A Smart Early Warning Method and System for Charging Piles Based on Hierarchical Co-evolution

Abstract

This invention discloses a hierarchical collaborative evolution-based intelligent early warning method and system for charging piles, relating to the field of charging data processing technology. The method and system include: S1, real-time acquisition of charging pile monitoring data and preprocessing of the data; S2, setting a judgment period, assessing the collaborative risk of the current period, generating risk level labels, and constructing a charging risk dataset; S3, calling the charging risk dataset, constructing an isolated forest model, an LSTM prediction model, and an XGBoost classification model for samples of different risk levels, and training the models; S4, dynamically updating the risk levels based on the real-time risk level labels and the trained models, and triggering an intelligent hierarchical early warning process. This solves the problem in existing technologies where the lack of hierarchical modeling and collaborative judgment mechanisms for different risk levels makes it difficult to balance the efficiency and accuracy of risk warnings.

Description

Technical Field

[0001] This invention relates to the field of charging data processing technology, specifically to a method and system for intelligent early warning of charging piles based on hierarchical collaborative evolution. Background Technology

[0002] With the development of the new energy vehicle industry, the number of electric vehicles on the road continues to rise, and charging piles, as their supporting infrastructure, are rapidly becoming widespread. Due to the drastic changes in physical quantities such as current, voltage, and temperature during charging, charging piles face various potential safety hazards, such as poor contact, overheating, and short circuits. Therefore, building a safe and efficient charging pile early warning system has become a key focus of the industry. Especially in scenarios with different fault levels, how to accurately distinguish between low-risk and high-risk scenarios, and how to combine multi-source monitoring data to achieve dynamic early warning, are the key technical challenges of current early warning systems.

[0003] For example, invention publication number CN117708733A discloses a safety early warning method and system for charging piles. It includes: real-time monitoring of preset target charging pile operating status data; filtering the operating status data to obtain filtered operating status data; performing shallow anomaly analysis on the filtered operating status data to obtain shallow anomaly identification results; performing predictive value analysis on the filtered operating status data using an autoregressive moving average model to obtain a predicted dataset corresponding to the filtered operating status data; performing threshold analysis on the operating status data using the predicted dataset to obtain threshold analysis results, and generating target early warning information based on the threshold analysis results; and calculating the emergency response time of the target early warning information to obtain the target emergency response time. This invention aims to improve the accuracy of safety early warnings for charging piles.

[0004] For example, the invention disclosed in CN117520748A provides a method for early warning of power module faults in electric vehicle charging piles, including the following steps: collecting current difference data; preprocessing the current difference data based on an autoregressive moving average model; constructing a many-to-one GRU subnetwork using an improved GRU network for each type of current difference data, constructing a new improved GRU network, and inputting the calculation results of the many-to-one GRU subnetwork into the newly constructed improved GRU network to construct an improved multi-layer GRU network; and using the trained improved multi-layer GRU network for early warning. This invention's improved GRU, by optimizing the calculation method of candidate hidden states, further reduces the computational load while improving the GRU's ability to identify fault characteristics in the charging pile current error time-series signal, thus enhancing the early warning capability.

[0005] However, existing technologies primarily focus on the analysis and processing of single models or single data sources, lacking differentiated modeling mechanisms for different risk levels. This makes it difficult to balance rapid response to low-risk warnings with accurate identification of high-risk warnings. Furthermore, most methods fail to establish a collaborative mechanism between model prediction and rule verification, resulting in issues such as missed or false alarms even in scenarios with interference from multiple information sources.

[0006] Therefore, in order to address the above issues, there is an urgent need for a smart early warning method and system for charging piles based on hierarchical collaborative evolution. Summary of the Invention

[0007] Technical problems to be solved

[0008] To address the shortcomings of existing technologies, this invention provides a charging pile intelligent early warning method and system based on hierarchical collaborative evolution, which solves the problem that the lack of hierarchical modeling and collaborative judgment mechanisms for different risk levels in existing technologies makes it difficult to balance the efficiency and accuracy of risk early warning.

[0009] Technical solution

[0010] To achieve the above objectives, this invention provides the following technical solution: a charging pile intelligent early warning method and system based on hierarchical collaborative evolution, comprising: S1, real-time acquisition of charging pile monitoring data, sequentially performing time alignment, weighted sliding filtering, exponential smoothing, and normalization compression processing on the charging pile monitoring data to obtain preprocessed charging pile monitoring data; S2, setting a judgment period, evaluating the collaborative risk of the current period based on the preprocessed charging pile monitoring data, generating risk level labels, and structurally storing the risk level labels and corresponding monitoring data to construct a charging risk dataset; S3, calling the charging risk dataset, constructing an isolated forest model, an LSTM prediction model, and an XGBoost classification model for samples of different risk levels, and constructing a weighted bias metric and training expression function to achieve multi-model collaborative learning and training; S4, based on the real-time risk level labels and the trained models, performing dynamic updates of risk levels and triggering an intelligent hierarchical early warning process: Level 1 risk calls the isolated forest model to output a Level 1 early warning signal; Level 2 risk combines the LSTM and XGBoost models to output a Level 2 early warning signal; Level 3 risk fuses an image recognition convolutional network and a heatmap feature extraction network to output a Level 3 early warning signal.

[0011] Further, the specific steps for collecting real-time monitoring data of the charging pile, and sequentially performing time alignment, weighted sliding filtering, exponential smoothing, and normalization compression processing on the monitoring data of the charging pile to obtain the preprocessed monitoring data of the charging pile are as follows: Real-time collect the monitoring data of the charging pile. The monitoring data of the charging pile includes the temperature of the charging cable, charging current, charging voltage, grounding resistance, ambient temperature, fan speed, and fan current. Through constructing a multi-source data alignment mechanism based on the consistency constraint of the sliding time window and sampling frequency, perform synchronous fusion processing on the monitoring data of the charging pile in the time dimension. Suppress noise of the monitoring data of the charging pile through the sliding average filtering algorithm, and weaken the high-frequency fluctuations caused by sensor jitter and electromagnetic interference. Perform short-term trend modeling and dynamic smoothing on the monitoring data of the charging pile through the exponentially weighted moving average algorithm. Use a normalization algorithm based on Z-score standardization and interquartile range compression to perform normalization and unified dimension processing on the monitoring data of the charging pile.

[0012] Further, set a judgment period. The specific steps for evaluating the collaborative risk of the current period based on the preprocessed monitoring data of the charging pile are as follows: Set a fixed-width sliding window as a judgment period. Calculate the average value of the charging current, the average value of the grounding resistance, and the average value of the temperature difference between the charging cable temperature and the ambient temperature within each period, and calculate the difference between the maximum value and the minimum value of the charging current within the period to obtain the charging current fluctuation value. Extract the real-time monitoring data of the charging pile. Divide the charging current fluctuation value by the average value of the charging current, and then multiply the obtained ratio by the first-order time derivative of the real-time charging cable temperature to obtain the electrothermal dynamic response term. Subtract the ratio of the real-time fan current to the rated fan current from the ratio of the real-time fan speed to the rated fan speed, take the absolute value of the obtained difference and add one to obtain the fan operation offset amplification term. Divide the difference between the real-time cable temperature and the real-time ambient temperature by the average value of the temperature difference, and then add the ratio of the real-time grounding resistance to the average value of the grounding resistance to obtain the geothermal risk normalization term. Multiply the electrothermal dynamic response term, the fan operation offset amplification term, and the geothermal risk normalization term in sequence to obtain the collaborative risk assessment value.

