An automobile charging pile abnormality identification method and system

Abstract

The application discloses a kind of automobile charging pile abnormality identification method and system, belong to new energy technical field, the method includes, first, acquisition multi-source data;Second, based on the multi-source data collected, respectively extract first feature data and second feature data;Then, based on the first feature data extracted, using automobile charging abnormal gun identification model to identify charging gun abnormality;Then, based on the second feature data extracted, using automobile charging abnormal field station identification model to identify field station abnormality;Finally, based on the charging gun abnormality and / or field station abnormality identified, generate executable decision scheme.Through setting two kinds of abnormality identification model, hierarchical abnormality identification is realized, the accuracy of abnormality identification is improved, decision scheme is formed based on identification result, and the operation and maintenance efficiency and operation and maintenance reliability are improved.

Description

Technical Field

[0001] This application belongs to the field of new energy technology, and specifically relates to a method and system for identifying abnormalities in car charging piles. Background Technology

[0002] With the popularization of new energy vehicles, the operation and maintenance technology of charging piles, as core infrastructure, has developed from traditional manual inspection to data-driven intelligent operation. The industry generally uses sensors, big data analysis and basic algorithms to achieve fault monitoring and diagnosis, forming a pattern of multiple technical paths in parallel.

[0003] Currently, the more common solutions can be divided into three categories: First, rule-based monitoring based on sensor data, which collects real-time parameters by deploying sensors such as voltage, current, and temperature, and sets fixed thresholds to trigger alarms, such as automatically alarming when the voltage exceeds the range of 210-230V. However, this type of solution relies on manually preset rules, which is slow to identify latent anomalies and is prone to false alarms due to rigid thresholds. Second, fault diagnosis based on basic machine learning, which uses algorithms such as KNN (K-Nearest Neighbors) and SVM (Support Vector Machine) to model historical fault data and determine anomalies by comparing the distance of feature samples. For example, the KNN algorithm is used to calculate the neighbor distance of three-phase current characteristics to identify faults. However, the models are mostly for single equipment dimensions and lack the ability to aggregate and analyze at the site level, making it difficult to adapt to widely distributed charging networks. Third, safety monitoring based on visual recognition, which uses high-definition cameras combined with image algorithms to identify visible risks such as smoke and equipment damage. Although it can quickly locate sudden safety hazards, it has no ability to identify electrical anomalies such as charging circuit faults and power attenuation, and needs to be used in conjunction with other systems.

[0004] Existing technologies still suffer from the following problems: data processing is mostly single-dimensional cleaning, lacking AI-driven automatic feature generation and filtering mechanisms; model collaboration is insufficient, failing to form a closed loop for hierarchical identification and root cause reasoning of equipment and sites; operation and maintenance decisions rely on manual dispatching, failing to incorporate reinforcement learning to achieve dynamic resource scheduling, and are particularly difficult to adapt to the operational needs of wide-area distribution and large-scale expansion under the integrated model. The overall integrity and intelligence of the technical solutions still have room for improvement. Specifically, the following existing technologies are examples: Chinese patent CN 118322919 A ​​discloses a method, device, equipment, medium, and program product for determining abnormal states of charging piles. This technical solution determines abnormalities based on the mapping relationship between fault frequency and fault type. Specifically, if a single fault exceeds a preset number of occurrences, the component is identified as abnormal; if multiple faults are concentrated and mapped to a certain abnormal category, that category is identified as abnormal. New rules for determining "continuously abnormal charging piles" and "highly abnormal charging piles" are added: the former is determined by a combination of the number of consecutive charging failures and a threshold number of users involved, while the latter is determined by a combination of fault rate, power supply, and the proportion of users experiencing charging failures. However, the above abnormality determination relies on manually preset fixed thresholds and lacks dynamic adaptation capabilities. For example, it does not provide corresponding abnormality determinations for charging piles in different regions or with different service years; it lacks a collaborative identification logic for single-pile abnormalities and site abnormalities, only handling single-device faults in isolation, without establishing aggregated analysis to locate the root cause at the site level; and the maintenance plan and operation and maintenance scheduling lack intelligence.

[0005] Chinese patent CN 118182216 A discloses a method, system, electronic device, and medium for handling anomalies based on charging piles. The method uses the user terminal as the trigger point and achieves anomaly handling through three-way interaction between the user terminal, the management center, and the agent terminal, emphasizing both user-initiated triggering and human agent intervention. It provides solutions for charging pile status query, nearby recommendations, and charging standard verification for two scenarios: charging anomalies and charging anomalies. However, this solution relies heavily on user-initiated anomaly identification, resulting in significant latency; it also suffers from high dependence on human agents, low processing efficiency, and lacks an iterative optimization mechanism after anomaly handling.

[0006] Chinese patent CN 108549955 A discloses a method and device for determining the anomaly rate of charging piles. Focusing on the prediction of charging pile anomaly rates, it employs a dual-path prediction approach using historical data-driven (AR(q) model) and multivariate regression, covering scenarios with both sufficient and insufficient data. It introduces multi-dimensional variables such as operation and maintenance and weather to improve the comprehensiveness of anomaly rate prediction. However, it only performs anomaly identification without closed-loop management of subsequent processing; furthermore, the model input features are singular and do not integrate multimodal data. Chinese patent CN117565724A discloses a method for identifying abnormal states of charging piles. It calculates the power variation coefficients of two segments; if the effective band variation coefficient exceeds a preset threshold, it determines an access anomaly; if the loss band variation coefficient exceeds a threshold, it determines a loss anomaly. Simultaneously, it identifies multi-pile shared-meter usage by checking for "power exceeding the reported installation by 1.2 times," and identifies peak-shifting users by checking the "percentage of high-power measurement points during high-load periods." It distinguishes between "access anomalies" and "core loss anomalies" based on the power variation coefficient; it adds rules for identifying multi-pile shared-meter usage and peak-shifting users, expanding the anomaly identification scenarios. However, the anomaly determination relies on a single power variation coefficient, failing to consider characteristics such as voltage stability and fault code correlation, making it prone to misjudgment; it lacks site-level anomaly identification capabilities, and the maintenance solution lacks root cause analysis.

[0007] Chinese patent CN 119575032 A discloses a method for detecting charging anomalies in charging piles. This method integrates charging efficiency and communication failure rate to calculate an anomaly score S, and determines an anomaly based on a threshold (S<0.6), triggering an early warning (e.g., immediately stopping use and notifying maintenance if the health status is low). The method quantifies the health index using voltage deviation rate, current fluctuation rate, temperature, and aging degree to achieve equipment health classification. The anomaly score S dynamically adjusts the aging influence coefficient W in the health index calculation formula, with the score and index mutually correcting each other. However, the weights of the health index and anomaly score are fixed, lacking scenario adaptability and the ability to dynamically adjust according to different scenarios. The evaluation accuracy is inconsistent across different scenarios, and the method lacks the operational and maintenance scheduling capabilities for large-scale charging pile networks.

