A fault detection method for an electric vehicle charging station

Abstract

The application provides a fault detection method of an electric vehicle charging station, the method comprising: acquiring multiple types of data, the multiple types of data comprising equipment operation state data, environmental influence data and abnormal label information; performing standardization preprocessing on the multiple types of data to obtain processed multiple types of data; inputting the processed multiple types of data into a pre-established multi-source data feature enhancement and space-time fusion model to obtain intermediate data after feature enhancement and space-time fusion; inputting the intermediate data into a pre-established double-constraint abnormality detection and fault root cause decision model to determine a fault main propagation path and determine a fault root cause equipment according to the fault main propagation path; and determining operation and maintenance decision disposal information according to a pre-established type label library and operation and maintenance knowledge base, the fault root cause equipment and the fault main propagation path. At least reliable technical support can be provided for safe and stable operation of the charging station, and operation and maintenance cost and safety risk are significantly reduced.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle charging station operation and maintenance technology, and more specifically, to a fault detection method for electric vehicle charging stations. Background Technology

[0002] With the deepening implementation of the "dual carbon" goals and the explosive growth of the electric vehicle industry, charging stations, as the core infrastructure supporting the popularization of electric vehicles, are experiencing a continuous increase in network coverage density and operational load. According to data from the China Electric Vehicle Charging Infrastructure Promotion Alliance, as of June 2024, the total number of charging infrastructure units nationwide reached 6.3 million, a 45% increase compared to the same period in 2023. Among them, the number of public charging stations exceeded 150,000, with an average of over 12 million charging services per day. The continuous and stable operation of charging stations is not only directly related to users' charging experience and travel safety, but also of significant strategic importance for ensuring the stable operation of urban energy networks and promoting the high-quality development of the new energy vehicle industry.

[0003] Current charging station fault detection technologies are mainly divided into two categories: one is based on traditional statistical and threshold methods, such as Principal Component Analysis (PCA) and threshold judgment methods; the other is based on single data-driven models, such as Sparse Autoencoder (SAE) and Multilayer Perceptrons (MLP). Although these technologies have achieved fault detection functions to a certain extent, they still have significant technical shortcomings in the face of the diverse data sources, complex operating conditions, and multi-source nature of charging stations, making it difficult to meet actual operation and maintenance needs. Summary of the Invention

[0004] In view of this, the present invention discloses a fault detection method for electric vehicle charging stations, which can at least integrate physical constraints and spatiotemporal characteristics, adapt to small sample scenarios, and have both detection accuracy and traceability capabilities, providing reliable technical support for the safe and stable operation of charging stations and significantly reducing operation and maintenance costs and safety risks.

[0005] Specifically, the present invention is achieved through the following technical solutions:

[0006] This application proposes a fault detection method for electric vehicle charging stations, the method comprising:

[0007] Acquire various types of data, including equipment operating status data, environmental impact data, and anomaly label information;

[0008] The various data are standardized and preprocessed to obtain the processed data.

[0009] The processed data are input into a pre-established multi-source data feature enhancement and spatiotemporal fusion model to obtain intermediate data after feature enhancement and spatiotemporal fusion.

[0010] The intermediate data is input into a pre-established dual-constraint anomaly detection and fault tracing decision model to determine the main fault propagation path, and the root cause device of the fault is determined based on the main fault propagation path.

[0011] Based on the pre-established type tag library and operation and maintenance knowledge base, the fault root cause device, and the fault main propagation path, the operation and maintenance decision-making and handling information is determined.

[0012] Optionally, the various types of data undergo standardization preprocessing, including:

[0013] The various types of data are cleaned to obtain cleaned data.

[0014] By combining statistical criteria and physical constraints, outlier data is removed from various cleaned datasets, resulting in a variety of reliable data.

[0015] Spatiotemporal standardization is performed on the various reliable data to obtain the processed data.

[0016] Optionally, the data cleaning steps for the various types of data include:

[0017] The filling method is performed using time-series weighted interpolation.

[0018] The time is calculated using the following formula. missing values :

[0019] ;

[0020] in, For time decay weight, This is a decay coefficient, ensuring that data from adjacent time points contribute more to missing values; , Missing time points Normal data from consecutive moments;

[0021] The steps for performing spatiotemporal standardization on the various types of trusted data include:

[0022] The various trusted data are divided into segments according to a fixed time window to obtain a variety of trusted data in a unified format for time series data;

[0023] Construct a three-dimensional spatial correlation matrix of charging piles, power grid, and energy storage system to determine the electrical coupling relationships between devices;

[0024] The electrical coupling relationship between devices is determined by the following formula:

[0025] ;

[0026] in, For covariance, Standard deviation, Range of values The larger the absolute value, the better the equipment. and The stronger the correlation between the parameters.