[0013] Further, the specific steps for generating risk level labels and storing the risk level labels and corresponding monitoring data in a structured manner to construct a charging risk data set are as follows: Compare the collaborative risk assessment value S with multiple-level risk thresholds S1, S2, and S3 to generate risk level labels: When S ≤ S1, mark the current period as normal charging; when S1 < S ≤ S2, mark the current period as level-1 risk; when S2 < S < S3, mark the current period as level-2 risk; when S ≥ S3, mark the current period as level-3 risk. Store the risk level label, the monitoring data of the charging pile, and the collaborative risk assessment value of the current period in a structured manner to construct a charging risk data set.

[0014] Furthermore, by utilizing the charging risk dataset, an isolated forest model, an LSTM prediction model, and an XGBoost classification model were constructed for samples at different risk levels. A weighted bias metric and training expression function were also constructed. The specific steps for multi-model collaborative learning training are as follows: Based on the charging risk dataset, charging pile monitoring data and collaborative risk assessment values ​​corresponding to normal charging and Level 1 risk were extracted to construct an isolated forest model. The number and depth of decision trees in the model were determined through multiple rounds of training iterations to complete the isolated forest model training. Charging cable temperature, charging current, charging voltage, ambient temperature, and fan speed corresponding to Level 2 risk were extracted to construct an LSTM prediction model. A time-weighted loss function was constructed, and the LSTM prediction model was trained with the time-weighted loss function value as the optimization objective. The output risk forward parameter set for the current period, including predicted values ​​for charging cable temperature, charging current, charging voltage, ambient temperature, and fan speed, was generated. Based on the risk forward parameter set output by the LSTM prediction model and the actual values ​​of the corresponding period's charging pile monitoring data, an XGBoost classification model was constructed. Combined with the corresponding collaborative risk assessment values, the predicted training expression values ​​were obtained through gradient boosting and parameter optimization, serving as training features to complete the XGBoost classification model training.

[0015] Further, the specific steps for constructing the time-weighted loss function and outputting its value are as follows: Sum the squared differences between the predicted and actual values ​​of all risk forward parameters within the t-th secondary risk period to form a prediction bias metric; multiply the time decay coefficient by the current period index and take its opposite, then take the power function value of the resulting opposite with respect to the natural constant e to obtain the time decay factor; multiply the prediction bias metric by the time decay factor to obtain the weighted bias value; calculate the arithmetic mean of the weighted bias values ​​across all secondary risk periods to obtain the time-weighted loss function value.

[0016] Furthermore, the specific steps for obtaining the predicted training expression value through gradient boosting mechanism and parameter optimization are as follows: Subtract the predicted and actual values ​​of all risk forward parameters within the t-th secondary risk period to obtain the error term for each risk forward parameter; square each error term and divide it by the sum of the corresponding actual value and the minimum term to obtain the normalized error for each risk forward parameter; sum the normalized errors of all risk forward parameters to obtain the normalized error metric for the current period; multiply the risk sensitivity coefficient, the collaborative risk assessment value of the t-th secondary risk period, and the corresponding normalized error metric, then add one, and take the natural logarithm of the sum to obtain the risk-weighted training error; calculate the arithmetic mean of the risk-weighted training errors across all secondary risk periods to obtain the predicted training expression value.

[0017] Furthermore, based on the real-time risk level labels and the trained model, the specific steps for dynamically updating the risk level are as follows: When the risk level is at level 1 for M consecutive judgment periods, the current risk level is upgraded from level 1 to level 2; when the risk level is at level 2, if any forward risk parameter output by the LSTM prediction model exceeds the operating limit, the current risk level is upgraded to level 3; when the risk level is at level 3, if a manual misjudgment confirmation signal is received, a risk level reset operation is performed, the current risk level is marked as charging normal, and the risk level judgment process is restarted.

[0018] Further, the intelligent hierarchical early warning process is triggered: For Level 1 risk, the isolated forest model is invoked to output a Level 1 early warning signal; for Level 2 risk, the LSTM and XGBoost models are combined to output a Level 2 early warning signal; for Level 3 risk, the image recognition convolutional network and heatmap feature extraction network are fused to output a Level 3 early warning signal. The specific steps are as follows: When the risk level is Level 1, statistical calculations are performed on the charging current and charging voltage in the current cycle to construct a time-domain feature vector, which is then input into the isolated forest model to output a Level 1 early warning signal; when the risk level is Level 2, charging pile monitoring data from the previous K cycles are extracted, input into the LSTM prediction model, and the risk forward parameter set for the current cycle is output. The risk forward parameter set includes predicted values ​​for charging cable temperature, charging current, charging voltage, ambient temperature, and fan speed; the deviation between the risk forward parameter set and the actual values ​​corresponding to the current cycle is calculated, a deviation vector is constructed, and input into XGBoost. The classification model outputs a level-two warning signal. When the risk level is level three, it acquires the charging gun interface image and corresponding charging pile monitoring data for the current period. The charging gun interface image is input into an image recognition convolutional network for processing, and the image anomaly index is output. The charging pile monitoring data is converted into a heat map and input into a heat map feature extraction network, and the coupled mutation feature is output. The image anomaly index and the coupled mutation feature form a joint feature group and are input into a multi-layer convolutional neural network to output a level-three warning signal. The warning signal includes the risk level, model type, and trigger time slice. For a level-one warning signal, it is output through the local human-machine interface of the charging pile, indicating that the current operating status has deviated from the normal range and requires continuous monitoring. For a level-two warning signal, it triggers the on-site audible and visual alarm device and calls the remote communication interface to push the warning signal to the terminal equipment of the registered maintenance personnel. For a level-three warning signal, it immediately cuts off the current charging circuit based on the level-two warning.

[0019] The second aspect of this invention provides a hierarchical collaborative evolution-based intelligent early warning system for charging piles, characterized by comprising: a data acquisition and preprocessing module, a collaborative risk assessment module, a multi-model integrated training module, and an intelligent hierarchical early warning module, wherein: the data acquisition and preprocessing module is used to acquire charging pile monitoring data in real time, and sequentially perform time alignment, weighted sliding filtering, exponential smoothing, and normalization compression processing on the charging pile monitoring data to obtain preprocessed charging pile monitoring data; the collaborative risk assessment module is used to set a judgment period, assess the collaborative risk of the current period based on the preprocessed charging pile monitoring data, generate risk level labels, and structurally store the risk level labels and corresponding monitoring data to construct a charging risk assessment system. The system includes a risk dataset and a multi-model ensemble training module. This module calls upon the charging risk dataset to construct Isolation Forest, LSTM prediction, and XGBoost classification models for samples at different risk levels. It also constructs a weighted bias metric and training expression function to achieve collaborative learning and training of multiple models. An intelligent risk-level warning module is used to dynamically update risk levels based on real-time risk level labels and the trained models, triggering an intelligent risk-level warning process: Level 1 risk calls the Isolation Forest model to output a Level 1 warning signal; Level 2 risk combines the LSTM and XGBoost models to output a Level 2 warning signal; Level 3 risk fuses an image recognition convolutional network and a heatmap feature extraction network to output a Level 3 warning signal.

[0020] Beneficial effects

[0021] The present invention has the following beneficial effects:

[0022] (1) The intelligent early warning method and system for charging piles based on hierarchical collaborative evolution introduces a collaborative risk assessment mechanism, integrates multiple real-time indicators such as current fluctuation, temperature difference, grounding resistance, and fan status, obtains collaborative risk assessment values, and compares them with multi-level risk thresholds to generate risk level labels, thereby realizing the quantification and classification of risks in the charging process, avoiding false alarms and missed alarms caused by single threshold judgment, and effectively improving the accuracy and stability of risk judgment.