[0008] In summary, existing technologies often suffer from insufficient flexibility and adaptability in anomaly detection. They rely on manually preset fixed thresholds or single models, failing to dynamically adapt to the different fault characteristics of charging piles in different regions and with varying service lives, leading to frequent misjudgments or missed detections. Collaborative identification and root cause reasoning between equipment and charging stations are lacking, with a focus primarily on single-pile anomaly detection and a lack of aggregated analysis capabilities for single-pile anomalies, making it impossible to pinpoint common problems at the station level. Furthermore, the absence of an anomaly root cause reasoning mechanism increases maintenance and troubleshooting costs. The anomaly identification and handling closed loop is incomplete, with maintenance scheduling relying on manual dispatching and failing to dynamically allocate resources based on the distance and skill matching of maintenance personnel, resulting in low response efficiency. Multimodal data fusion and feature engineering are weak, and a continuous iterative optimization mechanism is lacking. Summary of the Invention

[0009] To address the aforementioned issues, this application provides a method and system for identifying abnormalities in car charging piles, which features high operational efficiency and high reliability.

[0010] The purpose of this invention is to provide a method for identifying anomalies in car charging stations, including: Collect data from multiple sources; Based on the collected multi-source data, the first feature data and the second feature data are extracted respectively; Based on the extracted first feature data, an abnormal charging gun identification model is used to identify abnormal charging guns. Based on the extracted second feature data, an abnormal charging station identification model is used to identify station anomalies. Based on the identified charging gun anomalies and / or site anomalies, an actionable decision-making scheme is generated. Optionally, the multi-source data includes charging equipment data sources, charging station data sources, charging order data sources, and alarm data sources, among which, The charging equipment data sources include the charging gun's plug-in status, start / stop status, working status, current, voltage, and temperature, as well as the charging pile's power, operating time, and fault codes. Data sources for charging stations include the station's start / stop status, power supply data, equipment layout, number of charging piles, and operator information; The charging order data source includes detailed information about the charging order, including the order number, charging start and end time, charging amount, charging cost, and payment method. Alarm data sources include charging equipment fault alarms, security alarms, and abnormal transaction alarms.

[0011] Optionally, based on the collected multi-source data, first feature data is extracted, which includes the number of insertions and removals. Continuous charging duration characteristics Current fluctuation amplitude characteristics ΔI, voltage stability characteristics and statistics on the frequency characteristics of abnormal alarms The extraction methods are as follows: By statistically analyzing the number of times the charging gun is plugged in and unplugged within a certain time window T, the characteristics of the number of plugging and unplugging times are extracted. ; By recording the duration of continuous charging during each charging process, the continuous charging duration feature is extracted. ; By calculating the current fluctuation amplitude during the charging process, the current fluctuation amplitude feature ΔI is extracted: ΔI = max(I) min(I) In the formula, I is the current sequence during the charging process; Voltage stability characteristics are extracted using the following calculation formula. :

[0012] In the formula, It is the voltage value at the i-th sampling time. The average voltage is n, and the number of sampling points is n. By extracting statistical features of abnormal alarm frequency based on the number of abnormal alarms received by the charging gun within the time window T, we can identify these features. .

[0013] Optionally, based on the extracted first feature data, an abnormal charging gun identification model is used to identify charging gun anomalies. This includes using a deep learning model based on a long short-term memory network to identify charging gun anomalies, specifically including... The extracted first feature data are combined into a feature vector. And convert it into time series data: , where t represents the time step and m is the number of feature types; Time series data is input into a deep learning model based on a long short-term memory network, and the output is a two-dimensional probability vector y=[y1,y2], which represents the probability that the charging gun is in a normal state and the probability that it is in an abnormal state. Here, y1 is the probability that the charging gun is in a normal operating state, and y2 is the probability that the charging gun is in an abnormal operating state. The constraint relationship is: y1+y2=1. Set an anomaly probability threshold , when y2> This directly indicates that the charging gun is malfunctioning.

[0014] Optionally, it also includes training and optimizing the deep learning model based on the Long Short-Term Memory network, including, The cross-entropy loss function is used to measure the difference between the model's predictions and the true labels. The cross-entropy loss function is:

[0015] In the formula, N is the number of training samples. This is the true label of the i-th sample, where a value of 0 indicates normal and a value of 1 indicates abnormal. It is the probability of an anomaly predicted by the model for the i-th sample; The Adam optimizer is selected to update the parameters of the deep learning model based on the Long Short-Term Memory network, and the model is trained by minimizing the loss function.

[0016] Optionally, it also includes using operational experience to assist in determining charging gun malfunctions, specifically including: Acquire feature data, including abnormal orders, offline data, number of alarms, and demand fulfillment rate; Determine whether abnormal orders, offline data, alarm quantity, and demand fulfillment rate meet the corresponding preset abnormal conditions. If there is characteristic data that meets the preset abnormal conditions, then there is a charging gun abnormality, and the charging gun abnormality data is obtained.

[0017] Optionally, based on the collected multi-source data, second feature data is extracted, which includes average charging power features. Charging pile utilization characteristics Alarm Concentration Characteristics Power supply stability characteristics And the surrounding environmental factors E, the extraction methods are as follows: By calculating the average charging power of charging stations over a certain period of time, the average charging power characteristics are extracted. :

[0018] In the formula, is the charging power of the i-th charging pile, and n is the number of charging piles; By calculating the ratio of the number of charging piles actually in use to the total number of charging piles, the utilization rate of charging piles is extracted. :

[0019] In the formula, This represents the actual number of charging stations in use. This represents the total number of charging stations; By calculating the distribution concentration of alarms on different charging piles, alarm concentration characteristics are extracted. :

[0020] in, It represents the number of alarms for the i-th charging station. This represents the number of alarms for the j-th charging station. This is the average number of alarms; By calculating the standard deviation of the power supply, the characteristics of power supply stability can be obtained. :

[0021] In the formula, It is the power supply sequence of the i-th charging pile; The characteristics E of surrounding environmental factors are obtained based on ambient temperature and humidity.

[0022] Optionally, based on the extracted second feature data, an abnormal charging station identification model is used to identify station anomalies, including an ensemble learning model based on gradient boosting decision trees. The extracted second feature data is input, and the ensemble learning model based on gradient boosting decision trees outputs the original predicted values. In the formula, M is the number of iteration rounds; An ensemble learning model based on gradient boosting decision trees converts predicted values ​​into category labels L by setting a threshold, where L=1 indicates anomaly and L=0 indicates normality. Original predicted value threshold setting :

[0023] In the formula, The average of the original forecast values ​​for historically operating facilities. This represents the maximum value of the original forecast for historically operating facilities. These are the weighting coefficients; when > At that time, it was directly determined that there was an anomaly at the station.

[0024] Optionally, it also includes training and optimizing the ensemble learning model based on gradient boosting decision trees, including, Using the logarithmic loss function:

[0025] where y∈{ 1,1} are the true labels, and f(x) is the model prediction of the ensemble learning model based on gradient boosting decision tree; By adjusting the parameters of the gradient boosting decision tree ensemble learning model: the number of trees, the depth of the trees, and the learning rate β; Cross-validation is used to select the optimal combination of parameters.

[0026] Optionally, it also includes using station operation experience to assist in determining station anomalies, including, Collect the rated power data of each piece of equipment in the station as the benchmark value for judging power anomalies; Based on the start and end times of charging orders, filter scenarios where multiple charging guns are used simultaneously at the current power station; For scenarios involving simultaneous use, the earliest common usage time is determined by the intersection of the order times of each charging gun; and based on this time point, the current and voltage data of the corresponding order are matched from the data table and the real-time power is calculated. The maximum power of each charging gun under a single station is extracted and summed according to the station dimension. The power ratio is obtained by comparing the sum of the maximum power of a single station with the rated power of the station equipment. If the power percentage is less than the preset value, the site is considered abnormal.