[0027] Optionally, a multi-source data feature enhancement and spatiotemporal fusion model can be established through the following steps:

[0028] We designed a physically constrained sparse autoencoder and a PC-SAE loss function that embed the physical laws of the charging system, and added a combination strategy of oversampling and undersampling to output a balanced sample set based on multiple training samples.

[0029] The method employs a fusion of temporal feature extraction using a multi-scale temporal convolutional network (MS-TCN) and spatial feature aggregation using a graph attention network (GAT), combined with cross-modal attention, to output fused features by balancing the spatiotemporal coupling of the captured data of each sample in the sample set.

[0030] Optionally, the physical constraint sparse autoencoder of the physical laws of the charging system includes an encoder, a decoder and a physical constraint layer;

[0031] The PC-SAE loss function is:

[0032] ;in, , , These are the loss weighting coefficients;

[0033] The reconstruction loss is determined by the following formula. :

[0034] ;

[0035] in, Reconstructed data output from the PC-SAE network. For the sample size, The time window length, For feature dimensions These are standardized time-series samples;

[0036] The sparsity penalty loss is determined by the following formula. :

[0037] ;

[0038] ;

[0039] ;

[0040] Among them, target sparsity For hidden layer 2 The output of each neuron It is the Sigmoid activation function. For hidden layer 2 The average activation probability of each neuron, the hidden layer 2 is formed in the PC-SAE network structure, and is used to extract core fault features by applying sparse constraints;

[0041] The constrained generated samples satisfy the power physics formula:

[0042] ;

[0043] Where 1.05 is the line loss coefficient;

[0044] The physical loss is determined by the following formula. :

[0045] ;

[0046] in, These are the reconstructed values ​​of power, voltage, and current output from PC-SAE, respectively.

[0047] Optionally, temporal features of multi-scale TCN can be extracted using the following formula:

[0048] ;

[0049] in, , This is a 1D convolution operation with a kernel size of 3. For the expansion coefficient, select... To capture features with different time granularities, such as 1ms current spikes, 5ms voltage fluctuations, and 20ms power changes; For bias terms, The activation function is used to enhance the model's nonlinear fitting ability; multi-scale temporal features. ;

[0050] The formula for calculating attention weights is as follows:

[0051] ;

[0052] ;

[0053] in, For attention vectors, This is the weight matrix. For nodes The set of neighboring nodes, For activation functions;

[0054] The formula for outputting spatial features is as follows:

[0055] ;

[0056] in, It is the Sigmoid activation function. ;

[0057] Among them, time series features Spatial features The weighted fusion formula is:

[0058] ;

[0059] In this context, temporal features are used for queries, and spatial features are used for keys and values. ,

[0060] , Using spatial features as the query and temporal features as the key and value, i.e. , , ; The key dimension is half the feature dimension. For feature splicing operations; To output fused features.

[0061] Optionally, a dual-constraint anomaly detection and fault tracing decision-making model can be established through the following steps:

[0062] Establish a fully connected network to determine the probability of anomalies at each time step;

[0063] A dual-constraint loss function is constructed that integrates data classification accuracy and conformity to physical laws. By constraining the deviation of physical parameters of abnormal samples, the dual-constraint anomaly detection and fault tracing decision model can identify normal fluctuations.

[0064] A Bayesian dynamic threshold adjustment strategy is established. Through this dynamic adjustment mechanism, a dynamic threshold is determined so that the dual-constraint anomaly detection and fault tracing decision model can adapt to the complex working conditions of the charging station.

[0065] The anomaly probability of each sample at each time step is compared with the dynamic threshold to determine the output sample-level detection result;

[0066] Based on the Shapley additive interpretation model, the contribution of each input feature to the anomaly detection result is determined in order to identify the key features that dominate the anomaly;

[0067] A three-dimensional spatial correlation matrix of the charging pile-grid-energy storage system is constructed based on the standardized preprocessing of the aforementioned data. By combining feature contribution, a fault propagation graph is constructed, and the main propagation path and root cause device of the fault are determined by calculating the fault propagation probability between devices.

