabilities, the intelligence level and overall efficiency of charging pile fault repair have been significantly improved. It collects and deeply integrates Flash partition health status, electrical parameter timing events, and system load indicators, utilizing an advanced multimodal timing fusion model for accurate diagnosis and predictive maintenance. Based on the fault diagnosis results, it dynamically generates repair strategies and executes them automatically. Finally, through a federated learning mechanism, the repair results are fed back to the model for continuous iterative optimization. This achieves a leap from passive response to proactive prediction, from static rules to dynamic learning, and from modular fragmentation to collaborative autonomy, effectively improving repair efficiency, system reliability, and operational economy.

[0064] The dataset construction module is specifically used for: the server to collect a large amount of historical operating status data from different charging piles, including at least Flash partition status information, electrical parameter timing events, and system load indicators; The Flash partition status information includes at least partition integrity verification data, storage medium physical status data, and partition usage and performance data. The partition integrity verification data includes at least CRC checksums (used for integrity verification of critical static data such as the boot sector and application firmware area), and file system metadata status (such as the integrity of inodes and the consistency of superblocks). The storage medium physical status data includes at least bad block information (including the number of marked bad blocks, the growth rate of bad blocks, and the physical location of bad block distribution (e.g., whether they are concentrated in critical data areas, which is the direct basis for determining whether "bad blocks in the data area exceed the threshold"), wear leveling statistics (the number of erases for each physical block; average wear level and maximum wear level), and data retention capability indication (indirectly determining whether data has decayed due to prolonged lack of refresh by reading the voltage level or bit error rate of the storage unit). The partition usage and performance data includes at least the available space ratio and fragmentation level (insufficient space or severe fragmentation will reduce write performance and exacerbate wear), and read / write operation latency (an abnormal increase in read / write speed may indicate storage medium performance degradation or controller failure). The electrical parameter timing events include at least core power circuit parameters, charging control and communication events, and protection device action events (action timing of fuses, contactors, and relays: recording their opening and closing times and performing correlation analysis with electrical parameter abnormal events to distinguish between real faults and false alarms); the core power circuit parameters include at least voltage and current waveforms (instantaneous overvoltage / undervoltage, overcurrent; voltage drop / surge, current spike; waveform distortion), and power module status (such as IGBT drive signals, switching frequency, radiator temperature, etc., used to diagnose "power module IGBT signal abnormalities"); the charging control and communication events include at least charging handshake protocol sequences (message timing and response time for communication with the vehicle BMS; faults such as "communication timeout with BMS" need to be judged by analyzing such event sequences), and network messages (timing, packet loss rate, and error frames of key control commands and status feedback messages). The system load metrics include at least the following: computing resource utilization, system process and service status, communication stack load, and task scheduling information (readiness delay and scheduling delay of real-time tasks; excessive delay may indicate high system load, affecting the real-time performance of critical control loops). The computing resource utilization includes at least CPU load (not only overall utilization, but also CPU usage of each task / process to identify "abnormal exit of critical processes" or resource preemption issues), and memory usage (total memory usage, available memory, and memory fragmentation; memory leaks are a common cause of system instability). The system process and service status includes at least the heartbeat of critical daemons (whether network services, payment services, monitoring agents, etc., are running normally) and the watchdog timer status ("watchdog timeout reset" is the final manifestation of excessive system load or deadlock). The communication stack load includes at least the number of network connections (the number of concurrent connections with the backend server, payment gateway, other stubs, etc.), network bandwidth utilization and packet throughput, and TCP / IP stack error statistics (such as retransmission rate, connection errors, etc., used to diagnose the root cause of "interrupted heartbeat connection with the server"). The historical operational status data are preprocessed, including at least data cleaning, data integration and alignment, and feature engineering. The preprocessed historical operational status data are annotated with at least diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, and predictive maintenance suggestions. A dataset is constructed based on the annotated historical operational status data. The data cleaning includes at least missing value handling (i.e., checking for missing values ​​such as null values, NaN (non-numeric), or specifically labeled missing values ​​in the dataset, and taking measures such as deletion or interpolation), and outlier handling (using statistical methods, business rule methods, or DBSCAN clustering to identify outliers, and taking measures such as correction, deletion, or retention and labeling). The data integration and alignment includes at least timestamp alignment (unifying all different types of data streams to the same time base; for data with different collection frequencies (such as millisecond-level voltage waveforms and second-level CPU loads), resampling is required), and data association (associating the collected status data with historical fault event records, which is the key to building a supervised learning dataset). The feature engineering includes at least feature extraction (e.g., extracting the bad block growth rate from "bad block information"; extracting the average erase / write count and maximum erase / write count from "wear leveling statistics"; calculating insufficient space warning from "available space ratio" (e.g., 1 if available space < 10%, otherwise 0); extracting the mean, standard deviation, peak value, valley value, and ripple coefficient from voltage and current waveforms). The system performs various functions, including: 1) 2) Fast Fourier Transform (FFT) to extract the main frequency component and harmonic components, which are highly effective for assessing the health status of power devices; 3) Encoding of the "charging handshake protocol sequence" and "protection device action events" to extract features such as protocol interaction timeout counts and protection action frequency; 4) Extracting sliding window statistics from CPU and memory data, such as the average load, peak load, and trend (slope) over the past 5 minutes; 5) Extracting the error rate from "TCP / IP stack error statistics"; and 6) Extracting the number of restarts from "watchdog timer status" and performing feature transformations (normalization / standardization). Normalization: Transform features of different dimensions and orders of magnitude (such as voltage values ​​of several hundred volts, CPU load of 0-100%) to the same scale. Common methods include Min-Max normalization (to the [0,1] interval) and Z-Score standardization (mean of 0, variance of 1), which can accelerate model convergence and improve stability; Discretization: Segment continuous values, such as dividing the number of erases in "wear balance statistics" into "low", "medium" and "high" risk levels; One-hot encoding: Convert categorical variables (such as "fault type") into binary vectors to facilitate model understanding.