[0027] Optionally, based on the identified charging gun anomalies and / or site anomalies, actionable decision-making solutions are generated, including generating maintenance work orders and operational optimization strategies, including... Based on geographical location, match the nearest maintenance personnel, and at the same time establish a skill tag library for maintenance personnel to match personnel with corresponding skills based on the type of abnormal charging gun; Generate a work order, which includes the work order number, site information, abnormal charging gun information, handling suggestions, time limit requirements, and feedback entry point; Generate a site optimization strategy report, which includes an overview of the anomaly, root cause analysis, specific strategies, expected results, responsible persons, and an implementation timeline, and push it to the site operation and management platform.

[0028] Optionally, it also includes visualization through a visualization module, including, Based on the map interface, it provides a basic map layer, a station distribution layer, and an abnormal charging gun distribution layer. Users can show or hide the corresponding content by checking the layer name, mark each abnormal point, display brief information when the mouse hovers over it, and pop up a details window when the point is clicked. It also provides point details query, which includes abnormal data, historical processing records, and current processing progress. The data dashboard interface displays the anomaly identification accuracy rate, work order completion rate, and site anomaly rate in real time. The large model question-answering module asks questions in natural language and generates structured answers.

[0029] Optionally, it also includes iterative optimization of the vehicle charging abnormality charging gun identification model and the vehicle charging abnormality site identification model, including, Collect feedback data, including operation and maintenance verification feedback data, user feedback data, and operation data within a preset time after the anomaly handling is completed; The collected feedback data is labeled; An incremental training algorithm is used. After training, the accuracy of the new model on the validation set is calculated. If the accuracy is improved by ≥ a predetermined percentage compared with the original model, the model parameters of the anomaly recognition module are updated. Otherwise, the training parameters are readjusted and the model is trained again. Then, the model is returned to the corresponding recognition model.

[0030] Another objective of this application is to provide an abnormal identification system for car charging stations, including, The data acquisition module is used to collect data from multiple sources. The data preprocessing module is used to extract the first feature data and the second feature data based on the collected multi-source data. The anomaly identification module is used to identify charging gun anomalies based on the extracted first feature data using a car charging anomaly charging gun identification model; and to identify station anomalies based on the extracted second feature data using a car charging station anomaly identification model. The management decision module is used to generate actionable decision-making schemes based on identified charging gun anomalies and / or site anomalies. Optionally, it also includes a visualization module, which includes a map interface, a data dashboard interface, and a large model Q&A module. The map interface provides a basic map layer, a site distribution layer, and an abnormal charging gun distribution layer. Users can show or hide the corresponding content by checking the layer names. Each abnormal point is marked, and brief information is displayed when the mouse hovers over it. Clicking on the point will pop up a details window. The interface also provides a point details query that includes abnormal data, historical processing records, and current processing progress. The data dashboard interface is used to display the anomaly identification accuracy rate, work order completion rate, and site anomaly rate in real time. The large-scale question-answering module is used to generate structured answers based on natural language questions.

[0031] Optionally, it also includes a model iterative optimization module for performing the following steps: Collect feedback data, including operation and maintenance verification feedback data, user feedback data, and operation data within a preset time after the anomaly handling is completed; The collected feedback data is labeled; An incremental training algorithm is used. After training, the accuracy of the new model on the validation set is calculated. If the accuracy is improved by ≥ a predetermined percentage compared with the original model, the model parameters of the anomaly recognition module are updated. Otherwise, the training parameters are readjusted and the model is trained again. Then, the model is returned to the corresponding recognition model.

[0032] Compared with the prior art, this application has the following advantages: This invention constructs an optimized technology system integrating data-driven and intelligent decision-making. It extracts feature data from multi-source data, inputs it into corresponding recognition models, and identifies anomalies, providing core technical support for efficient operation and maintenance and high-quality expansion of charging services. Furthermore, it establishes a complete anomaly handling closed loop to improve operation and maintenance response efficiency. Specifically, this invention dynamically matches resources based on the distance, skills, and load of operation and maintenance personnel, automatically generating tiered work orders containing anomaly details, handling suggestions, and time limits. This effectively shortens the response time for first-level anomaly work orders, significantly improves work order completion rates, and substantially enhances user experience.

[0033] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 A schematic flowchart of an abnormal identification method for car charging piles in an embodiment of this disclosure is shown. Figure 2 A schematic flowchart of a data preprocessing method according to an embodiment of this disclosure is shown; Figure 3 A flowchart illustrating an embodiment of this disclosure for generating an executable decision scheme is shown. Figure 4 A schematic diagram of a visualization method according to an embodiment of this disclosure is shown; Figure 5 A schematic diagram of a model iterative optimization process according to an embodiment of this disclosure is shown; Figure 6 A schematic diagram of an abnormal identification system for a car charging station according to an embodiment of this disclosure is shown. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0037] like Figure 1 As shown in the figure, this disclosure presents a method for identifying anomalies in car charging piles. The method includes: first, collecting multi-source data; second, extracting first feature data and second feature data based on the collected multi-source data; then, using the extracted first feature data, identifying charging gun anomalies using a car charging anomaly charging gun identification model; then, using the extracted second feature data, identifying charging station anomalies using a car charging station anomaly identification model; and finally, generating an executable decision plan based on the identified charging gun anomalies and / or charging station anomalies. This method constructs an optimized technical system integrating data-driven and intelligent decision-making. By processing multi-source data through feature engineering and inputting it into the corresponding identification model, it obtains anomaly information, providing core technical support for the efficient operation and high-quality expansion of charging services.

[0038] Specifically, multi-source data acquisition achieves data collection and full coverage through device sensors, smart terminals, and third-party interfaces. For example, Table 1 shows the data sources for charging equipment and charging stations.

[0039] Table 1 Data Sources for Charging Equipment and Charging Stations

[0040] It should be noted that order data and alarm data are obtained on the business platform, i.e., through user-end mini-programs, etc. Furthermore, the collected data is transmitted and stored. The MQTT (Message Queuing Telemetry Transport) protocol is used for data transmission between the sensors and the cloud. This protocol is lightweight, has low bandwidth consumption, and is suitable for real-time data transmission. User app data is transmitted using the HTTPS (Hypertext Transfer Protocol Secure) protocol to ensure data security. Then, edge computing preprocessing is employed. An edge computing gateway is deployed at each site to perform preliminary filtering of the collected real-time data, reducing the data processing pressure on the cloud. The gateway's cache capacity supports 72 hours of offline data storage to prevent data loss due to network interruptions. Finally, a time-series database is used to store time-sensitive data such as charging gun voltage, current, and power. A relational database stores structured data such as user order information, maintenance work orders, and site basic information, while an object database stores unstructured data such as user complaint texts, fault code screenshots, and raw environmental monitoring logs. Uniqueness is ensured through UUID (Universally Unique Identifier) ​​naming.

[0041] In this embodiment of the disclosure, big data related to car charging piles is collected extensively from multiple data sources. The multi-source data includes charging equipment data sources, charging station data sources, charging order data sources, and alarm data sources. Among them, the charging equipment data sources include the charging gun's plug-in status, start / stop status, working status, current, voltage, and temperature, as well as the charging pile's power, running time, and fault codes, covering the real-time operating data of the charging gun and the charging pile. This data can be collected in real time through the sensors and monitoring modules built into the charging equipment.

