Semi-supervised learning-based direct current charging pile health state assessment method

Abstract

The invention discloses a semi-supervised learning-based direct-current charging pile health state evaluation method, and aims to solve the problems that a direct-current charging pile operation state comprehensive evaluation system is incomplete, data dimensions are not comprehensive enough, and data preprocessing is difficult in a large-scale real scene in the prior art. Through a semi-supervised learning method of data preprocessing, feature index design, feature index correlation analysis and hybrid consistency regularization and pseudo label generation, key health indexes are extracted from multi-dimensional data of the DC charging pile, and a comprehensive health state evaluation system is established. And an efficient operation state classification model is trained by using a semi-supervised learning method, so that the accuracy and reliability of fault diagnosis are improved. According to the method, the task load of labeling the equipment by operation and maintenance personnel can be effectively reduced, the operation and maintenance efficiency is improved, and stable operation and safe use of the charging pile are ensured.

Description

Technical Field

[0001] The present invention relates to the technical field of electric vehicle charging infrastructure, and particularly to a method for evaluating the health status of a DC charging pile based on semi-supervised learning. Background Art

[0002] With the rapid development of the global electric vehicle industry, DC charging piles have become an important part of electric vehicle charging infrastructure, assuming the dual responsibilities of enhancing the user charging experience and ensuring grid stability. The charging pile is not only the core equipment to support the green travel of electric vehicles, but also plays a key role in regulating grid load and improving energy efficiency. Therefore, the operation stability and reliability of charging piles are crucial for the sustainable development of the electric vehicle industry.

[0003] DC charging piles have significant advantages in promoting green transportation. However, their high-power output and complex structural design also result in a relatively high failure frequency and a wide variety of failure types. Especially during the charging process, equipment failures may lead to serious safety accidents, such as battery fires and electrical fires. These problems not only threaten the safety of users, but also pose challenges to the stability of the grid system. Therefore, the operation status monitoring and fault diagnosis of DC charging piles have become the focus of attention for electric vehicle users, grid companies, and regulatory agencies.

[0004] Currently, the operation and maintenance of DC charging piles mainly include:

[0005] 1. Fault detection methods based on traditional signal processing and algorithms. This technical solution mainly relies on traditional signal processing techniques, such as wavelet transform and Fourier transform, combined with classical machine learning algorithms (such as support vector machine SVM, decision tree, etc.) for charging pile fault detection. By performing frequency-domain or time-domain analysis on the electrical signals (such as voltage and current waveforms) collected during the operation of the charging pile, fault features are extracted, and then classified and diagnosed through machine learning algorithms. This method can accurately detect electrical anomalies or hardware failures, such as current overload and short circuit, and has a good diagnostic effect for simple fault modes.

[0006] Fault detection methods based on traditional signal processing and algorithms have the advantages of certain real-time performance and low computational cost. However, in the face of complex fault types and multi-dimensional monitoring, their performance is relatively limited, and they lack automation and adaptability. Therefore, this method is suitable for fault detection in simple scenarios, but its applicability is relatively limited in the large-scale application and changing environment of modern charging piles.

[0007] 2. Fault diagnosis method based on deep learning, which uses deep learning technologies (such as deep neural network DNN, convolutional neural network CNN, etc.) to diagnose the faults of charging piles. By classifying faults for simulated data or actual operation data, it can identify the fault types of the equipment under various operating conditions. This method can extract fault features by learning a large amount of historical data, improving the accuracy of fault diagnosis.

[0008] This method usually only considers the current operation data, ignoring other important factors that may affect the health status of the charging pile, such as temperature changes and user behavior, resulting in incomplete diagnosis. At the same time, deep learning models have high requirements for data quality and sample size. When dealing with noise or outliers in real-world data, problems such as inaccurate diagnosis or insufficient robustness may occur. In practical applications, deep learning models require a large amount of computing resources, and the training process is complex, which may limit their practical applications in resource-constrained environments. Summary of the Invention