[0065] In the charging pile fault diagnosis model creation module, the feature extraction layer is constructed based on the Flash feature extraction module, the electrical parameter feature extraction module, and the system load feature extraction module; The Flash feature extraction module is used to extract first local features from Flash partition state information through a first 1D convolutional neural network, and to integrate global temporal information into the first local features through a Transformer encoder to obtain a Flash feature sequence. The first 1D convolutional neural network is used to capture local temporal patterns of Flash data, such as short-term changes in partition integrity verification data. The Transformer encoder is used to capture long-term dependencies, such as trends in storage medium physical state data. The advantages include the ability to effectively handle the temporal dynamics and multi-scale features of Flash data.

[0066] The electrical parameter feature extraction module is used to extract second local features from electrical parameter time-series events through a second 1D convolutional neural network, and to capture first time dependencies from the second local features through a long short-term memory network (LSTM) to obtain an electrical parameter feature sequence; the second 1D convolutional neural network is used to extract high-frequency local features such as voltage and current waveforms; the LSTM is used to model long-term events in the sequence, such as the time context of protection device action events; the advantages include strong modeling capability for electrical parameter time-series events and adaptability to variable-length sequences.

[0067] The system load feature extraction module is used to extract third local features from system load indicators through a third 1D convolutional neural network, and to capture second time dependencies from the third local features through a gated recurrent unit (GRU) to obtain a system load feature sequence; the third 1D convolutional neural network is used to capture short-term fluctuations in indicators such as computing resource utilization; the GRU is used to efficiently learn the temporal evolution of system processes and service states; the advantages include high computational efficiency and suitability for processing real-time system load data.

[0068] The timing alignment layer is used to perform time alignment operations on Flash feature sequences, electrical parameter feature sequences, and system load feature sequences using a dynamic time warping algorithm, and uses deformable convolution to adjust sequence length and phase; advantages include the ability to handle modal data with different sampling rates and improve the accuracy of fusion.

[0069] The modal fusion layer is used to extract intermodal correlation features from the Flash feature sequence, electrical parameter feature sequence, and system load feature sequence output by the timing alignment layer through a cross-attention mechanism. The gating fusion unit dynamically adjusts the weights of each modal feature (Flash feature sequence, electrical parameter feature sequence, and system load feature sequence) based on the correlation features to obtain a fused feature sequence. The advantages include adaptive weight allocation and enhanced modal complementarity.

[0070] The feature enhancement layer is constructed based on a self-supervised contrastive learning module and a knowledge graph enhancement module. The self-supervised contrastive learning module is used to construct positive sample pairs (data with similar fault states) and negative sample pairs (data with different fault states) through a contrastive loss function, and to perform enhancement operations on the fused feature sequence to obtain a first-level enhanced feature sequence. The advantages include improving the robustness of features against noise and outliers without the need for additional annotation.

[0071] The knowledge graph enhancement module is used to interact with a preset knowledge graph of charging pile component relationships (such as fault propagation paths) and a first-level enhanced feature sequence through a graph neural network, and to integrate semantic information using an attention mechanism to obtain a second-level enhanced feature sequence. The advantages include injecting domain knowledge and improving the accuracy and interpretability of diagnosis.

[0072] The diagnostic prediction layer is constructed based on a multi-task diagnostic prediction module and an uncertainty estimation module; The multi-task diagnostic prediction module is used to aggregate the secondary enhanced feature sequences into a context vector through an attention pooling layer, and output diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, and predictive maintenance suggestions through multiple fully connected layers; the attention pooling layer is used to adaptively focus on key time steps; the fully connected layer is used to perform multi-task prediction; the advantages include shared feature representation, improved training efficiency and generalization ability.

[0073] The uncertainty estimation module is used to evaluate the confidence level of diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, and predictive maintenance suggestions using Monte Carlo dropout technology, and outputs fault diagnosis results carrying diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, predictive maintenance suggestions, and confidence levels; advantages include providing reliability indicators and enhancing the practicality of the model.