[0042] The data source for charging stations includes overall operational information, such as the station's start / stop status and power supply status; it also collects static data such as equipment layout, number of charging piles, and operator information. This data can be provided by the station's monitoring system and environmental monitoring equipment. The charging order data source includes detailed information about the charging orders, such as order number, charging start and end time, charging amount, charging cost, and payment method. This data records users' charging behavior and transaction details and can be obtained from the charging operation management system. The alarm data source is a collection of various charging-related alarm information, including charging equipment fault alarms, safety alarms, and abnormal transaction alarms. The alarm data source can come from the charging equipment's own alarm system, safety monitoring system, and transaction risk control system. Currently, different alarm levels can be set. For example, Level 1 alarms: faults endangering safety such as charging gun short circuits, voltage surges / dips >20%, and equipment smoke, must be pushed to the maintenance personnel's terminal in real time; Level 2 alarms: faults affecting operation such as charging power fluctuations >10%, gun body temperature >43℃, and offline time >20 minutes, pushed within 5 minutes; Level 3 alarms: faults that can be delayed, such as offline time <10 minutes and minor current fluctuations (±5%), summarized and pushed within 30 minutes. The above alarm threshold settings are not limited to these; adjusting the thresholds according to the application is applicable to this disclosure.

[0043] like Figure 2 As shown, the method also includes data preprocessing, including data cleaning, data integration, and feature engineering.

[0044] Data cleaning consists of three main parts: missing value imputation, outlier removal, and format correction. The specific operations are as follows: (1) Fill missing values ​​and supplement different data values ​​for different data sources: For charging gun operation data, linear interpolation is used to fill the missing values. If more than 5 data points are missing consecutively, the average value of the same type of charging gun in the same period is used to fill the missing values. If user behavior data is missing, it is filled by linking the order number to the payment platform interface. If it cannot be filled, it is marked and removed.

[0045] (2) Remove outliers. Process numerical data based on the 3σ principle. Calculate the mean μ and standard deviation σ of a certain indicator (such as charging power) and remove data that exceeds the range of [μ-3σ, μ+3σ]. Data with logical anomalies is removed based on business rules. For example, data with a charging time of less than 1 minute and a charging amount of more than 100 yuan does not conform to the actual charging scenario and is directly marked as abnormal and removed.

[0046] (3) Correct the format, including unifying the time format, converting all time data to UTC (Coordinated Universal Time) timestamps to avoid data confusion caused by time zone differences; standardizing the text format, removing special characters from user complaint texts and uniformly converting them to UTF-8 encoding (8-bit Unicode Transformation Format, a variable-length Unicode character encoding method); and unifying the format of fault codes.

[0047] Data integration includes establishing a correlation index, using timestamps and device IDs as core correlation keys, and data alignment to align data from different collection frequencies to the same time granularity.

[0048] Feature engineering involves the initial extraction, encoding, and standardization of features.

[0049] (1) Preliminary feature extraction: The initially extracted features include charging gun time-series features, trend features, and site aggregation features. Charging gun time-series features include short-term fluctuation features, obtained by calculating the standard deviation of power fluctuations per unit time and counting the number of consecutive charging failures within a specified time window; short-term fluctuation features quantify the stability of the charging process. Trend features are obtained by calculating the average voltage change rate within a specified time period and counting the proportion of orders with abnormal charging durations in the total orders; this feature captures the long-term trend of the charging process. Site aggregation features are obtained by calculating the proportion of abnormal charging guns per day to the total number of guns, and the proportion of abnormal gun (i.e., abnormal charging gun) runtime to the total runtime; these calculation results reflect the overall health status of the site equipment. In addition, the average charging waiting time during peak hours is also calculated, combined with user complaint rates, to assess the operational service quality of the site. Initial feature extraction is performed on the collected multi-source data to lay the foundation for subsequent feature extraction before model calculation. Feature extraction during model calculation involves further screening and refinement of the feature data, improving data reliability and further enhancing recognition accuracy.

[0050] (2) Feature coding and standardization: Categorical feature encoding: For categorical variables such as fault codes and region codes, one-hot encoding is used, with the encoding dimension consistent with the number of values ​​for the categorical feature, transforming discrete data into a numerical form suitable for machine learning models.

[0051] Numerical feature standardization: Min-Max standardization (deviation standardization) is used to map numerical features to the [0,1] interval. The formula is as follows: Standardized eigenvalue = (Original eigenvalue - Minimum eigenvalue) / (Maximum eigenvalue - Minimum eigenvalue) The maximum and minimum values ​​of the features are determined by statistical analysis of the entire historical data to ensure data scale consistency and improve model training efficiency and accuracy.

[0052] The aforementioned data preprocessing enhances multimodal data fusion and AI (Artificial Intelligence) feature engineering, thereby improving the coverage of data dimensions.

[0053] In this embodiment of the disclosure, obtaining the first feature data based on the collected multi-source data includes, The number of times a charging gun is plugged in and unplugged within a certain time window T is counted to reflect the frequency of use of the charging gun and to obtain the characteristics of the number of plugging and unplugging. This feature reflects the frequency of use of the charging gun; excessively high or low plugging / unplugging counts may indicate an abnormal situation.

[0054] Record the duration of continuous charging by the charging gun during each charging process to obtain continuous charging duration characteristics. An unusually long continuous charging time may indicate a problem with the charging gun or the vehicle.

[0055] Calculate the current fluctuation amplitude during charging and obtain the current fluctuation amplitude characteristic ΔI: ΔI = max(I) min(I) In the formula, I is the current sequence during the charging process; a large current fluctuation may indicate a circuit malfunction in the charging gun.

[0056] Voltage stability is measured using the standard deviation of voltage to obtain voltage stability characteristics. :

[0057] In the formula, It is the voltage value at the i-th sampling time. The average voltage is n, and the number of sampling points is n. Poor voltage stability may cause damage to the charging gun or abnormal charging.

[0058] Based on the number of abnormal alarms received by the charging gun within the time window T, obtain the abnormal alarm frequency characteristic statistics. Frequent alarms are usually a sign that there is a potential problem with the charging gun.

[0059] The method of identifying charging gun malfunctions using a car charging gun malfunction detection model includes a deep learning model based on Long Short-Term Memory (LSTM) networks. Specifically, LSTM (Long Short-Term Memory) is a special type of recurrent neural network (RNN) suitable for processing time-series data and capable of capturing long-term dependencies in the charging process. Its structure includes... Input layer: Combines the extracted features into a feature vector. As input. For time series data, the input format is... Where t represents the time step and m is the number of features. This disclosure allows the extracted first feature data to be combined into a feature vector. Converting data to time series involves the following steps: The feature vectors are concatenated sequentially along the time dimension. First, the time granularity is determined (e.g., data is collected once every 1 minute), and multiple sets of the first feature vectors are continuously collected. ; Arrange the first feature vectors collected at different time steps in chronological order: the feature vector of the first time step (t=1) is... ,correspond arrive (m=5, i.e., 5 features); the feature of the second time step (t=1) is ,correspond arrive .

[0060] Finally, the data is pieced together to form a time series: , where t represents the time step, for example, if 10 minutes of data are collected, t=10, and m is the number of feature types.