[0023] (2) This intelligent early warning method and system for charging piles based on hierarchical collaborative evolution adopts differentiated models for different risk levels: an isolated forest is used for low-risk anomaly detection, an LSTM prediction model is used for time-series prediction of secondary risks, and an XGBoost classification model is used for classification of complex features. By utilizing weighted bias metric and training expression function, multi-model collaborative learning is achieved, realizing rapid response in low-risk scenarios and accurate identification in high-risk scenarios.

[0024] (3) The intelligent early warning method and system for charging piles based on hierarchical collaborative evolution measures the prediction deviation through a time-weighted loss function and optimizes the LSTM model by combining a time decay factor; it characterizes the impact of collaborative risk and prediction error on the classification results by predicting training expression values, providing a more risk-sensitive training index for the XGBoost model, improving the discrimination and interpretability of the classification results for actual operational risks, thereby significantly improving the reliability and stability of the system in practical applications.

[0025] (4) This intelligent early warning method and system for charging piles based on hierarchical collaborative evolution matches different model paths according to the risk level: Level 1 risk outputs early warning prompts, Level 2 risk triggers on-site alarms and remote notifications, and Level 3 risk integrates image recognition and heat map feature extraction to output emergency warnings and cut off the circuit. This hierarchical early warning mechanism realizes a progressive response from low risk to high risk, improves the timeliness and reliability of safety protection, and ensures the safety of charging pile operation and the timeliness of early warning handling. Attached Figure Description

[0026] Figure 1 Here is a flowchart of a charging pile intelligent early warning method based on hierarchical collaborative evolution;

[0027] Figure 2 This is a structural diagram of a charging pile intelligent early warning system based on hierarchical collaborative evolution.

[0028] Figure 3 A bar chart showing the collaborative risk assessment values ​​and risk levels;

[0029] Figure 4 This is a schematic diagram of the intelligent early warning process. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Please see Figures 1-4This invention provides a technical solution: a charging pile intelligent early warning method and system based on hierarchical collaborative evolution, comprising: S1, real-time acquisition of charging pile monitoring data, sequentially performing time alignment, weighted sliding filtering, exponential smoothing and normalization compression processing on the charging pile monitoring data to obtain preprocessed charging pile monitoring data; S2, setting a judgment period, evaluating the collaborative risk of the current period based on the preprocessed charging pile monitoring data, generating risk level labels, and structurally storing the risk level labels and corresponding monitoring data to construct a charging risk dataset; S3, calling the charging risk dataset, constructing an isolated forest model, an LSTM prediction model and an XGBoost classification model for samples of different risk levels respectively, and constructing a weighted bias metric and training expression function to achieve multi-model collaborative learning training; S4, based on the real-time risk level labels and the trained models, performing dynamic risk level updates and triggering an intelligent hierarchical early warning process: Level 1 risk calls the isolated forest model to output a Level 1 early warning signal; Level 2 risk combines the LSTM and XGBoost models to output a Level 2 early warning signal; Level 3 risk fuses an image recognition convolutional network and a heatmap feature extraction network to output a Level 3 early warning signal.

[0032] Specifically, the charging pile monitoring data is collected in real time, and time alignment, weighted sliding filtering, exponential smoothing, and normalized compression are performed sequentially on the preprocessed charging pile monitoring data. The specific steps to obtain the preprocessed charging pile monitoring data are as follows: First, charging pile monitoring data is collected in real time, including charging cable temperature, charging current, charging voltage, grounding resistance, ambient temperature, fan speed, and fan current. Among them, the charging cable temperature data is obtained through thermocouple sensors installed inside the charging cable, reflecting the heat accumulation of the cable conductor and insulation layer; the charging current signal is collected in real time through a high-precision current transformer to characterize the current intensity fluctuation during the charging process; the instantaneous value of the charging voltage is obtained through the voltage sampling circuit inside the charging pile to analyze the output stability; the grounding resistance is measured through an ohmmeter in the grounding line to determine the electrical safety status; the ambient temperature is obtained through a temperature sensor installed in the charging environment to reflect the impact of external temperature control on the system's thermal management; the fan speed data is collected through a Hall effect speed sensor embedded in the fan drive unit to evaluate the heat dissipation response speed; and the fan current change is obtained through a current sensor in the fan power supply line to determine the degree of heat dissipation load. After completing the real-time acquisition of the aforementioned multi-source data, a multi-source data alignment mechanism based on a sliding time window and sampling frequency consistency constraint is constructed to synchronously fuse the charging pile monitoring data in the time dimension, ensuring the accuracy of the correspondence of various monitoring signals on the time axis. Subsequently, a moving average filtering algorithm is used to suppress noise in the charging pile monitoring data, reducing high-frequency fluctuations caused by sensor jitter and electromagnetic interference, and improving signal stability. Furthermore, an exponentially weighted moving average algorithm is used to model short-term trends and dynamically smooth the charging pile monitoring data, preserving key trends while suppressing short-period fluctuations. Finally, a normalization algorithm based on Z-score standardization and interquartile range compression is used to normalize and unify the dimensions of the charging pile monitoring data, ensuring comparability and numerical stability of data in subsequent model calculations.

[0033] In this implementation scheme, real-time acquisition and independent processing of charging pile monitoring data ensure that each type of physical quantity has a clear acquisition path and physical meaning, effectively supporting subsequent feature engineering and modeling accuracy. Specifically, a multi-source data alignment mechanism based on a sliding time window and sampling frequency constraints is designed to achieve precise synchronization of each monitoring variable in the time dimension, avoiding the loss of temporal information caused by sampling offsets. A moving average filtering algorithm is combined to suppress high-frequency noise and improve signal stability. An exponentially weighted moving average algorithm is introduced to model and dynamically smooth short-term trends, effectively reducing the interference of short-term disturbances on the model. Furthermore, a normalization algorithm combining Z-score standardization and interquartile range compression is used to achieve dimensional uniformity, making data from different dimensions comparable and embeddable. These preprocessing schemes enhance the adaptability and numerical stability of charging pile monitoring data in anomaly detection and risk modeling scenarios.

[0034] Specifically, the steps for setting a judgment period and assessing the collaborative risk of the current period based on preprocessed charging pile monitoring data are as follows: First, a fixed-width sliding window is set as a judgment period, with the window length matching the data sampling frequency to ensure temporal consistency of all charging pile monitoring data within the same judgment period. The average charging current is calculated within the period to obtain the average charging current for the current period; the average grounding resistance within the period is also calculated simultaneously to obtain the average grounding resistance; simultaneously, the point-by-point difference sequence between the charging cable temperature and the ambient temperature within the period is extracted, and the average of the difference sequence is calculated as the average temperature difference value; based on this, the difference between the maximum and minimum charging current within the period is calculated and defined as the charging current fluctuation value. Second, real-time charging pile monitoring data at the current moment is extracted, and the charging current fluctuation value of the period is divided by the average charging current within the period to obtain the relative current fluctuation rate; then, the real-time charging cable temperature is extracted and the corresponding first-order time derivative is calculated to obtain the charging cable temperature change rate; the relative current fluctuation rate is multiplied by the charging cable temperature change rate to form an electrothermal dynamic response term, used to measure the coupling degree between current disturbance and thermal change. Then, the real-time fan current and rated fan current are extracted and their ratio is calculated; simultaneously, the real-time fan speed and rated fan speed are extracted and their ratio is calculated; the absolute value of the difference between the two ratios is taken and one is added to construct a fan operation offset amplification term, which is used to characterize the degree to which the fan operation deviates from the rated operating condition. Next, the difference between the real-time cable temperature and the real-time ambient temperature is divided by the average temperature difference value, and the ratio of the real-time grounding resistance to the average grounding resistance is added to obtain a geothermal risk normalization term, which comprehensively reflects the impact of temperature gradient and grounding resistance offset in the geothermal path on the system's thermal safety. Finally, the electrothermal dynamic response term, the fan operation offset amplification term, and the geothermal risk normalization term are multiplied sequentially to obtain a collaborative risk assessment value, which is used to quantitatively characterize the degree of multi-source risk coupling and potential early warning level during the current charging cycle.