[0068] Optionally, the double-constraint training loss function is: ;

[0069] in, These are the loss weighting coefficients;

[0070] The cross-entropy loss is calculated using the following formula. :

[0071] ;

[0072] in, For the sample At any moment The true label is 1 for abnormal and 0 for normal.

[0073] This represents the anomaly probability output by the model.

[0074] The physical constraint loss is determined by the following formula. :

[0075] ;

[0076] in, This is an indicator function that takes a value of 1 when the sample is an anomaly, and 0 otherwise;

[0077] , These are reference values ​​for the rated operating voltage and current of the charging pile. , These are the normalized deviations of voltage and current relative to their rated values, respectively.

[0078] Optionally, the Bayesian dynamic threshold adjustment strategy includes a threshold update formula and an error judgment rule;

[0079] The threshold update formula is as follows:

[0080] ;

[0081] in:

[0082] The initial detection threshold is determined based on the optimal F1 score of the training set; This is the error attenuation coefficient; For the front The cumulative value of the detection error at any given time; For threshold;

[0083] Among them, the detection error index Quantification The formula for the deviation between the time-of-flight detection result and the actual situation is:

[0084] ;

[0085] when When, it indicates the first There is always a possibility of false positives where normal results are mistakenly identified as abnormal, or false negatives where abnormal results are mistakenly identified as normal; when When the time is right, it indicates that the test result is correct.

[0086] Optionally, the contribution of each input feature to the anomaly detection result can be quantified using the following formula:

[0087] ;

[0088] in, For fusion features The first in One characteristic, To remove the first The set of remaining features after each feature, To make the first The fused features are obtained by replacing each feature with a baseline value, where the baseline value is the mean value of the features under normal operating conditions. For expectation operator, Indicates the first The average contribution of each feature to the anomaly probability, a positive value indicates that the feature promotes anomaly detection, a negative value indicates that it inhibits anomaly detection, and the larger the absolute value, the stronger the contribution.

[0089] The probability of fault propagation is calculated using the following formula:

[0090] ;

[0091] in, For the fault from the device Transmitted to device The probability of;

[0092] For equipment The SHAP contribution value corresponding to the core features reflects the device's... The likelihood of a malfunction occurring;

[0093] Spatial Incidence Matrix medium equipment and The correlation coefficient reflects the electrical coupling strength between the two.

[0094] To be compatible with equipment The set of all related devices, based on the spatial association matrix. Confirmed, correlation coefficient equipment ;

[0095] The denominator is all associated device pairs. The sum of the contribution correlation products is used for normalization to ensure... .

[0096] Among them, the path cumulative probability maximization criterion is used to determine the main propagation path of the fault;

[0097] The steps for determining the main propagation path of the fault using the path cumulative probability maximization criterion include:

[0098] Starting with the device corresponding to the top-1 feature of SHAP contribution and ending with the device whose signal first shows an anomaly in anomaly detection, based on the spatial correlation matrix... Generate all possible device connection paths;

[0099] For each candidate path, the product of the propagation probabilities between adjacent devices along the path is calculated as the cumulative propagation probability of that path. ;

[0100] Selecting cumulative propagation probability The longest connection path is taken as the main fault propagation path, and the starting node of the main fault propagation path is the root cause device of the fault.

[0101] The cumulative propagation probability of each path is determined by the following formula. :

[0102] .

[0103] This application proposes a fault detection method for electric vehicle charging stations. Addressing the technical shortcomings of related technologies in multi-source heterogeneous data processing, small-sample fault identification, physical consistency assurance, and fault tracing, this application constructs an innovative framework that deeply integrates physical mechanisms and data-driven approaches. Through multi-source data acquisition and standardized preprocessing, it achieves unified formatting of high-frequency electrical parameters, mid-frequency environmental parameters, and low-frequency equipment status data. Furthermore, it proposes a physically constrained sparse autoencoder network, embedding the power balance law of the charging system into the feature learning process, and combining a sample balancing strategy to enhance small-sample fault features. Subsequently, a multi-scale spatiotemporal attention fusion module accurately captures the temporal evolution and spatial propagation characteristics of faults. Finally, it constructs a dual-constraint anomaly detection and fault tracing decision-making mechanism, jointly optimizing the detection model by combining data classification loss and physical constraint loss, and integrating the Shapley Additive Interpretation Model. By combining the feature contribution of exPlanations (SHAP) with spatial propagation probability, the root cause of faults can be located. This effectively overcomes the problems of low detection rate, high false alarm rate, slow response, and weak source tracing of traditional methods. The proposed method has a high fault detection rate, low false alarm rate, rapid response, and high source tracing accuracy. It has high precision, strong robustness, and physical interpretability, providing reliable technical support for the safe and stable operation of charging stations and significantly reducing operation and maintenance costs and safety risks. Attached Figure Description