[0061] LSTM layer: An LSTM unit contains an input gate, a forget gate, and an output gate. Its core computation process is as follows: Input Gate:

[0062] Forgotten Gate:

[0063] Output gate:

[0064] Cell renewal status:

[0065] New cell state:

[0066] Hidden state:

[0067] Where b is the bias vector, σ is the sigmoid function (a sigmoid monotonically increasing activation function), and ⊙ denotes element-wise multiplication. It is the hidden state from the previous moment. This represents the cell state at the previous time step, and W is the weight matrix. Representing input features We learn the influence of input features on "whether new information is allowed to enter" by constructing the weight matrix of the input gate. Indicates the hidden state in the previous moment. We learn the influence of historical information on the current opening and closing of the input gate by analyzing the weight matrix of the input gate. Representing input features We learn the influence of input features on "whether historical information is forgotten" by using the weight matrix of the forget gate. Indicates the hidden state in the previous moment. We learn the influence of historical information on the current opening and closing of the forget gate by examining the weight matrix of the forget gate. Representing input features The weight matrix of the output gate is used to learn the influence of input features on "whether to output the current cell state". Hidden state in the previous moment The weight matrix to the output gate; Input features The weight matrix for cell renewal states; Hidden state in the previous moment The weight matrix for cell renewal state.

[0068] Output layer: A fully connected layer combined with a softmax activation function is used to output the probability of the charging gun being in a normal or abnormal state. Assume the output of the output layer is y=[y1,y2], where y1 is the probability of the normal state, y2 is the probability of the abnormal state, and y1+y2=1.

[0069] Calculated using the softmax function: , i=1,2, where z i z j is the original output value of the fully connected layer, i is the category index, so i=1 corresponds to the normal state and i=2 corresponds to the abnormal state; j is the loop variable for summation.

[0070] Specifically, the extracted first feature data is combined into a feature vector. And convert it into time series data: , where t represents the time step, m is the number of feature types, and the first feature data has a total of 5 feature data types, i.e. m=5; Time-series data is input into the aforementioned deep learning model based on a Long Short-Term Memory (LSTM) network, which outputs the probabilities of the charging gun being in a normal or abnormal state. Further, during anomaly detection, the output layer of the LSTM model outputs a two-dimensional probability vector y=[y1,y2] through a softmax activation function, where y1 is the probability of the charging gun being in a normal operating state (range: [0,1]); y2 is the probability of the charging gun being in an abnormal operating state (range: [0,1]); and the constraint is y1+y2=1. Then, the anomaly detection threshold and rules are applied. To avoid false positives and false negatives, and based on the actual operating scenarios of the charging gun, the following judgment criteria were set through historical anomaly data verification and cross-validation optimization: a default anomaly probability threshold was set. However, it is not limited to this and can be dynamically adjusted according to the accuracy requirements of site operation and maintenance. For example, it can be reduced to 0.25 for safety-sensitive scenarios and increased to 0.35 for efficiency-priority scenarios. Anomaly detection rule: When the model output y2> That is, when the probability of an anomaly exceeds the threshold, the charging gun is directly determined to be abnormal.

[0071] In this embodiment of the disclosure, the method further includes training and optimizing a deep learning model based on a long short-term memory network, including: The cross-entropy loss function is used to measure the difference between the model's predictions and the true labels. The cross-entropy loss function is:

[0072] In the formula, N is the number of training samples. It is the true label of the i-th sample (0 indicates normal, 1 indicates abnormal). It is the probability of an anomaly predicted by the model for the i-th sample; The Adam optimizer is chosen to update the model's parameters, and the model is trained by minimizing the loss function. The parameter update formula for the Adam optimizer is as follows: Calculate the gradient:

[0073] First-order moment estimation:

[0074] Second-order moment estimation:

[0075] Deviation correction:

[0076] Parameter update:

[0077] Where θ is the model parameter, α is the learning rate, and β1 and β2 are the decay rates. It is a small constant to avoid the denominator being zero, and t is the number of training steps.

[0078] This embodiment of the disclosure also includes using operational experience to assist in determining charging gun malfunctions. Specifically, this includes: first, acquiring feature data, including abnormal orders, offline data, alarm counts, and demand fulfillment rates; then, determining whether the abnormal orders, offline data, alarm counts, and demand fulfillment rates all meet preset conditions. If any feature does not meet the preset conditions, it indicates that a charging gun malfunction exists, and charging gun malfunction data is acquired. For example, Table 2 shows the key indicators in the judgment process.

[0079] Table 2 Key Indicators for Auxiliary Judgment of Charging Gun Abnormalities

[0080] The system directly uses historical order data, alarm data, and other relevant data for more accurate determination, and serves as a backup logic to supplement data when identifying abnormal guns.

[0081] In this embodiment of the disclosure, the extraction of second feature data based on collected multi-source data includes, Calculate the average charging power of charging stations over a certain period of time to obtain the characteristics of average charging power. :

[0082] In the formula, is the charging power of the i-th charging pile, and n is the number of charging piles; abnormal changes in the average charging power may reflect problems with the power supply or equipment of the charging station.

[0083] Calculate the ratio of the number of charging piles actually in use to the total number of charging piles to obtain the charging pile utilization rate characteristics. :

[0084] In the formula, This represents the actual number of charging stations in use. This refers to the total number of charging piles; both excessively low and excessively high utilization rates may indicate abnormal station operations.

[0085] Calculate the distribution concentration of alarms on different charging piles to obtain alarm concentration characteristics. :

[0086] in, It represents the number of alarms for the i-th charging station. This represents the number of alarms for the j-th charging station. This represents the average number of alarms; a high concentration of alarms may indicate that some charging stations have common problems.

[0087] The power supply stability characteristics are obtained based on the standard deviation of the power supply. :

[0088] In the formula, It is a power supply sequence; unstable power supply may affect the normal operation of the entire station.

[0089] Based on ambient temperature and humidity, the characteristics E of surrounding environmental factors are obtained. Environmental data can be quantified and encoded as feature input.

[0090] Based on the extracted second feature data, the vehicle charging station anomaly identification model is used to identify station anomalies. This includes an ensemble learning model based on Gradient Boosting Decision Tree (GBDT) to identify station anomalies. GBDT is an ensemble learning algorithm based on decision trees, which builds a strong learner by iteratively training multiple decision trees.

[0091] Initialize the model: First, initialize a constant prediction value. ,in It is a loss function. It's a real label. It is a constant.

[0092] Iterative training: In each iteration m, calculate the negative gradient (residual): ,in This is the prediction from the previous model. Fit a decision tree. To predict these residuals, we minimize the loss function: ,in It is the learning rate, used to control the contribution of each tree.

[0093] Update the model:

[0094] Prediction: The final model prediction is , where M is the number of iterations. For classification problems, the predicted value can be converted into a category label by setting a threshold.

[0095] In this embodiment of the disclosure, the extracted second feature data is input, and the ensemble learning model based on gradient boosting decision tree outputs the original predicted value. In the formula, M is the number of iteration rounds; the continuous value output after M rounds of decision tree iteration has no fixed range and reflects the quantitative score of the degree of anomaly of the site. The larger the value, the higher the anomaly risk.

[0096] An ensemble learning model based on gradient boosting decision trees converts predicted values ​​into category labels L by setting a threshold, where L=1 indicates anomaly and L=0 indicates normality. Based on the overall operational characteristics of the charging stations (such as the number of charging piles, average daily order volume, and power supply capacity), and through industry benchmark data calibration and multi-scenario testing, tiered judgment criteria are established.