[0035] The specific formula for calculating the collaborative risk assessment value is as follows:

[0036] ;

[0037] In the formula, This represents the collaborative risk assessment value. This indicates the value of charging current fluctuation. This represents the average charging current. Indicates the real-time temperature of the charging cable. Indicates the real-time ambient temperature. Indicates real-time fan speed. Indicates the rated fan speed. Indicates real-time fan current. Indicates the rated fan current. This represents the average temperature difference. Indicates the real-time grounding resistance. This represents the average grounding resistance.

[0038] In this embodiment, Table 1 is a collaborative risk assessment value data table, listing the monitoring parameters and corresponding collaborative risk assessment values ​​for five judgment periods. The monitoring parameters include: charging current fluctuation value, average charging current, charging cable temperature change rate, real-time fan current, rated fan current, real-time fan speed, rated fan speed, charging cable real-time temperature, ambient real-time temperature, average temperature difference, and real-time grounding resistance and average grounding resistance. Specific data are as follows: In judgment period 1, the charging current fluctuation value is 2.1, the average charging current is 20, the charging cable temperature change rate is 0.8, the real-time fan current is 0.90, the rated fan current is 1.0, the real-time fan speed is 1600, the rated fan speed is 1600, the charging cable real-time temperature is 50, the ambient real-time temperature is 35, the average temperature difference is 10, the real-time grounding resistance is 0.30, the average grounding resistance is 0.5, and the corresponding collaborative risk assessment value is 0.19; In judgment period 2, the charging current fluctuation value is 3.2, the average charging current is 22, the charging cable temperature change rate is 2. ... The temperature change rate was 1.1, the real-time fan current was 0.95, the rated fan current was 1.0, the real-time fan speed was 1500, the rated fan speed was 1600, the real-time charging cable temperature was 52, the real-time ambient temperature was 36, the average temperature difference was 10, the real-time grounding resistance was 0.35, the average grounding resistance was 0.5, and the corresponding collaborative risk assessment value was 0.37. In judgment period 3, the charging current fluctuation value was 4.5, the average charging current was 21, the charging cable temperature change rate was 1.5, the real-time fan current was 1.00, the rated fan current was 1.0, and the real-time fan speed was 140. 0, rated fan speed is 1600, real-time charging cable temperature is 60, real-time ambient temperature is 34, average temperature difference is 10, real-time grounding resistance is 0.45, average grounding resistance is 0.5, and the corresponding collaborative risk assessment value is 1.27; In judgment period 4, the charging current fluctuation value is 2.5, the average charging current is 20, the charging cable temperature change rate is 0.9, the real-time fan current is 0.85, the rated fan current is 1.0, the real-time fan speed is 1550, the rated fan speed is 1600, the real-time charging cable temperature is 49, the real-time ambient temperature is 33, and the average temperature difference is... 10. The real-time grounding resistance is 0.32, the average grounding resistance is 0.5, and the corresponding collaborative risk assessment value is 0.28. In judgment period 5, the charging current fluctuation value is 3.9, the average charging current is 21, the charging cable temperature change rate is 1.3, the real-time fan current is 1.10, the rated fan current is 1.0, the real-time fan speed is 1350, the rated fan speed is 1600, the real-time temperature of the charging cable is 58, the real-time ambient temperature is 34, the average temperature difference is 10, the real-time grounding resistance is 0.43, the average grounding resistance is 0.5, and the corresponding collaborative risk assessment value is 0.99.

[0039] Table 1. Collaborative Risk Assessment Values ​​Data Table

[0040]

[0041] like Figure 3 As shown, the chart displays the collaborative risk assessment values ​​and risk level judgment results corresponding to five judgment periods, used to evaluate the comprehensive risk level and safety status of the charging process under multi-source charging pile monitoring data. The chart uses a color-coded bar chart to represent the collaborative risk assessment value S for each judgment period, with different risk levels distinguished by color: green indicates normal charging, blue indicates level one risk, orange indicates level two risk, and red indicates level three risk, visually reflecting the risk level under different judgment periods. Three dashed threshold lines are set in the chart as references for risk level division: the gray dashed line represents the level one risk threshold S1, used to identify the initial collaborative risk warning boundary; the black dashed line represents the level two risk threshold S2, defining the medium-intensity risk warning range; and the purple dashed line represents the level three risk threshold S3, used to identify high-risk collaborative risk states. As shown in the figure: the collaborative risk assessment value for judgment period 1 is 0.19, which is lower than S1, and is judged as normal charging; the collaborative risk assessment value for judgment period 2 is 0.37, which is between S1 and S2, belonging to level 1 risk; the collaborative risk assessment value for judgment period 3 is 1.27, which exceeds S3, and has entered the level 3 high-risk warning zone; the collaborative risk assessment value for judgment period 4 is 0.28, which is slightly higher than S1, and is still level 1 risk; the collaborative risk assessment value for judgment period 5 is 0.99, which is between S2 and S3, belonging to level 2 risk. Figure 3 This effectively demonstrates the risk identification performance of the collaborative risk assessment mechanism in complex working conditions under different judgment periods. Multi-level risk color differentiation and quantitative threshold definition jointly enhance the intuitiveness and accuracy of risk identification, providing visual support for dynamic risk level updates and intelligent early warning decision-making.

[0042] In this embodiment, the present invention constructs a periodic statistical feature extraction mechanism based on the average charging current, average grounding resistance, average temperature difference, and charging current fluctuation. Combined with key monitoring indicators such as real-time charging cable temperature, fan current, fan speed, and grounding resistance, it comprehensively depicts the multi-source heat, electricity, and heat dissipation states during the charging process. The resulting collaborative risk assessment value can dynamically reflect the complex potential risks during the charging process, providing accurate, quantitative, and time-consistent input for subsequent risk level judgment, model selection, and early warning response paths. This significantly enhances the sensitivity, accuracy, and hierarchical control foundation of risk perception.

[0043] Specifically, the specific steps for generating risk level labels, storing the risk level labels and corresponding monitoring data in a structured manner, and constructing a charging risk data set are as follows: First, based on the collaborative risk assessment value S extracted within each fixed sliding judgment period, an accurate comparison is made with multi-level risk thresholds, where the multi-level risk thresholds include a first-level risk threshold S1, a second-level risk threshold S2, and a third-level risk threshold S3. The specific label generation rules are as follows: When the collaborative risk assessment value S ≤ the first-level risk threshold S1, the current judgment period is marked as the normal charging level, indicating that the charging status is stable and the risk is controllable within the current period; when the first-level risk threshold S1 < S ≤ the second-level risk threshold S2, the current judgment period is marked as the first-level risk level, which is used to indicate that there are minor abnormalities during the charging process but no direct safety threat is constituted; when the second-level risk threshold S2 < S < the third-level risk threshold S3, the current judgment period is marked as the second-level risk level, representing a medium risk and requiring enhanced monitoring; when the collaborative risk assessment value S ≥ the third-level risk threshold S3, the current judgment period is marked as the third-level risk level, indicating that it has entered a high-risk state and a strong intervention mechanism needs to be triggered. Subsequently, the risk level label, charging pile monitoring data, and the calculated collaborative risk assessment value S within the judgment period are jointly formed into a structured record entry and stored in a unified data structure to form a charging risk data set with hierarchical label annotation, which is used to support subsequent multi-model training and intelligent early warning processes.