[0104] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0105] Figure 1 A flowchart illustrating a fault detection method for an electric vehicle charging station provided in an embodiment of this application;

[0106] Figure 2 A flowchart illustrating another fault detection method for electric vehicle charging stations provided in this application;

[0107] Figure 3 Internal structure diagram of the multi-source data feature enhancement and spatiotemporal fusion model provided in this application;

[0108] Figure 4 The core structural diagram of the dual-constraint anomaly detection and fault tracing decision-making model provided in this application;

[0109] Figure 5 Performance verification comparison chart of the physically constrained sparse autoencoder PC-SAE feature enhancement module provided for this application;

[0110] Figure 6Performance comparison chart of the dual-constraint anomaly detection DC-AD module provided in this application. Detailed Implementation

[0111] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of systems and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0112] The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to be limiting of the disclosure. The singular forms “a,” “the,” and “the” as used in this disclosure and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any and all possible combinations of one or more of the associated listed items.

[0113] It should be understood that although the terms first, second, third, etc., may be used in this disclosure to describe various information, such information should not be limited to these terms. These terms are used only to distinguish information of the same type from one another. For example, without departing from the scope of this disclosure, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0114] This application relates to the field of electric vehicle charging station operation and maintenance technology, and more specifically, to a fault detection method for electric vehicle charging stations.

[0115] With the deepening implementation of the "dual carbon" goals and the explosive growth of the electric vehicle industry, charging stations, as the core infrastructure supporting the popularization of electric vehicles, are experiencing a continuous increase in network coverage density and operational load. According to data from the China Electric Vehicle Charging Infrastructure Promotion Alliance, as of June 2024, the total number of charging infrastructure units nationwide reached 6.3 million, a 45% increase compared to the same period in 2023. Among them, the number of public charging stations exceeded 150,000, with an average of over 12 million charging services per day. The continuous and stable operation of charging stations is not only directly related to users' charging experience and travel safety, but also of significant strategic importance for ensuring the stable operation of urban energy networks and promoting the high-quality development of the new energy vehicle industry.

[0116] Current charging station fault detection technologies mainly fall into two categories: one is based on traditional statistical and threshold methods, such as Principal Component Analysis (PCA) and threshold judgment methods; the other is based on single data-driven models, such as Sparse Autoencoder (SAE) and Multilayer Perceptrons (MLP). Although these technologies have achieved fault detection functions to some extent, they still have significant technical shortcomings when facing the diverse data sources, complex operating conditions, and challenging data characteristics of charging stations, making it difficult to meet actual operation and maintenance needs. Specific problems are as follows:

[0117] First, there are shortcomings in handling the heterogeneity of multi-source data: charging station data encompasses electrical parameters, environmental parameters, and equipment status, making it difficult for traditional methods to uniformly model the spatiotemporal coupling relationships, leading to feature loss. Second, small sample fault features are not significant: the proportion of fault samples at charging stations is low, and fault features are easily submerged in normal fluctuations, making existing models prone to overfitting or missed detections. Third, the lack of physical constraints leads to a high false alarm rate: most data-driven models rely solely on data distribution for modeling, without embedding the physical laws of the charging system, misjudging normal fluctuations as faults, with false alarm rates often exceeding 3%. Fourth, the fault tracing capability is weak: it can only detect anomalies but cannot locate the root cause of the fault, increasing operation and maintenance costs. These technical deficiencies result in existing detection systems generally suffering from three lows and one weakness: low fault detection rate, high false alarm rate, low response speed, and weak tracing capability, making them unable to adapt to the high-density, high-load operation requirements of charging stations.

[0118] Based on this, and considering the core characteristics of charging stations such as heterogeneity of multi-source data, scarcity of fault samples, and regularity of physical operation, a fault detection method is designed that integrates physical constraints and spatiotemporal features, adapts to small sample scenarios, and has both detection accuracy and traceability capabilities. This method addresses the bottlenecks of existing technologies and has become a key technical problem that urgently needs to be solved in the current field of electric vehicle charging station operation and maintenance.