[0097] Original predicted value threshold setting :

[0098] In the formula, The average of the original forecast values ​​for historically operating facilities. This represents the maximum value of the original forecast for historically operating facilities. These are the weighting coefficients. It can be set to 0.7, but is not limited to this setting; when > At that time, it was directly determined that there was an anomaly at the station.

[0099] In this embodiment of the disclosure, the method further includes training and optimizing the ensemble learning model based on gradient boosting decision trees, including: For classification problems, the log loss function is used:

[0100] where y∈{ {1,1} are the true labels, and f(x) is the model prediction; The performance of the ensemble learning model based on gradient boosting decision trees is optimized by adjusting the parameters, including the number of trees, the depth of the trees, and the learning rate β. Cross-validation is used to select the optimal parameter combination. For example, the training data is divided into k folds, and different folds are used as the validation set in turn, while the remaining folds are used as the training set. The average validation loss is calculated to evaluate the performance of different parameter combinations, where k is an integer.

[0101] By implementing two anomaly detection models, the flexibility and accuracy of anomaly detection are improved, reducing the false positive and false negative rates. Specifically, an LSTM deep learning model captures timing anomalies in charging guns, while a GBDT model analyzes aggregated anomalies at the charging station. Incremental learning is used to dynamically optimize model parameters and thresholds, adapting to regional differences and variations in the failure frequency of new and old charging piles without manual rule adjustments. This effectively improves anomaly detection accuracy and significantly reduces the false positive rate, minimizing ineffective maintenance due to false positives and the amplification of equipment failures due to false negatives.

[0102] In this embodiment of the disclosure, the method further includes using operational experience to assist in determining station anomalies. This includes: first, collecting basic rated power data of each device at the station as a benchmark for power anomaly determination; second, filtering scenarios where multiple charging guns are used simultaneously at the current station based on the start and end times of charging orders; then, for simultaneous use scenarios, determining the earliest common usage time point by the intersection of the order times for each charging gun; then, based on this time point, matching the current and voltage data of the corresponding orders from the data table and calculating the real-time power, extracting the maximum power value of each charging gun at a single station, and summing them by station dimension; then, comparing the sum of the maximum power values ​​at a single station with the rated power of the station equipment to obtain the power percentage; finally, if the power percentage is less than a preset percentage, the station is considered abnormal.

[0103] For example, as shown in Table 3, decision-making models can be supplemented based on operational experience.

[0104]

[0105] Furthermore, as shown in Table 3, after identifying a site anomalies according to the judgment rules, the number of anomalies under different abnormal terminals can also be determined. This auxiliary judgment, based directly on platform historical orders, alarms, and other relevant data, is more accurate and serves as a fallback logic to supplement data when identifying abnormal sites. Further, the GBDT model inputs features such as average charging power, alarm Gini coefficient, and power supply stability, and combines these with a second judgment based on the rule that the full load power ratio is ≤95%, thus providing a technical solution for locating the root cause of site-level anomalies. It should be noted that charging gun malfunctions and site malfunctions are interconnected, with local and overall issues being linked. Charging gun malfunctions are local equipment malfunctions, while site malfunctions are overall malfunctions. Multiple charging guns malfunctioning consecutively or in a concentrated manner will trigger the determination of site malfunctions. Site malfunctions may also cause multiple charging guns to malfunction simultaneously.

[0106] The two identification models described above enable collaborative identification and accurate root cause reasoning for equipment and sites, reducing maintenance and troubleshooting costs. They construct a collaborative architecture between a single-gun anomaly identification model and a site anomaly identification model, locating site-level issues through the percentage of anomalies in individual piles and deviations in operational indicators. Simultaneously, a lightweight large model is introduced, which outputs root cause probabilities and handling suggestions simply by inputting anomaly characteristics and historical records. Maintenance personnel can identify faulty components and solutions without on-site inspections, significantly shortening the time required for each maintenance check and reducing labor costs.

[0107] In this embodiment of the disclosure, such as Figure 3 As shown, based on the identified charging gun anomalies and site anomalies, an executable decision-making solution is generated, including generating maintenance work orders and operation optimization strategies, including... The work order generation unit matches the nearest maintenance personnel based on geographical location and establishes a skill tag library for maintenance personnel, matching personnel with corresponding skills based on the type of faulty gun. To consider load balancing, if multiple maintenance personnel meet the distance and skill requirements, the personnel with the fewest current work orders are selected to avoid a backlog of work orders for any single individual. Then, a work order is generated, including the work order number, site information, faulty gun information, handling suggestions, time limits, and feedback entry point. Finally, notifications are sent via the system app and SMS, and push logs are recorded in the cloud-based work order system to ensure traceability.

[0108] The strategy optimization unit generates a site optimization strategy report, including an overview of the anomaly, root cause analysis, specific strategies, expected results, responsible persons, and an execution schedule, and pushes it to the site operation management platform. For example, the site optimization strategy report is generated in PDF format. Preferably, the large model automatically generates optimization solutions. For example, the input is: "root cause of site anomalies (e.g., concentrated equipment aging) + site size (e.g., 20 guns) + historical optimization records"; the output is: a customized solution (e.g., "1. Prioritize replacing 10 guns that have been in use for more than 5 years; 2. Add one gun aging inspection per month; 3. Conduct maintenance during off-peak hours (0-6 am)"), and automatically generates a PDF "Site Optimization Execution Manual"; verification: after the solution is generated, the large model compares the optimization effects of similar historical sites and provides expected improvement indicators (e.g., "expected site anomaly rate to decrease from 25% to 8%)". Thus, the above methods construct a complete closed loop for anomaly handling, improve the efficiency of operation and maintenance response, and dynamically match resources based on the distance, skills, and load of operation and maintenance personnel. This disclosure automatically generates hierarchical work orders containing anomaly details, handling suggestions, and time limit requirements, effectively shortening the response time of first-level anomaly work orders, significantly improving the work order completion rate, and significantly enhancing the user experience.

[0109] In this embodiment of the disclosure, the method further includes visualization, employing a dual-interface design of "GIS (Geographic Information System) map + data dashboard," supporting access from both web and mobile devices. The interface layout is shown in Figure 4 (visual interface layout diagram). The map interface includes a basic map layer, a station distribution layer, and an anomaly gun distribution layer. Users can show or hide corresponding content by selecting layer names. Anomaly point markings are provided for each anomaly point; brief information is displayed when the mouse hovers over the point, and a details window pops up when the point is clicked. Point details queries include anomaly data, historical processing records, and current processing progress. Then, the data dashboard interface includes real-time display of core indicators, wherein: Anomaly detection accuracy = (Number of anomalies manually verified as true / Total number of anomalies identified by the model) × 100%; Work order completion rate = (Number of work orders completed within the time limit / Total number of work orders) × 100%; Station anomalous rate = (Number of stations with current anomalous issues / Total number of stations) × 100%; Additionally, this section supports exporting detailed exception reports.

[0110] Furthermore, the large-scale model question-answering module generates structured answers based on natural language questions. For example, users can ask questions in natural language (such as "Why has the anomaly rate at this site surged in the past 3 days?" or "How can we reduce the detection of charging gun anomalies?"); the response logic involves the large-scale model associating anomaly identification logs, work order records, and historical data to generate structured answers (such as "The anomaly rate at this site surged due to oxidation of the contact points of 3 charging guns; it is recommended to prioritize the handling of hardware maintenance personnel"), and voice interaction is supported.