[0044] In this implementation plan, by comparing the collaborative risk assessment value with multi-level risk thresholds, clear risk level labels are generated, and combined with the complete charging pile monitoring data and collaborative risk assessment value within the corresponding judgment period, structured storage is carried out to construct a charging risk data set. This mechanism realizes the organic unity of hierarchical identification, label generation, and data set construction of the risks during the charging process. It not only improves the organization efficiency and time series integrity of the charging monitoring data but also provides a risk data basis with labels for subsequent model training, enhances the learning ability and adaptability of the system in scenarios with multiple risk levels, and effectively supports the accurate modeling and intelligent early warning response of charging risks.

[0045] Specifically, the charging risk dataset is used to construct an isolated forest model, an LSTM prediction model, and an XGBoost classification model for samples of different risk levels, and a weighted bias metric and training expression function are constructed to achieve multi-model collaborative learning training. The specific steps are as follows: Based on the structured storage of the charging risk dataset, the charging cable temperature, charging current, charging voltage, grounding resistance, ambient temperature, fan speed, and fan current, as well as the corresponding collaborative risk assessment values, are extracted during the judgment period corresponding to normal charging and level 1 risk, to construct the input sample set of the isolated forest model; combined with the ability of the isolated forest model to identify risk anomalies, the number and maximum depth of decision trees in the model are automatically adjusted by controlling the sample sampling ratio of each decision tree and using multiple rounds of training iterations, and finally the isolated forest model used for level 1 risk period anomaly detection is output. Subsequently, for the assessment period at the secondary risk level, five time-series variables—charging cable temperature, charging current, charging voltage, ambient temperature, and fan speed—were extracted from the charging risk dataset to construct the input data sequence for the LSTM prediction model. A time-weighted loss function based on a time-step weighting factor was constructed, assigning different weights to the prediction bias at different time steps to enhance the model's learning ability for key time periods. Using the time-weighted loss function value as the optimization objective, the model parameters were updated through backpropagation to train and obtain the risk forward parameter set for the current period. The risk forward parameter set includes predicted values ​​for charging cable temperature, charging current, charging voltage, ambient temperature, and fan speed. Finally, by combining the risk forward parameter set output by the LSTM prediction model with the actual values ​​of the charging pile monitoring data for the current period, dimensional alignment and time-series pairing were performed to construct a feature difference set for model training. The corresponding collaborative risk assessment value for the current period was introduced as an auxiliary feature input into the XGBoost classification model, and multiple rounds of iterative optimization were performed through a gradient boosting mechanism to dynamically adjust the feature splitting nodes and weight parameters. Meanwhile, the predicted training expression value is used as the supervision signal of the loss function. The predicted training expression value is composed of feature difference, collaborative risk assessment value and risk sensitivity coefficient. It guides the optimization direction of the model and strengthens the ability to identify high-risk sample feature combinations, thereby improving the discrimination accuracy and robustness of the XGBoost classification model in the secondary risk identification scenario.

[0046] In this implementation plan, a data classification mechanism for charging pile monitoring is constructed to address different risk levels. An isolated forest model, an LSTM prediction model, and an XGBoost classification model are trained separately. A weighted bias metric and training expression function are used to achieve collaborative training of multiple models, effectively improving the models' ability to discriminate different risk scenarios and their prediction accuracy. Overall, this collaborative modeling and training process effectively enhances the adaptability, stability, and intelligence of the early warning system under tiered response conditions, providing high-quality multi-model support for subsequent dynamic risk level assessment and intelligent tiered early warning output.

[0047] Specifically, the steps for constructing and outputting the time-weighted loss function are as follows: For each judgment period at the second-level risk level, extract the predicted and actual monitored values ​​of the risk forward parameter group consisting of charging cable temperature, charging current, charging voltage, ambient temperature, and fan speed. Calculate the corresponding variable differences, square all variable differences, and sum them to obtain the overall prediction deviation metric for the current period. Multiply the time decay coefficient by the index number of the current period and take the opposite number. Use the resulting opposite number as the exponent to calculate a power function value with the natural constant e as the base, forming the time decay factor for the current period. This factor represents the decay trend of the importance of the risk forward parameters to model training over time. Multiply the prediction deviation metric for the current period by the corresponding time decay factor to obtain the weighted deviation value after time-weighted adjustment. Calculate the arithmetic mean of the weighted deviation values ​​for each judgment period at the second-level risk level, and output the time-weighted loss function value used to measure the global training effect of the LSTM prediction model. Specifically, by analyzing the evolution trend of collaborative risk assessment values ​​in the historical judgment period of known level 2 risk, the least squares method is used to fit the distribution of these collaborative risk assessment values ​​under different period indices to obtain the time decay coefficient, which is used to control the influence of different periods on the total loss in the time-series weighted loss function. The value range of the time decay coefficient is [0,1].

[0048] The specific formula for calculating the time-weighted loss function value is as follows:

[0049] ;

[0050] In the formula, This represents the time-weighted loss function value. This indicates the total number of secondary risk periods within the assessment period. This represents the predicted value of the forward parameter for the i-th risk within the t-th secondary risk period. This represents the actual value of the forward parameter of the i-th risk within the t-th secondary risk. Indicates the time decay coefficient. This indicates a secondary risk cycle index.

[0051] In this embodiment, the present invention introduces a time decay factor to weight the prediction deviation of each judgment period, thereby enhancing the sensitivity to the evolutionary features of the time dimension during model training and effectively avoiding the mean bias problem of the influence of different time nodes on the training results. At the same time, by using the risk forward parameter set as the prediction object, the model can accurately capture the dynamic trends and potential abnormal signals within the secondary risk period, improving the LSTM prediction model's ability to identify and respond to high-risk evolutionary trends. This provides more risk-sensitive training feature inputs for the subsequent training of the XGBoost classification model, thereby further enhancing the accuracy and robustness of the multi-model collaborative mechanism in complex risk scenarios.

[0052] Specifically, the steps for obtaining the predicted training expression value through gradient boosting and parameter optimization are as follows: First, for each judgment period under secondary risk, the corresponding risk forward parameters are extracted; the predicted and actual values ​​of all risk forward parameters in the t-th secondary risk period are subtracted respectively to obtain the error term for each risk forward parameter; after squaring each error term, it is divided by the sum of the corresponding actual value and the minima to obtain the normalized error of each risk forward parameter; the normalized errors of all risk forward parameters are summed to obtain the normalized error metric for the current period, thereby avoiding the imbalance problem between indicators of different dimensions. Subsequently, the risk sensitivity coefficient, the collaborative risk assessment value in the t-th secondary risk period, and the normalized error metric are multiplied sequentially, and the product is incremented by one and the natural logarithm is taken to obtain the risk-weighted training error for the current period, enhancing the model's focus on training high-risk samples. Finally, the arithmetic mean of the risk-weighted training errors across all secondary risk periods is taken to obtain the predicted training expression value for the current stage. This value guides the XGBoost classification model to enhance its ability to identify and distinguish high-risk feature patterns during parameter updates, thereby improving the overall model's generalization performance in complex risk scenarios. The risk sensitivity coefficient is based on the nonlinear fitting relationship between the collaborative risk assessment value and the actual failure rate of secondary risk samples in historical charging pile monitoring data. An exponentially weighted moving average algorithm is used for fitting, and a stable solution is selected through cross-validation. The risk sensitivity coefficient ranges from [0,1]. The minima refer to a very small, non-zero positive constant used to prevent division by zero during division operations, ensuring the numerical stability and continuity of the calculation process.