[0119] Based on this, this application proposes a fault detection method for electric vehicle charging stations.

[0120] Please see Figure 1 This application discloses a flowchart illustrating a fault detection method for electric vehicle charging stations. Specifically, the fault detection method for electric vehicle charging stations proposed in this application includes:

[0121] S101, acquires various types of data.

[0122] The various types of data include equipment operating status data, environmental impact data, and anomaly label information.

[0123] S102, perform standardization preprocessing on the various data to obtain processed data.

[0124] S103, input the processed multiple data into a pre-established multi-source data feature enhancement and spatiotemporal fusion model to obtain intermediate data after feature enhancement and spatiotemporal fusion.

[0125] S104, input the intermediate data into the pre-established dual-constraint anomaly detection and fault tracing decision model, determine the main fault propagation path, and determine the root cause device of the fault based on the main fault propagation path.

[0126] S105, determine the operation and maintenance decision-making and handling information based on the pre-established type tag library and operation and maintenance knowledge base, the fault root cause device and the fault main propagation path.

[0127] Specifically, the various types of data undergo standardization preprocessing, including:

[0128] The various types of data are cleaned to obtain cleaned data.

[0129] By combining statistical criteria and physical constraints, outlier data is removed from various cleaned datasets, resulting in a variety of reliable data.

[0130] Spatiotemporal standardization is performed on the various reliable data to obtain the processed data.

[0131] The data cleaning steps for the various types of data include:

[0132] The filling method is performed using time-series weighted interpolation.

[0133] The time is calculated using the following formula. missing values :

[0134] ;

[0135] in, For time decay weight, This is a decay coefficient, ensuring that data from adjacent time points contribute more to missing values; , Missing time points Normal data at adjacent time points.

[0136] The steps for performing spatiotemporal standardization on the various types of trusted data include:

[0137] The various trusted data are divided into segments according to a fixed time window to obtain a variety of trusted data in a unified format for time series data;

[0138] Construct a three-dimensional spatial correlation matrix of charging piles, power grids, and energy storage systems to determine the electrical coupling relationships between the devices.

[0139] The electrical coupling relationship between devices is determined by the following formula:

[0140] ;

[0141] in, For covariance, Standard deviation, Range of values The larger the absolute value, the better the equipment. and The stronger the correlation between the parameters.

[0142] As an example, a comprehensive data acquisition system can be built based on the existing SCADA system of the charging station and the newly added high-frequency sensors. This system collects three core types of data to cover equipment operating status, environmental impact, and fault label information. Specifically, this may include:

[0143] Electrical parameters: Charging pile output voltage Charging current Charging power The sampling frequency was set to 1kHz to capture high-frequency electrical transient characteristics;

[0144] Environmental and equipment status parameters: internal temperature of the charging pile ,humidity Vibration acceleration containing X / Y / Z triaxial components Sampling frequency 10Hz; characterizing equipment operating environment and mechanical condition; circuit breaker status. Charging gun contact status The sampling frequency is 1Hz, and the switching of operating conditions of key equipment is recorded.

[0145] Tag data: Fault types such as overvoltage, overcurrent, poor charging gun contact, abnormal temperature, etc., and timestamps of fault occurrence. This provides a basis for model training and performance verification through annotation.

[0146] Here, this application addresses the problem of random missing data in time series data by employing a time series weighted interpolation method to fill in the missing data at specific times. missing values By assigning higher weights to data from adjacent time points, we ensure that the filled values ​​conform to the temporal evolution of the data.

[0147] This application also combines statistical criteria and physical constraints to achieve precise removal of abnormal noise data. Based on the 3σ criterion, it removes outliers that deviate from the data mean by more than three standard deviations. Furthermore, it sets reasonable ranges for parameters, such as voltage, based on the physical operating limits of the equipment. and current Invalid data outside this range is removed to ensure the authenticity and validity of the data.

[0148] Furthermore, to capture transient characteristics of faults, this application segments the cleaned data into fixed time windows and sets the window length. Step, covering 200ms transient features, with specific time and space dimensions as follows:

[0149] Time dimension: To capture transient characteristics of faults, the cleaned data is divided into fixed time windows, and the window length is set. Step, covering 200ms transient features, to form time series samples. , For the sample size, As a feature dimension, it covers the core indicators of three types of data: electrical, environmental, and equipment status, and realizes the unified formatting of time series data.