[0009] The present invention proposes a semi-supervised learning method combining hybrid consistency regularization and pseudo-label generation, aiming to extract key health indicators from the multi-dimensional data of DC charging piles, establish a comprehensive health status evaluation system, and use semi-supervised learning methods to train an efficient operating state classification model, thereby improving the accuracy and reliability of fault diagnosis, providing effective support for the intelligent monitoring and fault warning of charging piles, and ensuring the stable operation and safe use of charging piles. The present invention can use semi-supervised learning algorithms to train a DC charging pile health status estimation model in scenarios of small-scale labeled data and large-scale unlabeled data. The purpose of the present invention is achieved through the following technical solutions:

[0010] A method for evaluating the health status of a DC charging pile based on semi-supervised learning, comprising the following steps:

[0011] S1. Data preprocessing: Perform data preprocessing on the original working data of the DC charging pile to improve the reliability of the data;

[0012] S1 includes the following sub-steps:

[0013] (1) Removing outliers: According to relevant national standards and industry regulations, remove serious faults or invalid data, and eliminate obviously abnormal data;

[0014] (2) Handling missing data: For key indicators with a large number of missing values, adopt a discard process to avoid negative impacts on the analysis results.

[0015] S2. Feature index design: Extract feature indexes from three dimensions of safety performance, electrical performance, and operation efficiency to characterize the actual operation state of the charging pile and identify potential fault situations;

[0016] The characteristic indicators in S2 include:

[0017] Safety performance evaluation indicators: average operating temperature, insulation failure frequency, equipment service life;

[0018] Electrical performance evaluation indicators: downtime duration, failure frequency, average load, power deviation;

[0019] Operating performance evaluation indicators: zero - energy order ratio, abnormal termination order ratio, average energy per order, cumulative operating time.

[0020] S3. Correlation analysis of characteristic indicators: Determine the characteristic indicators with strong correlation with the classification results of the charging pile operating status through correlation analysis as health factors;

[0021] The specific methods of S3 include:

[0022] First, perform Z - Score standardization processing on the index data. The formula is as follows:

[0023]

[0024] where, F i is the standardized value of the i - th index, S i is the original value of the i - th index, is the average value of this index, and σ is the standard deviation of this index.

[0025] The information gain IG(S i ) of each characteristic indicator is calculated by the following formula:

[0026] IG(S i ) = H(Y) - H(Y|S i )

[0027]

[0028] where, p(s ik ) is the probability that the feature S i takes the value s ik ; p(y j ) is the probability that the classification label Y takes the value y i ;

[0029] After calculating the information gain, select the feature with the highest information gain, that is, the indicator with the greatest contribution to the classification result, as the health factor for the health status classification of DC charging piles.

[0030] S4. Semi-supervised learning method using hybrid consistency regularization and pseudo-label generation: Utilize limited labeled data and unlabeled data to enhance the classification performance of the model by expanding the training set, and output the evaluation result of the health status of DC chargers.

[0031] The semi-supervised learning method of consistency regularization and pseudo-label generation includes:

[0032] (1) Use a multi-layer feedforward neural network model to classify the status of chargers.

[0033] (2) Use labeled data to train a preliminary model, generate pseudo-labels and expand the training set.

[0034] (3) Through the consistency regularization strategy, ensure that the model outputs consistent pseudo-labels when processing original samples and augmented samples.

[0035] (4) Optimize the model parameters through the backpropagation algorithm to train the evaluation model of the health status of DC chargers.

[0036] (5) Use the trained optimal model to output the evaluation result of the health status of DC chargers.

[0037] The technical effects of the present invention are:

[0038] (1) The method for estimating the health status of DC chargers based on semi-supervised learning can comprehensively analyze multi-dimensional data including electrical, operation, and safety indicators extracted from the operation data of DC chargers over a certain period, improving the accuracy of estimating the health status of DC chargers.

[0039] (2) The method for estimating the health status of DC chargers based on semi-supervised learning can train a high-accuracy charger health status estimation model relying only on a small part of labeled data, reducing the task volume of equipment tagging by operation and maintenance personnel and improving the operation and maintenance efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] Figure 1 It is a flowchart of the semi-supervised learning method of hybrid consistency regularization and pseudo-label generation of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0041] Combine the embodiments to illustrate the specific technical solutions of the present invention.