[0053] The specific formula for calculating the predicted training expression value is as follows:

[0054] ;

[0055] In the formula, This represents the predicted training expression value. This indicates the total number of secondary risk periods within the assessment period. Indicates the risk sensitivity coefficient. This represents the collaborative risk assessment value for the t-th secondary risk period. This represents the predicted value of the forward parameter for the i-th risk within the t-th secondary risk period. This represents the actual value of the forward parameter of the i-th risk within the t-th secondary risk. Indicates a minus term.

[0056] In this implementation scheme, a prediction training expression value calculation mechanism is constructed to enhance the reinforcement learning and robust generalization capabilities of the intelligent early warning model for charging piles on high-risk samples. The acquired prediction training expression values ​​with risk perception capabilities provide training features for the XGBoost classification model that balance accuracy and risk coverage. This ensures that the XGBoost classification model can accurately identify high-risk feature patterns in multi-source heterogeneous and highly dynamic perturbation risk scenarios, thereby improving the accuracy, stability, and safety protection capabilities of the entire early warning system under complex charging conditions.

[0057] Specifically, based on the real-time generated risk level labels and the trained Isolation Forest model, LSTM prediction model, and XGBoost classification model, the specific steps for dynamically updating the risk level are as follows: When the risk level labels generated in M ​​consecutive judgment cycles are all Level 1 risk, it is considered that there is a continuous mild abnormal trend in the current charging process. At this time, the current risk level is upgraded from Level 1 to Level 2, triggering the Level 2 risk response mechanism; where M is a positive integer constant; when the current risk level is in Level 2 risk state, the LSTM prediction model is called in real time to obtain the predicted values ​​of all risk forward parameters in the next cycle. If any predicted value exceeds the operating limit threshold, it indicates that there is a sudden high-risk hidden danger. The current risk level is upgraded to Level 3, and the Level 3 risk intervention mechanism is executed in conjunction; when the risk level is in Level 3 state, a misjudgment confirmation signal is received after manual review. Then, a risk level reset operation is performed, the risk level label is reset to normal charging, and the complete judgment process from judgment cycle setting, charging pile monitoring data processing to risk level label generation is restarted.

[0058] This implementation scheme achieves an adaptive control mechanism for charging risk status through dynamic updates of charging risk levels, enabling flexible responses to risk changes in a dynamic charging environment. When a persistent Level 1 risk status is detected, potential mild risk trends can be identified in a timely manner, significantly improving the sensitivity of early warning responses. When the risk level escalates to Level 3, it possesses the ability to respond quickly to sudden severe risks, further enhancing the effectiveness of the risk prevention and control system. Furthermore, by introducing a manual misjudgment confirmation mechanism, the risk level can be reasonably reset under extreme conditions, effectively ensuring the stability and controllability of the risk assessment mechanism in complex charging scenarios. Overall, this dynamic risk level update mechanism constructs a complete closed-loop control process from risk discovery, trend identification, emergency response to misjudgment correction, improving the adaptability, reliability, and practicality of this invention in the safe operation of charging piles.

[0059] Specifically, the intelligent hierarchical early warning process is triggered, dynamically selecting the corresponding model architecture and outputting an early warning signal based on the real-time risk level label. This includes the following steps: First, when the risk level of the current judgment period is Level 1, the charging current and charging voltage within the period are extracted, and statistical features such as mean, variance, and peak value are calculated to construct a standardized time-domain feature vector. This vector is then input into the trained Isolation Forest model, outputting a Level 1 early warning signal. Second, when the risk level of the current judgment period is Level 2, five types of charging pile monitoring data (charging cable temperature, charging current, charging voltage, ambient temperature, and fan speed) from the previous K judgment periods are extracted from the current moment to form a time series input. This input is then fed into the trained LSTM prediction model to predict the risk forward parameter set for the current period. Here, K is a positive integer constant. Subsequently, the predicted value and the actual value of the current period are sequentially differenced to construct a risk deviation vector. The risk deviation vector is used as input features into the XGBoost classification model and combined with the collaborative risk assessment value of the current period to generate a Level 2 early warning signal, indicating a moderate-intensity potential operational anomaly in the current period. When the risk level for the current assessment period is Level 3, a multimodal fusion early warning process based on images and monitoring data is executed. Specifically, this includes: First, acquiring images of the charging gun interface from a high-resolution image sensor for the current period, inputting them into an image recognition convolutional network to extract typical risk features such as loose connections, interface dust accumulation, and structural obstruction, and outputting image anomaly indicators. Simultaneously, constructing a two-dimensional heatmap matrix image from all charging pile monitoring data within the current period, and inputting it into a heatmap feature extraction network to extract spatial anomaly patterns such as sudden temperature rises and fluctuation imbalances, obtaining coupled mutation feature vectors. Subsequently, combining the image anomaly indicators and coupled mutation features into a joint feature group, inputting it into a multi-layer convolutional neural network for multimodal feature fusion and high-dimensional representation reconstruction, outputting a Level 3 early warning signal, indicating a serious charging safety risk. Furthermore, all risk level early warning signals are output in a structured format, containing three core pieces of information: risk level label, trigger model type, and early warning trigger time slice. Specifically: For Level 1 warning signals, a warning is issued through the local human-machine interface of the charging pile, reminding the on-duty personnel that the current operating status has deviated from the normal range and requires enhanced continuous monitoring; for Level 2 warning signals, the local audible and visual alarm device is immediately triggered, and the warning information is pushed to the terminal equipment of registered maintenance personnel through the remote communication interface to ensure rapid response under medium-risk conditions; for Level 3 warning signals, based on the Level 2 warning response, a charging control power-off command is further executed to achieve a safe disconnection of the current circuit and prevent serious risks from escalating further.

[0060] This implementation plan constructs a risk-level driven intelligent hierarchical early warning process, realizing a multi-model collaborative charging risk identification and hierarchical response mechanism, significantly improving the multimodal feature analysis capability of charging pile monitoring data. At risk level one, the isolated forest model is used to rapidly detect anomalies in the time-domain feature vector constructed from charging current and charging voltage, enhancing the early identification capability of mild risks. At risk level two, the LSTM prediction model and XGBoost classification model are integrated, and dynamic evaluation is performed based on the deviation vector constructed from the risk forward parameter set and actual observations, enhancing the accurate perception capability of moderate-intensity operational fluctuations. At risk level three, the convolutional recognition of charging gun interface images and coupled mutation detection of heatmap feature extraction networks are combined, utilizing joint feature sets to achieve high-order correlation judgment between images and monitoring data, thereby improving the rapid response capability under severe risk conditions. Early warning signals at each level form a hierarchical output channel through a local interactive interface, a remote communication interface, and a power-off control mechanism, constructing a closed-loop response system covering mild early warnings, audible and visual alarms, and circuit disconnection. Overall, this early warning process is characterized by high accuracy, high robustness, and high real-time performance, providing effective support for complex risk situation awareness and safety control in charging scenarios.

[0061] like Figure 2 As shown, the second aspect of the present invention provides a charging pile intelligent early warning system based on hierarchical collaborative evolution, characterized in that it includes: a data acquisition and preprocessing module, a collaborative risk assessment module, a multi-model integrated training module, and an intelligent hierarchical early warning module, wherein: the data acquisition and preprocessing module is used to acquire charging pile monitoring data in real time, and sequentially perform time alignment, weighted sliding filtering, exponential smoothing, and normalization compression processing on the charging pile monitoring data to obtain preprocessed charging pile monitoring data; the collaborative risk assessment module is used to set a judgment period, assess the collaborative risk of the current period based on the preprocessed charging pile monitoring data, generate risk level labels, and structurally store the risk level labels and corresponding monitoring data to construct a charging pile intelligent early warning system. The system includes: a charging risk dataset; a multi-model ensemble training module, which calls upon the charging risk dataset to construct an isolated forest model, an LSTM prediction model, and an XGBoost classification model for samples of different risk levels, and constructs a weighted bias metric and training expression function to achieve multi-model collaborative learning and training; and an intelligent graded early warning module, which performs dynamic updates of risk levels based on real-time risk level labels and the trained models, and triggers the intelligent graded early warning process: for level 1 risks, the isolated forest model is called to output a level 1 early warning signal; for level 2 risks, the LSTM and XGBoost models are combined to output a level 2 early warning signal; and for level 3 risks, an image recognition convolutional network and a heatmap feature extraction network are fused to output a level 3 early warning signal.