[0074] The diagnostic conclusions include fault status indicators (clearly determining the current system status, such as "normal", "warning", "fault") and fault types (Flash storage faults: such as "boot partition CRC check error", "data area bad block exceeds threshold", "file system index corruption"; power supply electrical faults: such as "DC side overvoltage protection", "AC input undervoltage", "power module IGBT overheating"; communication faults: such as "communication timeout with BMS", "heartbeat connection interruption with server"; software / system faults: such as "watchdog timeout reset", "critical process abnormal exit", "system overload"). The fault details include at least the faulty component (accurate to the smallest operable unit, such as: "Flash chip 2, block address 0x8000-0x8FFF", "current sensor of phase A charging module", "SSL layer of TCP / IP communication stack"), fault severity level (assessing the urgency of the fault's impact on the system, such as "fatal (system crash)", "serious (loss of function)", "moderate (performance degradation)", "warning (requires attention)"), and fault characteristic values ​​(providing specific values ​​that trigger diagnosis, such as: "peak current reaches 52A (threshold 50A)", "Flash block erase count is 98,752 times (warning threshold 100,000 times)", "CPU utilization rate is above 95% for 3 consecutive minutes"). The repair recommendations include at least recommended repair actions (providing specific executable instructions, such as: "Enable Flash redundant backup partition and perform master-slave switch"; "Remap bad blocks in the data area and attempt to repair the file system"; "Perform a soft reboot of the charging module and reduce the current charging power to 80%"; "Restart the network service process") and repair strategy parameters (providing parameterized guidance for the repair actions, such as: "After reboot, the charging current limit is 30A"; "Set the priority of data migration to high").

[0075] Associated fault risk warnings indicate secondary or associated faults that may be caused by the current fault; for example, when a "Flash boot sector error" is diagnosed, a warning should be issued that "the system may fail to start, causing all charging services to be interrupted"; when a "communication module fault" is diagnosed, a warning should be issued that "fault data cannot be reported, affecting remote monitoring".

[0076] Predictive maintenance recommendations provide forward-looking warnings; for example: "The current Flash wear rate is high, and the predicted remaining lifespan is 45 days. Preventive maintenance is recommended." or "The power module's heat dissipation efficiency has decreased, and overheating risk is expected in high-temperature environments."

[0077] Confidence level is a probability value given to the reliability of the fault diagnosis result, such as "diagnosis confidence level: 92%"; this helps to adopt a more conservative repair strategy (such as requesting manual intervention) when the confidence level is low.

[0078] The formula for the loss function is: ; Where L represents the loss value of the loss function; Indicates multitasking loss; λ represents the comparison loss; λ represents the hyperparameter of the weighted balancing comparison loss; K represents the total number of tasks, with a value of 5, representing the five prediction tasks: diagnostic conclusion, fault details, repair suggestions, associated fault risk warning, and predictive maintenance suggestions; k represents the task index. This represents the weight of the k-th task; represents the loss of the k-th task; c represents the category index; This represents the true label of the k-th task and the c-th category; B represents the probability distribution of the prediction for the c-th category of the k-th task; B represents the batch size; sim() represents the cosine similarity function; z i z represents the feature vector of the i-th sample; i+ represents the positive sample feature of the i-th sample; τ represents the temperature parameter, used to control the sharpness of the similarity distribution, with a value of 0.1; z j Let z represent the feature vector of the j-th sample. i For negative sample features.

[0079] The charging pile fault diagnosis model training module is specifically used for: The server uses k-fold cross-validation to divide the dataset into training, validation, and test sets in a ratio of 7:1.5:1.5. The charging pile fault diagnosis model is trained using the training set until the loss value of the loss function is less than a preset loss threshold. The F1-Score, precision, and recall are calculated using the validation set to validate the trained charging pile fault diagnosis model. If the validation fails, the training set is expanded and training continues. If the validation passes, AUC-ROC, AUC-PR, and accuracy are calculated using the test set to test the validated charging pile fault diagnosis model. If the test fails, the training set is expanded and training continues. If the test passes, training ends.

[0080] The charging pile fault diagnosis model deployment module is specifically used for: The server sequentially compresses the tested charging pile fault diagnosis model through knowledge distillation, dynamic pruning, and quantization techniques. After fine-tuning and performance verification of the compressed charging pile fault diagnosis model, it is deployed to the charging pile. The fault diagnosis module is specifically used for: During operation, the charging pile collects real-time operating status data. Through multi-threading technology, the real-time operating status data is preprocessed, including at least data cleaning, data integration and alignment. The preprocessed real-time operating status data and historical operating status data of the local machine in a preset period are input into the deployed charging pile fault diagnosis model for inference. The output includes at least the diagnosis conclusion, fault details, repair suggestions, associated fault risk warning, predictive maintenance suggestions and confidence level of the fault diagnosis result. The fault repair module is specifically used for: The charging pile analyzes the fault diagnosis results. When the diagnosis conclusion is a fault, it determines whether the repair suggestion has the execution authority and execution capability. If so, the fault repair operation is automatically executed based on the repair suggestion, and the fault repair result is recorded. If not, a fault notification is generated based on the fault diagnosis results, and the fault notification is pushed to the management terminal in real time. The fault repair results should include at least a summary of the repair execution, a detailed description of the repair operation process, post-repair verification data, resource consumption and impact assessment, correlation analysis and suggested revisions; The repair execution summary is a high-level overview of the repair actions and their final status, equivalent to a quick report, including: Original fault identifier: a unique ID or timestamp associated with the original "fault diagnosis result" that triggered this repair; Repair action trigger time: the precise time point at which the charging pile decided to execute the automatic repair; Executed repair operation: which one or more of the "repair suggestions" were specifically executed, such as: "performed a soft reboot of the charging module", "completed the Flash primary / backup partition switch"; Final repair status: the final result of this repair operation, usually an enumerated value.