[0111] like Figure 5 As shown, the method further includes iterative optimization of the vehicle charging anomaly gun identification model and the vehicle charging anomaly site identification model. This includes: first, collecting feedback data, including operation and maintenance verification feedback data, user feedback data, and operational data within a preset time after anomaly handling; then, labeling the collected feedback data; finally, using an incremental training algorithm, after training, calculating the accuracy of the new model on the validation set. If the accuracy is improved by ≥ a predetermined percentage compared to the original model, the model parameters of the anomaly identification module are updated; otherwise, the training parameters are readjusted and training is repeated, then returned to the corresponding identification model. The preset time is 7 days, but not limited to this; 5 days, etc., are also applicable to this invention. The preset percentage can be 2%, but not limited to this; 3%, etc., are also applicable to this invention. A continuous iteration mechanism and a large-scale adaptation architecture are established to adapt to large-scale wide-area networks, shorten the model iteration cycle, adapt to widely distributed charging networks, and reduce the operating cost of a single site.

[0112] like Figure 6 As shown in the embodiments of this disclosure, an abnormal identification system for car charging piles capable of performing the above-described method is also introduced. The system includes a data acquisition module, a data preprocessing module, an abnormal identification module, and a management decision module. The data acquisition module is used to collect multi-source data; the data preprocessing module is used to extract first feature data and second feature data based on the collected multi-source data; the abnormal identification module is used to identify charging gun abnormalities using a car charging abnormality gun identification model based on the extracted first feature data; and is used to identify charging station abnormalities using a car charging abnormality station identification model based on the extracted second feature data; the management decision module is used to generate an executable decision scheme based on the identified charging gun abnormalities and / or charging station abnormalities. In this embodiment of the disclosure, such as Figure 4 As shown, the system also includes a visualization module, which comprises a map interface, a data dashboard interface, and a large model question-and-answer module. The map interface provides a basic map layer, a site distribution layer, and an abnormal gun distribution layer. Users can show or hide the corresponding content by checking the layer names. Each abnormal point is marked, and brief information is displayed when the mouse hovers over it. Clicking on the point will pop up a details window. The interface also provides a point details query that includes abnormal data, historical processing records, and current processing progress. The data dashboard interface is used to display the anomaly identification accuracy rate, work order completion rate, and site anomaly rate in real time. The large-scale question-answering module is used to generate structured answers based on natural language questions.

[0113] In this embodiment of the disclosure, such as Figure 5 As shown, the system also includes a model iterative optimization module, used to perform the following steps: Collect feedback data, including operation and maintenance verification feedback data, user feedback data, and operation data within a preset time after the anomaly handling is completed; The collected feedback data is labeled; An incremental training algorithm is used. After training, the accuracy of the new model on the validation set is calculated. If the accuracy is improved by ≥ a predetermined percentage compared with the original model, the model parameters of the anomaly recognition module are updated. Otherwise, the training parameters are readjusted and the model is trained again. Then, the model is returned to the corresponding recognition model.

[0114] The aforementioned system addresses the problems of low efficiency, high cost, and difficulty in cross-regional management caused by the wide distribution of charging pile facilities, resulting in traditional manual inspections. It brings multi-dimensional value to the business. First, in terms of efficiency, through AI-based hierarchical identification models and reinforcement learning-based operation and maintenance scheduling, the response time for first-level anomaly work orders is significantly shortened, the number of sites managed per maintenance personnel is increased, and cross-regional operation and maintenance efficiency is improved, adapting to the needs of large-scale charging network management. Second, in terms of cost, eliminating manual inspections effectively reduces the number of frontline maintenance personnel, saving labor costs; early warning of hidden anomalies reduces the charging gun scrap rate, saving annual equipment repair and scrapping costs, amplifying cost advantages. Finally, in terms of operation and strategy, improved anomaly identification accuracy, reduced user complaint rate, and increased charging gun availability enhance the user experience.

[0115] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for identifying anomalies in car charging stations, characterized in that, include, Collect data from multiple sources; Based on the collected multi-source data, the first feature data and the second feature data are extracted respectively; Based on the extracted first feature data, an abnormal charging gun identification model is used to identify abnormal charging guns. Based on the extracted second feature data, an abnormal charging station identification model is used to identify station anomalies. Based on the identified charging gun anomalies and / or site anomalies, an actionable decision-making scheme is generated.

2. The method for identifying abnormalities in car charging piles according to claim 1, characterized in that, Multi-source data includes charging equipment data sources, charging station data sources, charging order data sources, and alarm data sources, among which, The charging equipment data sources include the charging gun's plug-in status, start / stop status, working status, current, voltage, and temperature, as well as the charging pile's power, operating time, and fault codes. Data sources for charging stations include the station's start / stop status, power supply data, equipment layout, number of charging piles, and operator information; The charging order data source includes detailed information about the charging order, including the order number, charging start and end time, charging amount, charging cost, and payment method. Alarm data sources include charging equipment fault alarms, security alarms, and abnormal transaction alarms.

3. The method for identifying abnormalities in car charging piles according to claim 2, characterized in that, Based on the collected multi-source data, the first feature data is extracted, which includes the number of insertions and removals. Continuous charging duration characteristics Current fluctuation amplitude characteristics ΔI, voltage stability characteristics and statistics on the frequency characteristics of abnormal alarms The extraction methods are as follows: By statistically analyzing the number of times the charging gun is plugged in and unplugged within a certain time window T, the characteristics of the number of plugging and unplugging times are extracted. ; By recording the duration of continuous charging during each charging process, the continuous charging duration feature is extracted. ; By calculating the current fluctuation amplitude during the charging process, the current fluctuation amplitude feature ΔI is extracted: ΔI=max(I) min(I) In the formula, I is the current sequence during the charging process; Voltage stability characteristics are extracted using the following calculation formula. : In the formula, It is the voltage value at the i-th sampling time. The average voltage is n, and the number of sampling points is n. By extracting statistical features of abnormal alarm frequency based on the number of abnormal alarms received by the charging gun within the time window T, we can identify these features. .

4. The method for identifying abnormalities in car charging stations according to claim 3, characterized in that, Based on the extracted first feature data, an abnormal charging gun identification model is used to identify charging gun anomalies. This includes a deep learning model based on a long short-term memory network to identify charging gun anomalies, specifically including... The extracted first feature data are combined into a feature vector. And convert it into time series data: , where t represents the time step and m is the number of feature types; Time series data is input into a deep learning model based on a long short-term memory network, and the output is a two-dimensional probability vector y=[y1,y2], which represents the probability that the charging gun is in a normal state and the probability that it is in an abnormal state. Here, y1 is the probability that the charging gun is in a normal operating state, and y2 is the probability that the charging gun is in an abnormal operating state. The constraint relationship is: y1+y2=1. Set an anomaly probability threshold , when y2> This directly indicates that the charging gun is malfunctioning.

5. The method for identifying abnormalities in car charging stations according to claim 4, characterized in that, This also includes training and optimizing deep learning models based on long short-term memory networks, including... The cross-entropy loss function is used to measure the difference between the model's predictions and the true labels. The cross-entropy loss function is: In the formula, N is the number of training samples. This is the true label of the i-th sample, where a value of 0 indicates normal and a value of 1 indicates abnormal. It is the probability of an anomaly predicted by the model for the i-th sample; The Adam optimizer is selected to update the parameters of the deep learning model based on the Long Short-Term Memory network, and the model is trained by minimizing the loss function.