[0062] like Figure 4The diagram illustrates the multi-level early warning mechanism based on risk level determination in this invention. The overall process begins with data acquisition. Real-time collection of charging pile monitoring data extracts charging cable temperature, charging current, charging voltage, ambient temperature, fan speed, and fan current, which are then input into the collaborative risk assessment module to generate the collaborative risk assessment value and corresponding risk level label for the current period. After risk level determination, the process proceeds to either Level 1, Level 2, or Level 3 early warning stages based on the determination result: When the risk level is determined to be Level 1, the Level 1 early warning process is executed, calling the isolated forest model to identify the charging status for the current period. If the risk level remains at Level 1 for three consecutive periods, a risk escalation mechanism is triggered, upgrading the risk level from Level 1 to Level 2. When the risk level is determined to be Level 2, the Level 2 early warning process is executed. Charging pile monitoring data from previous periods is extracted and input into the LSTM prediction model, outputting a risk forward parameter set. This set is combined with actual values ​​to construct a deviation vector, which is then input into the XGBoost classification model to generate an early warning signal. If any forward risk parameter output by the LSTM prediction model exceeds the operating limit, a risk escalation mechanism is triggered, raising the risk level to Level 3. If all forward risk parameters are within the operating limit, the current risk level is maintained, and data acquisition continues to monitor trend changes. When the risk level is determined to be Level 3, a Level 3 early warning process is executed. Images of the charging gun interface and charging pile monitoring data are acquired and input into the image recognition convolutional network and heatmap feature extraction network, respectively. Abnormal image indicators and coupled mutation features are extracted and fused to construct a joint feature set to generate a Level 3 early warning signal. At this stage, an emergency handling mechanism is triggered, immediately cutting off the current charging circuit to prevent fault propagation. In the Level 3 risk stage, a manual misjudgment confirmation signal mechanism is also introduced. If a misjudgment confirmation signal is received, a risk level reset operation is performed, resetting the current risk level to normal charging and re-entering the risk judgment process, thus achieving closed-loop feedback control.

[0063] This implementation plan effectively enhances the dynamic identification, accurate assessment, and tiered response capabilities for charging safety risks by introducing a hierarchical, collaboratively evolving intelligent early warning system for charging piles. By deeply integrating the multidimensional features of charging pile monitoring data with collaborative risk assessment values, and constructing targeted model structures and training strategies based on risk levels, a closed-loop response mechanism from data perception to model inference and then to risk intervention is achieved. While maintaining the stability and real-time nature of risk assessment, this system enhances the sensitivity and adaptability to hidden risks in complex charging scenarios, providing more reliable, intelligent, and practical technical support for the actual operation of charging piles.

[0064] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0065] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A charging pile intelligent early warning method based on hierarchical collaborative evolution, characterized in that, Includes the following steps: S1 collects charging pile monitoring data in real time, and performs time alignment, weighted sliding filter, exponential smoothing and normalization compression processing on the charging pile monitoring data in sequence to obtain pre-processed charging pile monitoring data. S2, set the judgment period, evaluate the collaborative risk of the current period based on the preprocessed charging pile monitoring data, generate risk level labels, and store the risk level labels and corresponding monitoring data in a structured manner to build a charging risk dataset; The specific steps for setting the judgment period and assessing the collaborative risk of the current period based on the preprocessed charging pile monitoring data are as follows: Set a fixed-width sliding window as a judgment period, calculate the average charging current, average grounding resistance, and average temperature difference between the charging cable and the ambient temperature in each period, and calculate the difference between the maximum and minimum charging current in the period to obtain the charging current fluctuation value. Extract real-time charging pile monitoring data, divide the charging current fluctuation value by the ratio of the average charging current, and multiply the resulting ratio by the first time derivative of the real-time charging cable temperature to obtain the electrothermal dynamic response term; subtract the ratio of the real-time fan speed to the rated fan speed from the ratio of the real-time fan current to the rated fan current, take the absolute value of the difference and add one to obtain the fan operation offset amplification term; divide the difference between the real-time cable temperature and the real-time ambient temperature by the average temperature difference value, and add the ratio of the real-time grounding resistance to the average grounding resistance to obtain the geothermal risk subordination term; multiply the electrothermal dynamic response term, the fan operation offset amplification term, and the geothermal risk subordination term in sequence to obtain the collaborative risk assessment value; S3 calls the charging risk dataset and constructs an isolated forest model, an LSTM prediction model, and an XGBoost classification model for samples with different risk levels. It also constructs a weighted bias metric and a training expression function to achieve multi-model collaborative learning and training. S4, based on real-time risk level labels and the trained model, performs dynamic updates of risk levels and triggers an intelligent hierarchical early warning process: for level 1 risks, the Isolation Forest model is invoked to output a level 1 early warning signal; for level 2 risks, the LSTM and XGBoost models are combined to output a level 2 early warning signal. The three-level risk fusion image recognition convolutional network and heatmap feature extraction network output a three-level early warning signal.

2. The intelligent early warning method for charging piles based on hierarchical collaborative evolution according to claim 1, characterized in that: The specific steps for acquiring real-time charging pile monitoring data and performing time alignment, weighted sliding filtering, exponential smoothing, and normalization compression on the charging pile monitoring data to obtain preprocessed charging pile monitoring data are as follows: Real-time collection of charging pile monitoring data, including charging cable temperature, charging current, charging voltage, grounding resistance, ambient temperature, fan speed, and fan current; By constructing a multi-source data alignment mechanism based on the consistency constraint of the sliding time window and sampling frequency, the charging pile monitoring data is synchronously fused in the time dimension; the sliding average filtering algorithm is used to suppress the noise of the charging pile monitoring data and weaken the high-frequency fluctuations caused by sensor jitter and electromagnetic interference; the exponentially weighted moving average algorithm is used to model the short-term trend and dynamically smooth the charging pile monitoring data; a normalization algorithm based on Z-score normalization and interquartile range compression is used to normalize the charging pile monitoring data and unify the dimension.

3. The intelligent early warning method for charging piles based on hierarchical collaborative evolution according to claim 1, characterized in that: The specific steps for generating risk level labels, storing the risk level labels and the corresponding monitoring data in a structured manner, and constructing a charging risk dataset are as follows: Compare the collaborative risk assessment value S with the multi-level risk thresholds S1, S2, and S3 to generate risk level labels: When S ≤ S1, mark the current period as normal charging; When S1 < S ≤ S2, mark the current period as a first-level risk; When S2 < S < S3, mark the current period as a second-level risk; When S ≥ S3, mark the current period as a third-level risk; Structurally store the risk level label, charging pile monitoring data, and collaborative risk assessment value of the current period to construct a charging risk dataset.