[0042] The present invention includes the following steps:

[0043] S1. Data preprocessing

[0044] The working data of the DC charging pile received by the charging pile cloud platform has low precision and high sparsity (the sampling time interval is 3 minutes). Therefore, it is necessary to first preprocess the original data, including removing outliers and handling missing data, to improve the reliability of subsequent analysis.

[0045] Outliers are usually caused by sensor failures or data input errors, which may seriously interfere with the learning process of subsequent models. The first step in handling outliers is to remove those severely faulty or invalid data according to relevant national standards and industry regulations; the second step is to eliminate obviously abnormal data, such as the case where the output voltage is less than zero. The principle of handling missing data is to retain valid data as much as possible while ensuring data quality. For key indicators with a large number of missing values (such as voltage and current), discard processing should be adopted to avoid negative impacts on the analysis results.

[0046] S2. Feature Index Design

[0047] The method proposed in the present invention extracts feature indexes from three dimensions of safety performance, electrical performance and operation efficiency to characterize the actual operation state of the charging pile and identify potential fault situations to improve the accuracy of the health state estimation of the DC charging pile.

[0048] The feature indexes specifically include:

[0049] 1. Safety Performance Evaluation Indexes

[0050] 1) Average Working Temperature (S 1 ):During the charging process, the equipment generates heat. Excessive temperature will accelerate component aging, reduce insulation performance, and even pose a fire risk. Therefore, monitoring this index can detect overheating problems early, carry out temperature reduction intervention in time, and improve the safety and reliability of the equipment.

[0051] 2) Insulation Fault Frequency (S 2 ):Frequent insulation faults may lead to serious accidents such as short circuits and electric leaks. Therefore, monitoring this index is the key to evaluating the safety performance of the charging pile. By tracking these faults, the root causes can be identified and the maintenance strategy can be optimized, thereby reducing the failure rate and improving the overall safety.

[0052] 3) Equipment Service Life (S 3 ):A longer service life usually means better adaptability and safety of the equipment in various environments. However, long-term operation may also lead to equipment wear, aging and performance degradation. Therefore, regular maintenance and evaluation are essential. Analyzing this index helps to formulate effective equipment replacement and upgrade strategies to ensure that the charging pile maintains the best safety performance throughout its service life.

[0053] 2. Electrical Performance Evaluation Indexes

[0054] 1) Downtime duration (S 4 ):Refers to the total time that the charging pile is out of operation due to faults, which directly reflects the reliability and availability of the charging pile. Therefore, monitoring and minimizing the downtime duration is crucial for improving the performance of the charging pile. Regular maintenance and fault diagnosis help reduce the downtime and ensure the efficient and stable operation of the equipment.

[0055] 2) Fault frequency (S 5 ):Indicates the total number of faults that occur within a specific time period and is an important indicator for evaluating the stability and failure rate of the equipment. A higher fault frequency may indicate design, manufacturing, or operation problems, posing potential safety risks. By analyzing this indicator, fault modes and root causes can be identified, and then targeted improvement measures can be taken.

[0056] 3) Average load (S 6 ):Refers to the ratio of the actual average power output of the charging pile within a certain time to its rated power, and the calculation formula is as follows:

[0057]

[0058] Among them, represents the actual power output of the charging pile during this period, and P 0 represents the rated power of the charging pile. This indicator reflects the load situation of the equipment. A higher average load may lead to overload, increased heat generation, and an increased risk of faults or safety hazards. Monitoring this indicator helps optimize the use of the charging pile and ensure that the equipment operates within a safe load range.

[0059] 4) Power deviation (S 7 ):Refers to the average deviation between the actual output power of the charging pile and the desired output power of the battery management system (BMS), and the calculation formula is as follows:

[0060]

[0061] Among them, represents the actual power output of the charging pile during this period, represents the power required by the BMS. This indicator is crucial for evaluating the accuracy and reliability of the charging pile. A significant power deviation may reduce the charging efficiency and may even damage the battery, so it needs to be monitored regularly.