[0081] The detailed repair process log records detailed data and event sequences during the repair process for in-depth analysis and problem diagnosis, including: Automated operation sequence logs: complete command-line logs or API call records of the execution of repair instructions, which are crucial for repairing failed debugging attempts; Key parameter snapshots: parameter settings of relevant systems during the execution of repair operations, such as: process status before reboot, number of bad blocks before switching partitions, adjusted power limit values, etc.; Manual intervention records (if any): if manual intervention is required or integrated during the repair process (e.g., administrator confirmation), the personnel involved, time, and operation details must be recorded.

[0082] Post-repair verification data records whether the repair is truly effective and needs to be verified with data. This part is the bridge connecting "action" and "effect", including: Post-repair system status re-inspection results: After the repair operation is completed, the charging pile will run a simplified diagnosis again, record the real-time status of key indicators, and compare it with that before the repair; Short-term stability monitoring summary: Within a short period of time after the repair (such as 5 minutes), the system continues to monitor relevant parameters and record whether there are any abnormalities again or whether it tends to stabilize.

[0083] Resource consumption and impact assessment records the impact of the repair operation itself on the system. It is used to evaluate the cost and optimization direction of the repair strategy, including: Repair time: The total time from the start of the repair to the completion of verification, which is important for evaluating the recovery time objective (RTO); System resource overhead: The CPU, memory, network bandwidth, etc. consumed in performing the repair operation (such as data migration, process restart); Business impact: Did the charging service get affected during the repair? For example: "The repair caused a 2-minute charging interruption" or "The repair was completed in the background with no business impact".

[0084] Correlation analysis and suggestion correction involve providing feedback and revising the initial diagnosis and suggestions based on the actual effects of the repair, demonstrating the system's learning ability. This includes: Diagnostic confidence verification: comparing the "confidence" of the "repair result" with the "diagnostic conclusion." If a high-confidence diagnosis fails to be repaired, it indicates that the model may have blind spots, which is an important negative feedback signal; Repair suggestion effectiveness rating: rating the effectiveness of the implemented "repair suggestions" (e.g., effective, moderately effective, ineffective). If the repair fails, new, practice-validated "repair suggestions" may be generated; Correlation risk warning verification: checking whether the previously predicted "correlation fault risk warning" has occurred to verify the accuracy of the risk prediction model.

[0085] The charging pile fault diagnosis model iteration module is specifically used for: The charging pile uses a federated learning mechanism to construct an incremental dataset based on the fault repair results and real-time operating status data at preset iteration intervals. The local charging pile fault diagnosis model is trained using the incremental dataset to obtain local model parameters. The local model parameters are then uploaded to the server for aggregation to obtain global model parameters fed back from the server. The local charging pile fault diagnosis model is updated based on the global model parameters to iteratively optimize the charging pile fault diagnosis model.

[0086] In summary, the advantages of this invention are: 1. A server collects a large amount of historical operational status data from different charging piles, including at least Flash partition status information, electrical parameter timing events, and system load indicators. This historical operational status data is preprocessed to construct a dataset. Then, the server creates a charging pile fault diagnosis model based on a feature extraction layer, timing alignment layer, modality fusion layer, feature enhancement layer, and diagnostic prediction layer. A loss function for the charging pile fault diagnosis model is set, and the dataset is divided into training, validation, and test sets for sequential training, validation, and testing. The tested charging pile fault diagnosis model is compressed and deployed to the charging piles. During operation, real-time operational status data is collected from the charging piles and input into the charging pile fault diagnosis model to obtain fault diagnosis results including at least diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, predictive maintenance suggestions, and confidence levels. The system automatically executes fault repair operations based on fault diagnosis results and records the repair results. Through a federated learning mechanism, the charging pile fault diagnosis model iteratively optimizes itself based on fault repair results and real-time operational status data. This involves integrating multimodal data such as Flash partition status information, electrical parameter timing events, and system load indicators into a dedicated charging pile fault diagnosis model, enabling accurate diagnosis and predictive early warning of complex faults. Furthermore, it automatically executes repair operations based on the fault diagnosis results and innovatively introduces a federated learning mechanism, allowing the charging pile fault diagnosis model to continuously self-optimize using local fault repair results. Ultimately, this forms an intelligent operation and maintenance system with accuracy, adaptability, and continuous evolution capabilities, fundamentally solving the problems of module fragmentation, rigid strategies, and lack of learning ability in traditional solutions. This significantly improves the intelligence level and overall efficiency of charging pile fault repair.