4. The intelligent early warning method for charging piles based on hierarchical collaborative evolution according to claim 1, characterized in that: The specific steps for calling the charging risk dataset, constructing an isolation forest model, an LSTM prediction model, and an XGBoost classification model for different risk level samples respectively, and constructing a weighted deviation metric and a training expression function to achieve multi-model collaborative learning and training are as follows: Based on the charging risk dataset, extract the charging pile monitoring data and collaborative risk assessment values corresponding to normal charging and first-level risks, construct an isolation forest model, and determine the number and depth of decision trees in the model through multiple rounds of training iterations to complete the training of the isolation forest model; Extract the charging cable temperature, charging current, charging voltage, ambient temperature, and fan speed corresponding to the second-level risk, construct an LSTM prediction model, and construct a time series weighted loss function. Use the time series weighted loss function value as the optimization target to train the LSTM prediction model, and output the risk forward parameter group of the current period. The risk forward parameter group includes the predicted values of the charging cable temperature, charging current, charging voltage, ambient temperature, and fan speed; Based on the risk forward parameter group output by the LSTM prediction model and the actual values of the charging pile monitoring data in the corresponding period, construct an XGBoost classification model, and combine the corresponding collaborative risk assessment value. Through the gradient boosting mechanism and parameter optimization, obtain the predicted training expression value as the training feature to complete the training of the XGBoost classification model.

5. The intelligent early warning method for charging piles based on hierarchical collaborative evolution according to claim 4, characterized in that: The specific steps for constructing the time series weighted loss function and outputting the time series weighted loss function value are as follows: The prediction deviation metric is calculated by summing the squared differences between the predicted and actual values ​​of all forward parameters for risk within the t-th secondary risk period. The time decay factor is obtained by multiplying the time decay coefficient by the current period index and taking the opposite number, and then taking the power function value of the opposite number with the natural constant e as the base. The prediction deviation metric is multiplied by the time decay factor to obtain the weighted deviation value. The time-series weighted loss function value is obtained by taking the arithmetic mean of the weighted deviation values ​​for all secondary risk periods.

6. The intelligent early warning method for charging piles based on hierarchical collaborative evolution according to claim 4, characterized in that: The specific steps for obtaining the predicted training expression value through gradient boosting mechanism and parameter optimization are as follows: Subtract the predicted and actual values ​​of all risk forward parameters in the t-th secondary risk period to obtain the error term for each risk forward parameter. Square each error term and divide it by the sum of the corresponding actual value and the minima to obtain the normalized error for each risk forward parameter. Sum the normalized errors of all risk forward parameters to obtain the normalized error metric for the current period. Multiply the risk sensitivity coefficient, the collaborative risk assessment value of the t-th secondary risk period, and the corresponding normalized error metric, add one, and take the natural logarithm of the sum to obtain the risk-weighted training error. The arithmetic mean of the risk-weighted training errors over all secondary risk periods is used to obtain the predicted training expression value.

7. The intelligent early warning method for charging piles based on hierarchical collaborative evolution according to claim 1, characterized in that: The specific steps for dynamically updating the risk level based on real-time risk level labels and the trained model are as follows: When the risk level is at level 1 for M consecutive judgment periods, the current risk level will be upgraded from level 1 to level 2. When the risk level is at level 2, if any forward risk parameter output by the LSTM prediction model exceeds the operating limit, the current risk level will be raised to level 3. When the risk level is at level 3, if a manual misjudgment confirmation signal is received, a risk level reset operation will be performed, marking the current risk level as normal charging and restarting the risk level judgment process.

8. The intelligent early warning method for charging piles based on hierarchical collaborative evolution according to claim 1, characterized in that: The specific steps for triggering the intelligent hierarchical early warning process are as follows: For Level 1 risk, the isolated forest model is invoked to output a Level 1 early warning signal; for Level 2 risk, the LSTM and XGBoost models are combined to output a Level 2 early warning signal; for Level 3 risk, an image recognition convolutional network and a heatmap feature extraction network are fused to output a Level 3 early warning signal. When the risk level is Level 1, statistical calculations are performed on the charging current and charging voltage in the current cycle to construct a time-domain feature vector, which is then input into the isolated forest model to output a Level 1 early warning signal. When the risk level is level 2, the charging pile monitoring data of the previous K periods are extracted, input into the LSTM prediction model, and the risk forward parameter set of the current period is output. The risk forward parameter set includes the predicted values ​​of charging cable temperature, charging current, charging voltage, ambient temperature and fan speed. The deviation value between the risk forward parameter set and the actual value corresponding to the current period is calculated, the deviation vector is constructed and input into the XGBoost classification model, and the level 2 early warning signal is output. When the risk level is level three, the charging gun interface image and the corresponding charging pile monitoring data for the current period are obtained. The charging gun interface image is input into the image recognition convolutional network for processing and the image anomaly index is output. The charging pile monitoring data is converted into a heat map and input into the heat map feature extraction network to output coupled mutation features. The image anomaly index and coupled mutation features form a joint feature group and are input into a multi-layer convolutional neural network to output a level three warning signal. The warning signal includes the risk level, model type, and trigger time slice; for the level 1 warning signal, it is output through the local human-machine interface of the charging pile, indicating that the current operating status has deviated from the normal range and needs to be continuously monitored. For Level 2 warning signals, the on-site audible and visual alarm device is triggered, and the warning signal is pushed to the terminal device of the registered maintenance personnel through the remote communication interface; For a Level 3 warning signal, the current charging circuit will be immediately cut off, based on the Level 2 warning signal.

9. A charging pile intelligent early warning system based on hierarchical collaborative evolution, characterized in that: include: The module comprises a data acquisition and preprocessing module, a collaborative risk assessment module, a multi-model integrated training module, and an intelligent hierarchical early warning module, among which: The data acquisition and preprocessing module is used to acquire charging pile monitoring data in real time, and to perform time alignment, weighted sliding filtering, exponential smoothing and normalization compression processing on the charging pile monitoring data in sequence to obtain preprocessed charging pile monitoring data. The collaborative risk assessment module is used to set the judgment period, assess the collaborative risk of the current period based on the preprocessed charging pile monitoring data, generate risk level labels, and store the risk level labels and corresponding monitoring data in a structured manner to construct a charging risk dataset. The specific steps for setting the judgment period and assessing the collaborative risk of the current period based on the preprocessed charging pile monitoring data are as follows: Set a fixed-width sliding window as a judgment period, calculate the average charging current, average grounding resistance, and average temperature difference between the charging cable and the ambient temperature in each period, and calculate the difference between the maximum and minimum charging current in the period to obtain the charging current fluctuation value. Extract real-time charging pile monitoring data, divide the charging current fluctuation value by the ratio of the average charging current, and multiply the resulting ratio by the first time derivative of the real-time charging cable temperature to obtain the electrothermal dynamic response term; subtract the ratio of the real-time fan speed to the rated fan speed from the ratio of the real-time fan current to the rated fan current, take the absolute value of the difference and add one to obtain the fan operation offset amplification term; divide the difference between the real-time cable temperature and the real-time ambient temperature by the average temperature difference value, and add the ratio of the real-time grounding resistance to the average grounding resistance to obtain the geothermal risk subordination term; multiply the electrothermal dynamic response term, the fan operation offset amplification term, and the geothermal risk subordination term in sequence to obtain the collaborative risk assessment value; The multi-model integrated training module is used to call the charging risk dataset, construct the Isolation Forest model, LSTM prediction model and XGBoost classification model for samples of different risk levels, and construct the weighted bias metric and training expression function to realize multi-model collaborative learning training. The intelligent hierarchical early warning module is used to dynamically update the risk level based on real-time risk level labels and the trained model, and trigger the intelligent hierarchical early warning process: for level 1 risk, the isolated forest model is invoked to output a level 1 early warning signal; for level 2 risk, the LSTM and XGBoost models are combined to output a level 2 early warning signal; for level 3 risk, the image recognition convolutional network and the heat map feature extraction network are fused to output a level 3 early warning signal.

Images