[0062] 3. Operating performance evaluation indicators

[0063] 1) Zero-energy order ratio (S 8):This indicator represents the proportion of orders with zero energy in the charging orders to the total number of orders. A higher proportion may indicate resource constraints, resulting in the charging piles being unable to complete the charging tasks normally, thereby reducing the operating efficiency. Monitoring this indicator helps the operators detect potential equipment failure problems in a timely manner and carry out necessary maintenance and optimization.

[0064] 2) Proportion of abnormally terminated orders (S 9 ):This indicator measures the proportion of orders abnormally terminated due to faults or other problems to the total number of orders. Abnormally terminated orders may reflect faults in the equipment or system, affecting the normal progress of the charging service. By analyzing this indicator, equipment faults can be identified and repaired in a timely manner to ensure the stable operation of the charging piles.

[0065] 3) Average energy per order (S 10 ):This indicator represents the average energy delivered by each charging order. A lower average energy per order may mean a lower utilization rate of the charging piles or the equipment failing to meet the charging needs of users. Monitoring and analyzing this indicator helps identify potential equipment failures or abnormal operations, thereby improving the use of the equipment and optimizing the service.

[0066] 4) Cumulative operating time (S 11 ):This indicator represents the total operating duration of the charging piles, reflecting the usage frequency of the charging piles and the reliability of the equipment. A longer cumulative operating time usually indicates a higher utilization rate of the equipment, but from another perspective, it also means that the equipment may face the risks of aging and failure. Therefore, monitoring this indicator helps evaluate the health status of the equipment, carry out maintenance and replacement in a timely manner, and ensure the long-term stability of the charging piles.

[0067] S3. Correlation analysis of characteristic indicators

[0068] In order to determine the characteristic indicators that have a strong correlation with the classification results of the operating status of the charging piles and then identify them as health factors, correlation analysis needs to be carried out. This process identifies the key indicators that significantly affect the status assessment of the charging piles by evaluating the statistical relationship between the selected characteristics and the classification results. These health factors will be able to improve the accuracy and reliability of the subsequent status classification model.

[0069] When conducting the correlation analysis between the characteristic indicators and the classification of the operating status of the charging piles, first perform Z-Score standardization processing on the indicator data. The formula is as follows:

[0070]

[0071] where, F i is the standardized value of the i-th indicator, S i is the original value of the i-th indicator, is the average value of this index, and σ is the standard deviation of this index. Based on the entropy concept in information theory, information gain quantifies the impact of each feature on the target category.

[0072] In this embodiment, the changes in features such as failure frequency, downtime duration, and power deviation have different degrees of influence on the charging pile status (i.e., the classification label). The information gain IG(S i ) of each feature index is calculated as follows:

[0073] IG(S i ) = H(Y) - H(Y|S i )

[0074]

[0075] where p(s ik ) is the probability that the feature S i takes the value s ik ; p(y j ) is the probability that the classification label Y takes the value y j ; after calculating the information gain, the feature with the highest information gain, that is, the index with the greatest contribution to the classification result, is selected as the health factor for classifying the health status of the DC charging pile.

[0076] S4. Semi-supervised learning method combining hybrid consistency regularization and pseudo-label generation

[0077] The present invention combines a semi-supervised learning method of hybrid consistency regularization and pseudo-label generation, aiming to effectively utilize limited labeled data and improve the classification performance of the model by expanding the training set. The structure is as Figure 1 shown.

[0078] The semi-supervised learning method of hybrid consistency regularization and pseudo-label generation mainly includes the following steps:

[0079] (1) Multi-layer feedforward neural network model

[0080] This method selects a multi-layer feedforward neural network as the model architecture for classifying the charging pile status. In each training, the model is trained using a mixed dataset of labeled data and pseudo-labeled data. The loss function combines cross-entropy loss and consistency regularization loss, and optimizes the classification performance and robustness through weighted combination. After each round of training, the model continues to be used for the next round of pseudo-label generation and training, gradually improving the performance of the model.

[0081] (2) Pseudo-label generation

[0082] First, use the labeled data to train a preliminary model to classify the operating status of the charging pile. After training, use the model to predict the unlabeled data and generate pseudo-labels. To ensure the high quality of the pseudo-labels, select those samples with higher prediction confidence and add them to the training set. In this way, by continuously generating pseudo-labels, the training set is gradually expanded, and the generalization ability of the model is improved.