[0087] 2. By collecting multimodal historical operating status data, including Flash partition status information, electrical parameter timing events, and system load indicators, and performing preprocessing such as data cleaning, integration, alignment, and feature engineering, a high-quality dataset is constructed. This comprehensive and refined data collection and processing method ensures that the input information for fault diagnosis is rich and reliable, and can capture subtle anomalies in the operation of charging piles, thereby significantly improving the accuracy and robustness of fault diagnosis, reducing the risk of false alarms and missed alarms, and demonstrating the advantages of the data-driven approach.

[0088] 3. The charging pile fault diagnosis model integrates a feature extraction layer, a temporal alignment layer, a modality fusion layer, a feature enhancement layer, and a diagnostic prediction layer. It also integrates a variety of advanced technologies such as 1D convolutional neural networks, Transformer, LSTM, GRU, dynamic time warping, cross-attention mechanism, self-supervised contrastive learning, and knowledge graph. This multi-technology fusion architecture can effectively process temporal data and multimodal information, extract deep features, and enhance the model's generalization ability. The diagnostic results not only include fault conclusions but also provide details, suggestions, and confidence levels, thus improving the comprehensiveness and interpretability of fault diagnosis.

[0089] 4. Automatic fault repair operations based on fault diagnosis results have been implemented. The charging pile can analyze the fault diagnosis results, determine the feasibility of the repair suggestions, and automatically execute the repair or generate a notification. This automated process reduces the need for manual intervention, improves fault response speed and processing efficiency, and provides data support for subsequent optimization by recording fault repair results. This demonstrates the practicality of intelligent operation and maintenance and helps to reduce the downtime and maintenance costs of charging piles.

[0090] 5. By adopting a federated learning mechanism, charging piles can iteratively optimize the model locally based on real-time data and repair results. Only local model parameters are uploaded to the server for aggregation, rather than the original data. This method effectively protects user privacy and data security, while realizing distributed continuous improvement of the model, avoiding the risks brought by data centralization, and ensuring that the model can adapt to the personalized operating environment of different charging piles, thus improving the scalability and long-term applicability of the technology.

[0091] 6. The model is compressed through knowledge distillation, dynamic pruning, and quantization techniques, and fine-tuned and verified before deployment to ensure that the fault diagnosis model runs efficiently on resource-constrained charging pile equipment. This compression and optimization strategy balances model performance and computational efficiency, enabling complex AI models to achieve real-time inference on edge devices, improving the practicality and promotion value of the technology, and making it suitable for the low latency requirements of large-scale charging pile networks.

[0092] 7. The diagnostic prediction layer adopts a multi-task diagnostic prediction module and an uncertainty estimation module, which can simultaneously output diagnostic conclusions, fault details, repair suggestions, risk warnings and maintenance suggestions, and assess confidence level. This multi-task and uncertainty assessment design enhances the reliability and credibility of diagnostic results, provides decision support for users, reduces the possibility of blind operation, and is particularly suitable for safety-critical charging pile scenarios, improving the overall system robustness.

[0093] 8. By regularly updating the model based on incremental data through a federated learning mechanism, the charging pile fault diagnosis model is continuously and adaptively optimized. This iterative learning capability enables the model to improve itself over time and with changes in the operating environment, adapting to new fault modes and avoiding the problem of model obsolescence. It reflects the dynamic evolutionary characteristics of artificial intelligence technology in industrial applications and has long-term technical value.

[0094] 9. By comprehensively collecting multimodal operating data and constructing an advanced fault diagnosis model, high-precision and automated charging pile fault diagnosis and repair are achieved. Its advantages lie in the integration of multi-source feature extraction, temporal alignment and knowledge enhancement technologies to ensure diagnostic reliability, combined with model compression to achieve efficient edge deployment, and continuous optimization of the model through a federated learning mechanism while protecting data privacy. Ultimately, a highly efficient operation and maintenance system integrating intelligent early warning, autonomous repair and adaptive learning is formed, which significantly improves the availability and maintenance efficiency of charging piles.