6. The method for identifying abnormalities in car charging piles according to claim 5, characterized in that, This also includes using operational experience to assist in identifying charging gun malfunctions, specifically including: Acquire feature data, including abnormal orders, offline data, number of alarms, and demand fulfillment rate; Determine whether abnormal orders, offline data, alarm quantity, and demand fulfillment rate meet the corresponding preset abnormal conditions. If there is characteristic data that meets the preset abnormal conditions, then there is a charging gun abnormality, and the charging gun abnormality data is obtained.

7. The method for identifying abnormalities in car charging piles according to any one of claims 1-6, characterized in that, Based on the collected multi-source data, a second feature data is extracted, which includes the average charging power feature. Charging pile utilization characteristics Alarm Concentration Characteristics Power supply stability characteristics And the surrounding environmental factors E, the extraction methods are as follows: By calculating the average charging power of charging stations over a certain period of time, the average charging power characteristics are extracted. : In the formula, is the charging power of the i-th charging pile, and n is the number of charging piles; By calculating the ratio of the number of charging piles actually in use to the total number of charging piles, the utilization rate of charging piles is extracted. : In the formula, This represents the actual number of charging stations in use. This represents the total number of charging stations; By calculating the distribution concentration of alarms on different charging piles, alarm concentration characteristics are extracted. : in, It represents the number of alarms for the i-th charging station. This represents the number of alarms for the j-th charging station. This is the average number of alarms; By calculating the standard deviation of the power supply, the characteristics of power supply stability can be obtained. : In the formula, It is the power supply sequence of the i-th charging pile; The characteristics E of surrounding environmental factors are obtained based on ambient temperature and humidity.

8. The method for identifying abnormalities in car charging piles according to claim 7, characterized in that, Based on the extracted second feature data, an abnormal charging station identification model is used to identify charging station anomalies. This includes an ensemble learning model based on gradient boosting decision trees for identifying charging station anomalies. The extracted second feature data is input, and the ensemble learning model based on gradient boosting decision trees outputs the original predicted values. In the formula, M is the number of iteration rounds; An ensemble learning model based on gradient boosting decision trees converts predicted values ​​into category labels L by setting a threshold, where L=1 indicates anomaly and L=0 indicates normality. Original predicted value threshold setting : In the formula, The average of the original forecast values ​​for historically operating facilities. This represents the maximum value of the original forecast for historically operating facilities. These are the weighting coefficients; when > At that time, it was directly determined that there was an anomaly at the station.

9. The method for identifying abnormalities in car charging piles according to claim 8, characterized in that, It also includes model training and optimization of ensemble learning models based on gradient boosting decision trees, including, Using the logarithmic loss function: where y∈{ 1,1} are the true labels, and f(x) is the model prediction of the ensemble learning model based on gradient boosting decision tree; By adjusting the parameters of the gradient boosting decision tree ensemble learning model: the number of trees, the depth of the trees, and the learning rate β; Cross-validation is used to select the optimal combination of parameters.

10. The method for identifying abnormalities in car charging piles according to claim 9, characterized in that, It also includes using station operation experience to assist in identifying station anomalies, including, Collect the rated power data of each piece of equipment in the station as the benchmark value for judging power anomalies; Based on the start and end times of charging orders, filter scenarios where multiple charging guns are used simultaneously at the current power station; For scenarios involving simultaneous use, the earliest common usage time is determined by the intersection of the order times of each charging gun; and based on this time point, the current and voltage data of the corresponding order are matched from the data table and the real-time power is calculated. The maximum power of each charging gun under a single station is extracted and summed according to the station dimension. The power ratio is obtained by comparing the sum of the maximum power of a single station with the rated power of the station equipment. If the power percentage is less than the preset value, the site is considered abnormal.

11. The method for identifying abnormalities in car charging piles according to claim 10, characterized in that, Based on the identified charging gun anomalies and / or site anomalies, actionable decision-making solutions are generated, including the generation of maintenance work orders and operational optimization strategies. Based on geographical location, match the nearest maintenance personnel, and at the same time establish a skill tag library for maintenance personnel to match personnel with corresponding skills based on the type of abnormal charging gun; Generate a work order, which includes the work order number, site information, abnormal charging gun information, handling suggestions, time limit requirements, and feedback entry point; Generate a site optimization strategy report, which includes an overview of the anomaly, root cause analysis, specific strategies, expected results, responsible persons, and an implementation timeline, and push it to the site operation and management platform.

12. The method for identifying abnormalities in car charging piles according to claim 11, characterized in that, It also includes visualization through a visualization module, including, Based on the map interface, it provides a basic map layer, a station distribution layer, and an abnormal charging gun distribution layer. Users can show or hide the corresponding content by checking the layer name, mark each abnormal point, display brief information when the mouse hovers over it, and pop up a details window when the point is clicked. It also provides point details query, which includes abnormal data, historical processing records, and current processing progress. The data dashboard interface displays the anomaly identification accuracy rate, work order completion rate, and site anomaly rate in real time. The large model question-answering module asks questions in natural language and generates structured answers.

13. The method for identifying abnormalities in car charging piles according to claim 12, characterized in that, This also includes iterative optimization of the vehicle charging abnormality charging gun identification model and the vehicle charging abnormality site identification model, which includes, Collect feedback data, including operation and maintenance verification feedback data, user feedback data, and operation data within a preset time after the anomaly handling is completed; The collected feedback data is labeled; An incremental training algorithm is used. After training, the accuracy of the new model on the validation set is calculated. If the accuracy is improved by ≥ a predetermined percentage compared with the original model, the model parameters of the anomaly recognition module are updated. Otherwise, the training parameters are readjusted and the model is trained again. Then, the model is returned to the corresponding recognition model.

14. A vehicle charging station anomaly identification system, characterized in that, include, The data acquisition module is used to collect data from multiple sources. The data preprocessing module is used to extract the first feature data and the second feature data based on the collected multi-source data. The anomaly detection module is used to identify charging gun anomalies based on the extracted first feature data and a car charging anomaly charging gun identification model. And a vehicle charging anomaly identification model is used to identify charging station anomalies based on the extracted second feature data; The management decision module is used to generate actionable decision-making solutions based on identified charging gun anomalies and / or site anomalies.

15. The vehicle charging pile anomaly identification system according to claim 14, characterized in that, It also includes a visualization module, which comprises a map interface, a data dashboard interface, and a large model Q&A module. The map interface provides a basic map layer, a site distribution layer, and an abnormal charging gun distribution layer. Users can show or hide the corresponding content by checking the layer names. Each abnormal point is marked, and brief information is displayed when the mouse hovers over it. Clicking on the point will pop up a details window. The interface also provides a point details query that includes abnormal data, historical processing records, and current processing progress. The data dashboard interface is used to display the anomaly identification accuracy rate, work order completion rate, and site anomaly rate in real time. The large-scale question-answering module is used to generate structured answers based on natural language questions.

16. The vehicle charging pile anomaly identification system according to claim 15, characterized in that, It also includes a model iterative optimization module, used to perform the following steps: Collect feedback data, including operation and maintenance verification feedback data, user feedback data, and operation data within a preset time after the anomaly handling is completed; The collected feedback data is labeled; An incremental training algorithm is used. After training, the accuracy of the new model on the validation set is calculated. If the accuracy is improved by ≥ a predetermined percentage compared with the original model, the model parameters of the anomaly recognition module are updated. Otherwise, the training parameters are readjusted and the model is trained again. Then, the model is returned to the corresponding recognition model.

Images