[0150] Spatial Dimension: Constructing a three-dimensional spatial relationship matrix of charging piles, power grid, and energy storage systems. Quantify the electrical coupling relationship between devices.

[0151] Specifically, a multi-source data feature enhancement and spatiotemporal fusion model is established through the following steps:

[0152] We designed a physically constrained sparse autoencoder and a PC-SAE loss function that embed the physical laws of the charging system, and added a combination strategy of oversampling and undersampling to output a balanced sample set based on multiple training samples.

[0153] The method employs a fusion of temporal feature extraction using a multi-scale temporal convolutional network (MS-TCN) and spatial feature aggregation using a graph attention network (GAT), combined with cross-modal attention, to output fused features by balancing the spatiotemporal coupling of the captured data of each sample in the sample set.

[0154] Among them, the physical constraint sparse autoencoder of the physical laws of the charging system includes an encoder, a decoder and a physical constraint layer.

[0155] The PC-SAE loss function is:

[0156] ;

[0157] in, , , This is the loss weighting coefficient.

[0158] The reconstruction loss is determined by the following formula. :

[0159] ;

[0160] in, Reconstructed data output from the PC-SAE network. For the sample size, The time window length, For feature dimensions These are standardized time-series samples.

[0161] The sparsity penalty loss is determined by the following formula. :

[0162] ;

[0163] ;

[0164] ;

[0165] Among them, target sparsity For hidden layer 2 The output of each neuron It is the Sigmoid activation function. For hidden layer 2 The average activation probability of each neuron, the hidden layer 2 is formed in the PC-SAE network structure, and is used to extract core fault features by applying sparse constraints.

[0166] The constrained generated samples satisfy the power physics formula:

[0167] ;

[0168] Where 1.05 is the line loss coefficient.

[0169] The physical loss is determined by the following formula. :

[0170] ;

[0171] in, These are the reconstructed values ​​of power, voltage, and current output from PC-SAE, respectively.

[0172] The temporal features of multi-scale TCN can be extracted using the following formula:

[0173] ;

[0174] in, , This is a 1D convolution operation with a kernel size of 3. For the expansion coefficient, select... To capture features with different time granularities, such as 1ms current spikes, 5ms voltage fluctuations, and 20ms power changes; For bias terms, The activation function is used to enhance the model's nonlinear fitting ability; multi-scale temporal features. .

[0175] The formula for calculating attention weights is as follows:

[0176] ;

[0177] ;

[0178] in, For attention vectors, This is the weight matrix. For nodes The set of neighboring nodes, This is the activation function.

[0179] The formula for outputting spatial features is as follows:

[0180] ;

[0181] in, It is the Sigmoid activation function. .

[0182] Among them, time series features Spatial features The weighted fusion formula is:

[0183] ;

[0184] Wherein: time-series features are used for queries, and spatial features are used for keys and values, i.e. , , Using spatial features as the query and temporal features as the key and value, i.e. , , ; The key dimension is half the feature dimension. For feature splicing operations; To output fused features.

[0185] Here, this application addresses the problem of insignificant features caused by the scarcity of small fault samples by designing a Physically Constrained Sparse Autoencoder (PC-SAE) that embeds the physical laws of the charging system, thereby achieving feature enhancement and sample balancing.

[0186] Specifically, the PC-SAE network structure includes an encoder, a decoder, and a physical constraint layer, with the following architecture:

[0187] The encoder adopts a progressive structure of input layer → hidden layer 1 → hidden layer 2. The input layer receives normalized time-series samples. (dimension) Hidden layer 1 contains 128 neurons and uses the ReLU activation function to achieve preliminary feature mapping; hidden layer 2 contains 32 neurons and extracts core fault features by applying sparse constraints.

[0188] Decoder: Employs a structure symmetrical to the encoder, consisting of hidden layer 2 → hidden layer 3 → output layer. Hidden layer 3 contains 128 neurons and uses the ReLU activation function; the output layer has a dimension of... The Sigmoid activation function is used to output the reconstructed data. ;

[0189] Physical constraint layer: Embedding the core power formula of the charging system The coefficient 1.05 represents the line loss correction factor, ensuring that the generated features and reconstructed data conform to physical operating laws and avoiding feature distortion caused by purely data-driven approaches. Specifically, this application employs a combination of oversampling and undersampling strategies to address the imbalance in the distribution of faulty and normal samples.