[0095] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for repairing charging pile faults, characterized in that: Includes the following steps: Step S1: The server collects historical operating status data from different charging piles, including at least Flash partition status information, electrical parameter timing events, and system load indicators, and preprocesses each of the historical operating status data to construct a dataset. Step S2: The server creates a charging pile fault diagnosis model based on the feature extraction layer, temporal alignment layer, modality fusion layer, feature enhancement layer, and diagnostic prediction layer, and sets the loss function of the charging pile fault diagnosis model. Step S3: The server divides the dataset into a training set, a validation set, and a test set to train, validate, and test the charging pile fault diagnosis model in sequence. Step S4: After compressing the tested charging pile fault diagnosis model, the server deploys it to the charging pile. Step S5: During operation, the charging pile collects real-time operating status data. After preprocessing the real-time operating status data, it is input into the deployed charging pile fault diagnosis model to obtain fault diagnosis results that include at least the diagnosis conclusion, fault details, repair suggestions, associated fault risk warning, predictive maintenance suggestions, and confidence level. Step S6: The charging pile automatically performs fault repair operations based on the fault diagnosis results and records the fault repair results; Step S7: The charging pile uses a federated learning mechanism to iteratively optimize the charging pile fault diagnosis model based on the fault repair results and real-time operating status data. In step S2, the feature extraction layer is constructed based on the Flash feature extraction module, the electrical parameter feature extraction module, and the system load feature extraction module; The Flash feature extraction module is used to extract first local features from Flash partition state information through a first 1D convolutional neural network, and to integrate global time information into the first local features through a Transformer encoder to obtain a Flash feature sequence. The electrical parameter feature extraction module is used to extract second local features from electrical parameter time-series events through a second 1D convolutional neural network, and to capture first time dependencies from the second local features through a long short-term memory network to obtain an electrical parameter feature sequence. The system load feature extraction module is used to extract third local features from system load indicators through a third 1D convolutional neural network, and to capture second time dependencies from the third local features through a gated recurrent unit to obtain a system load feature sequence. The timing alignment layer is used to perform time alignment operations on the Flash feature sequence, electrical parameter feature sequence and system load feature sequence through a dynamic time warping algorithm, and uses deformable convolution to adjust the sequence length and phase. The modal fusion layer is used to extract intermodal correlation features from the Flash feature sequence, electrical parameter feature sequence and system load feature sequence output by the timing alignment layer through a cross-attention mechanism. The gating fusion unit dynamically adjusts the weights of each modal feature according to the correlation features to obtain a fused feature sequence. The feature enhancement layer is constructed based on a self-supervised contrastive learning module and a knowledge graph enhancement module. The self-supervised contrastive learning module is used to construct positive and negative sample pairs through a contrastive loss function, and to perform enhancement operations on the fused feature sequence to obtain a first-level enhanced feature sequence. The knowledge graph enhancement module is used to interact the preset charging pile component relationship knowledge graph with the first-level enhanced feature sequence through a graph neural network, and to integrate semantic information using an attention mechanism to obtain the second-level enhanced feature sequence. The diagnostic prediction layer is constructed based on a multi-task diagnostic prediction module and an uncertainty estimation module; The multi-task diagnosis and prediction module is used to aggregate each secondary enhanced feature sequence into a context vector through an attention pooling layer, and output diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings and predictive maintenance suggestions through multiple fully connected layers. The uncertainty estimation module is used to evaluate the confidence level of diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, and predictive maintenance suggestions using Monte Carlo dropout technology, and outputs fault diagnosis results carrying diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, predictive maintenance suggestions, and confidence levels; The diagnostic conclusion includes the fault status indicator and the fault type; The fault details include at least the faulty component, the fault severity level, and the fault characteristic values; The repair recommendations include at least the recommended repair actions and repair strategy parameters.

2. The charging pile fault repair method as described in claim 1, characterized in that: Specifically, step S1 involves the server collecting historical operating status data from different charging piles, including at least Flash partition status information, electrical parameter timing events, and system load indicators. The Flash partition status information includes at least partition integrity verification data, storage medium physical status data, and partition usage and performance data. The partition integrity verification data includes at least a CRC checksum and file system metadata status. The physical state data of the storage medium includes at least bad block information, wear leveling statistics, and data retention capability indicators; the partition usage and performance data includes at least the available space ratio and fragmentation level, and read / write operation latency. The electrical parameter timing events include at least core power circuit parameters, charging control and communication events, and protection device action events; the core power circuit parameters include at least voltage and current waveforms and power module status; the charging control and communication events include at least charging handshake protocol sequences and network messages. The system load metrics include at least computing resource utilization, system process and service status, communication stack load, and task scheduling information; the computing resource utilization includes at least CPU load and memory usage; the system process and service status includes at least the heartbeat of critical daemons and the watchdog timer status; the communication stack load includes at least the number of network connections, network bandwidth utilization and packet throughput, and TCP / IP stack error statistics. The historical operational status data are preprocessed, including at least data cleaning, data integration and alignment, and feature engineering. The preprocessed historical operational status data are annotated with at least diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, and predictive maintenance suggestions. A dataset is constructed based on the annotated historical operational status data. The data cleaning includes at least missing value processing and outlier processing; The data integration and alignment includes at least timestamp alignment and data association; the feature engineering includes at least feature extraction and feature transformation.

3. The charging pile fault repair method as described in claim 1, characterized in that: Step S3 specifically involves: The server uses k-fold cross-validation to divide the dataset into training, validation, and test sets in a ratio of 7:1.5:1.