[0083] (3) Consistency regularization training

[0084] To improve the robustness of the model, the research introduced a consistency regularization strategy. This strategy requires the model to output consistent pseudo-labels when processing the original samples and the augmented samples. By comparing the prediction results of the original samples and the augmented samples, minimize the difference between the two, so as to ensure that the model can maintain stable predictions in the face of data perturbations, and improve the classification accuracy and stability.

[0085] (4) Output result

[0086] Finally, the optimal model obtained through training outputs the evaluation results of the health status of the DC charging pile. The results are divided into four levels: A, B, C, and D, corresponding to A: completely normal and operable state; B: latent fault and operable state; C: obvious fault and operable state; D: fault and inoperable state

[0087] According to the above steps, it is possible to extract health factors characterizing the health status of the DC charging pile based on the operating condition data of the DC charging pile received by the charging pile big data platform, and train a DC charging pile health status estimation model through semi-supervised learning based on a small number of labels given by the operation and maintenance personnel, output the health status rating of the DC charging pile, and further serve as a reference for the operation and maintenance personnel to maintain and repair the DC charging equipment, improve the operation and maintenance efficiency, and reduce the operation cost and the equipment failure risk.

Claims

1. A method for evaluating the health status of a DC charging pile based on semi-supervised learning, characterized in that: The following steps are involved: S1. Data preprocessing: preprocess the original working data of the DC charging pile to improve the reliability of the data; S2. Characteristic index design: Extract characteristic indicators from three dimensions: safety performance, electrical performance and operating efficiency to characterize the actual operating status of the charging pile and identify potential fault conditions; S3. Correlation analysis of characteristic indicators: through correlation analysis, characteristic indicators with strong correlation with the classification results of the charging pile operation status are determined as health factors; S4, semi-supervised learning method using hybrid consistency regularization and pseudo-label generation: using limited labeled data and unlabeled data, the classification performance of the model is improved by expanding the training set; Output the DC charging pile health status assessment results.

2. A method for evaluating the health status of a DC charging pile based on semi-supervised learning according to claim 1, characterized in that: S1 includes the following sub-steps: (1) Removing outliers: According to relevant national standards and industry regulations, remove serious faults or invalid data, and eliminate data with obvious abnormalities; (2) Handling missing data: For key indicators with a large number of missing values, discard them to avoid negative impact on the analysis results.

3. A method for evaluating the health status of a DC charging pile based on semi-supervised learning according to claim 1, characterized in that: The characteristic indicators in S2 include: Safety performance evaluation indicators: average operating temperature, insulation failure frequency, equipment service life; Electrical performance evaluation indicators: downtime, fault frequency, average load, power deviation; Operation performance evaluation indicators: proportion of zero-energy orders, proportion of abnormally terminated orders, average energy per order, and cumulative operating time.

4. A method for evaluating the health status of a DC charging pile based on semi-supervised learning according to claim 1, characterized in that: S3 specific methods include: First, the indicator data is standardized by Z-Score, and the formula is as follows: Among them, F i is the standardized value of the ith indicator, S i is the original value of the ith indicator, is the average value of the indicator, and σ is the standard deviation of the indicator; The information gain IG(S) of each feature index i ) is calculated as follows: IG(S i )=H(Y)-H(Y|S i ) Among them, p(s ik ) is the characteristic S i The value is s ik The probability of p(y j ) is the classification label Y with the value y j probability; After calculating the information gain, the feature with the highest information gain, that is, the indicator that contributes the most to the classification result, is selected as the health factor for the health status classification of the DC charging pile.

5. A method for evaluating the health status of a DC charging pile based on semi-supervised learning according to claim 1, characterized in that: S4 specific methods include: (1) Use a multi-layer feedforward neural network model to classify charging pile status; (2) Use the labeled data to train the preliminary model, generate pseudo labels and expand the training set; (3) Through the consistency regularization strategy, ensure that the model outputs consistent pseudo labels when processing original samples and enhanced samples; (4) Optimize model parameters through back propagation algorithm and train the DC charging pile health status assessment model; (5) Use the trained optimal model to output the health status assessment results of the DC charging pile.

Images