5. The charging pile fault diagnosis model is trained using the training set until the loss value of the loss function is less than a preset loss threshold. The F1-Score, precision, and recall are calculated using the validation set to validate the trained charging pile fault diagnosis model. If the validation fails, the training set is expanded and training continues. If the validation passes, AUC-ROC, AUC-PR, and accuracy are calculated using the test set to test the validated charging pile fault diagnosis model. If the test fails, the training set is expanded and training continues. If the test passes, training ends.

4. The charging pile fault repair method as described in claim 1, characterized in that: Step S4 specifically involves: The server sequentially compresses the tested charging pile fault diagnosis model through knowledge distillation, dynamic pruning, and quantization techniques. After fine-tuning and performance verification of the compressed charging pile fault diagnosis model, it is deployed to the charging pile. Step S5 specifically involves: During operation, the charging pile collects real-time operating status data. Through multi-threading technology, the real-time operating status data is preprocessed, including at least data cleaning, data integration and alignment. The preprocessed real-time operating status data and historical operating status data of the local machine in a preset period are input into the deployed charging pile fault diagnosis model for inference. The output includes at least the diagnosis conclusion, fault details, repair suggestions, associated fault risk warning, predictive maintenance suggestions and confidence level of the fault diagnosis result. Step S6 specifically involves: The charging pile analyzes the fault diagnosis results. When the diagnosis conclusion is a fault, it determines whether the repair suggestion has the execution authority and execution capability. If so, the fault repair operation is automatically executed based on the repair suggestion, and the fault repair result is recorded. If not, a fault notification is generated based on the fault diagnosis results, and the fault notification is pushed to the management terminal in real time. Step S7 specifically involves: The charging pile uses a federated learning mechanism to construct an incremental dataset based on the fault repair results and real-time operating status data at preset iteration intervals. The local charging pile fault diagnosis model is trained using the incremental dataset to obtain local model parameters. The local model parameters are then uploaded to the server for aggregation to obtain global model parameters fed back from the server. The local charging pile fault diagnosis model is updated based on the global model parameters to iteratively optimize the charging pile fault diagnosis model.

5. A charging pile fault repair system, characterized in that: Includes the following modules: The dataset construction module is used by the server to collect historical operating status data from different charging piles, including at least Flash partition status information, electrical parameter timing events, and system load indicators, and to preprocess the historical operating status data to construct the dataset. The charging pile fault diagnosis model creation module is used by the server to create a charging pile fault diagnosis model based on a feature extraction layer, a temporal alignment layer, a modal fusion layer, a feature enhancement layer, and a diagnosis prediction layer, and to set the loss function of the charging pile fault diagnosis model. The charging pile fault diagnosis model training module is used by the server to divide the dataset into a training set, a validation set, and a test set, so as to train, validate, and test the charging pile fault diagnosis model in sequence. The charging pile fault diagnosis model deployment module is used by the server to compress the tested charging pile fault diagnosis model and deploy it to the charging pile. The fault diagnosis module is used to collect real-time operating status data during the operation of the charging pile. After preprocessing the real-time operating status data, it is input into the deployed charging pile fault diagnosis model to obtain fault diagnosis results that include at least the diagnosis conclusion, fault details, repair suggestions, associated fault risk warning, predictive maintenance suggestions, and confidence level. The fault repair module is used to automatically perform fault repair operations based on the fault diagnosis results of the charging pile and record the fault repair results; The charging pile fault diagnosis model iteration module is used to iteratively optimize the charging pile fault diagnosis model based on the fault repair results and real-time operating status data through a federated learning mechanism. In the charging pile fault diagnosis model creation module, the feature extraction layer is constructed based on the Flash feature extraction module, the electrical parameter feature extraction module, and the system load feature extraction module; The Flash feature extraction module is used to extract first local features from Flash partition state information through a first 1D convolutional neural network, and to integrate global time information into the first local features through a Transformer encoder to obtain a Flash feature sequence. The electrical parameter feature extraction module is used to extract second local features from electrical parameter time-series events through a second 1D convolutional neural network, and to capture first time dependencies from the second local features through a long short-term memory network to obtain an electrical parameter feature sequence. The system load feature extraction module is used to extract third local features from system load indicators through a third 1D convolutional neural network, and to capture second time dependencies from the third local features through a gated recurrent unit to obtain a system load feature sequence. The timing alignment layer is used to perform time alignment operations on the Flash feature sequence, electrical parameter feature sequence and system load feature sequence through a dynamic time warping algorithm, and uses deformable convolution to adjust the sequence length and phase. The modal fusion layer is used to extract intermodal correlation features from the Flash feature sequence, electrical parameter feature sequence and system load feature sequence output by the timing alignment layer through a cross-attention mechanism. The gating fusion unit dynamically adjusts the weights of each modal feature according to the correlation features to obtain a fused feature sequence. The feature enhancement layer is constructed based on a self-supervised contrastive learning module and a knowledge graph enhancement module. The self-supervised contrastive learning module is used to construct positive and negative sample pairs through a contrastive loss function, and to perform enhancement operations on the fused feature sequence to obtain a first-level enhanced feature sequence. The knowledge graph enhancement module is used to interact the preset charging pile component relationship knowledge graph with the first-level enhanced feature sequence through a graph neural network, and to integrate semantic information using an attention mechanism to obtain the second-level enhanced feature sequence. The diagnostic prediction layer is constructed based on a multi-task diagnostic prediction module and an uncertainty estimation module; The multi-task diagnosis and prediction module is used to aggregate each secondary enhanced feature sequence into a context vector through an attention pooling layer, and output diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings and predictive maintenance suggestions through multiple fully connected layers. The uncertainty estimation module is used to evaluate the confidence level of diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, and predictive maintenance suggestions using Monte Carlo dropout technology, and outputs fault diagnosis results carrying diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, predictive maintenance suggestions, and confidence levels; The diagnostic conclusion includes the fault status indicator and the fault type; The fault details include at least the faulty component, the fault severity level, and the fault characteristic values; The repair recommendations include at least the recommended repair actions and repair strategy parameters.

6. The charging pile fault repair system as described in claim 5, characterized in that: The dataset construction module is specifically used for: the server to collect historical operating status data from different charging piles, including at least Flash partition status information, electrical parameter timing events, and system load indicators; The Flash partition status information includes at least partition integrity verification data, storage medium physical status data, and partition usage and performance data. The partition integrity verification data includes at least a CRC checksum and file system metadata status. The physical state data of the storage medium includes at least bad block information, wear leveling statistics, and data retention capability indicators; the partition usage and performance data includes at least the available space ratio and fragmentation level, and read / write operation latency. The electrical parameter timing events include at least core power circuit parameters, charging control and communication events, and protection device action events; the core power circuit parameters include at least voltage and current waveforms and power module status; the charging control and communication events include at least charging handshake protocol sequences and network messages. The system load metrics include at least computing resource utilization, system process and service status, communication stack load, and task scheduling information; the computing resource utilization includes at least CPU load and memory usage; the system process and service status includes at least the heartbeat of critical daemons and the watchdog timer status; the communication stack load includes at least the number of network connections, network bandwidth utilization and packet throughput, and TCP / IP stack error statistics. The historical operational status data are preprocessed, including at least data cleaning, data integration and alignment, and feature engineering. The preprocessed historical operational status data are annotated with at least diagnostic conclusions, fault details, repair suggestions, associated fault risk warnings, and predictive maintenance suggestions. A dataset is constructed based on the annotated historical operational status data. The data cleaning includes at least missing value processing and outlier processing; The data integration and alignment includes at least timestamp alignment and data association; the feature engineering includes at least feature extraction and feature transformation.

7. A charging pile fault repair system as described in claim 5, characterized in that: The charging pile fault diagnosis model training module is specifically used for: The server uses k-fold cross-validation to divide the dataset into training, validation, and test sets in a ratio of 7:1.5:1.

5. The charging pile fault diagnosis model is trained using the training set until the loss value of the loss function is less than a preset loss threshold. The F1-Score, precision, and recall are calculated using the validation set to validate the trained charging pile fault diagnosis model. If the validation fails, the training set is expanded and training continues. If the validation passes, AUC-ROC, AUC-PR, and accuracy are calculated using the test set to test the validated charging pile fault diagnosis model. If the test fails, the training set is expanded and training continues. If the test passes, training ends.

8. A charging pile fault repair system as described in claim 5, characterized in that: The charging pile fault diagnosis model deployment module is specifically used for: The server sequentially compresses the tested charging pile fault diagnosis model through knowledge distillation, dynamic pruning, and quantization techniques. After fine-tuning and performance verification of the compressed charging pile fault diagnosis model, it is deployed to the charging pile. The fault diagnosis module is specifically used for: During operation, the charging pile collects real-time operating status data. Through multi-threading technology, the real-time operating status data is preprocessed, including at least data cleaning, data integration and alignment. The preprocessed real-time operating status data and historical operating status data of the local machine in a preset period are input into the deployed charging pile fault diagnosis model for inference. The output includes at least the diagnosis conclusion, fault details, repair suggestions, associated fault risk warning, predictive maintenance suggestions and confidence level of the fault diagnosis result. The fault repair module is specifically used for: The charging pile analyzes the fault diagnosis results. When the diagnosis conclusion is a fault, it determines whether the repair suggestion has the execution authority and execution capability. If so, the fault repair operation is automatically executed based on the repair suggestion, and the fault repair result is recorded. If not, a fault notification is generated based on the fault diagnosis results, and the fault notification is pushed to the management terminal in real time. The charging pile fault diagnosis model iteration module is specifically used for: The charging pile uses a federated learning mechanism to construct an incremental dataset based on the fault repair results and real-time operating status data at preset iteration intervals. The local charging pile fault diagnosis model is trained using the incremental dataset to obtain local model parameters. The local model parameters are then uploaded to the server for aggregation to obtain global model parameters fed back from the server. The local charging pile fault diagnosis model is updated based on the global model parameters to iteratively optimize the charging pile fault diagnosis model.

Images