[0190] Oversampling involves generating physically compliant pseudo-samples from minority class fault samples via the PC-SAE network. The number of generated pseudo-samples is calculated as (number of majority class samples - number of minority class samples) × 0.8, which increases the number of minority class samples to ensure a reasonable proportion of fault samples.

[0191] Undersampling involves using the Near-Miss algorithm on the majority class of normal samples to retain key samples similar to faulty samples, thereby reducing the interference of redundant data from the majority class samples on model training.

[0192] The output is the balanced and enhanced data. PC-SAE Refining Features , This is the number of balanced samples. This application also integrates the temporal feature extraction of Multi-Scale Temporal Convolutional Network (MS-TCN) with the spatial feature aggregation of Graph Attention Network (GAT), and combines cross-modal attention to capture the spatiotemporal coupling relationship of data to construct Multi-Scale Spatio-Temporal Attention Fusion (MST-AF).

[0193] Specifically, the time-series feature extraction of multi-scale TCN uses three TCN branches with different expansion coefficients to cover the evolution of fault features at different time scales, such as 1ms current spikes and 10ms voltage fluctuations, as shown in the following formula:

[0194] ;

[0195] in, , This is a 1D convolution operation, with the kernel size set to 3. For the expansion coefficient, select... It enables the capture of features with different time granularities, such as 1ms current spikes, 5ms voltage fluctuations, and 20ms power changes. For bias terms, The activation function enhances the model's nonlinear fitting ability; the final output is multi-scale temporal features. Each of the three branches has 32 dimensions.

[0196] GAT spatial feature aggregation is based on spatial correlation matrix The GAT quantification method is used to quantify the spatial influence weights between devices, thereby achieving accurate aggregation of spatial features.

[0197] Node feature initialization: The PC-SAE refined features of each charging pile are used as graph node features to construct a node feature matrix. ,in This refers to the number of charging stations;

[0198] This application also employs a dual-head attention mechanism to achieve attention to temporal features. Spatial features The weighted fusion formula is as follows:

[0199] ;

[0200] The first focus of attention is on temporal features as the query and spatial features as the key and value. , , The fusion results highlight the time-series characteristics as the dominant factor.

[0201] The second focus is on using spatial features as the query and temporal features as the key and value. , , The fusion result highlights the spatial characteristics as the dominant factor.

[0202] The key dimension is half the feature dimension. For feature splicing operations;

[0203] Final output fused features This feature encompasses both multi-scale temporal fault information and spatial correlation information between devices, providing comprehensive and highly recognizable feature support for subsequent anomaly detection.

[0204] Specifically, a dual-constraint anomaly detection and fault tracing decision-making model is established through the following steps.

[0205] Establish a fully connected network to determine the probability of anomalies at each time step;

[0206] A dual-constraint loss function is constructed that integrates data classification accuracy and conformity to physical laws. By constraining the physical parameter deviations of abnormal samples, the dual-constraint anomaly detection and fault tracing decision model can identify normal fluctuations.

[0207] A Bayesian dynamic threshold adjustment strategy is established. Through this dynamic adjustment mechanism, a dynamic threshold is determined so that the dual-constraint anomaly detection and fault tracing decision model can adapt to the complex working conditions of the charging station.

[0208] The anomaly probability of each sample at each time step is compared with the dynamic threshold to determine the output sample-level detection result;

[0209] Based on the Shapley additive interpretation model, the contribution of each input feature to the anomaly detection result is determined in order to identify the key features that dominate the anomaly;

[0210] A three-dimensional spatial correlation matrix of the charging pile-grid-energy storage system is constructed based on the standardized preprocessing of the aforementioned data. By combining feature contribution, a fault propagation graph is constructed, and the main propagation path and root cause device of the fault are determined by calculating the fault propagation probability between devices.

[0211] The double-constraint training loss function is:

[0212] ;

[0213] in, This is the loss weighting coefficient.

[0214] The cross-entropy loss is calculated using the following formula. :

[0215] ;

[0216] in, For the sample At any moment The true label is 1 for abnormal and 0 for normal. This represents the anomaly probability output by the model.

[0217] Here, cross-entropy loss The binary cross-entropy loss is used to measure the difference between the model classification results and the true labels, ensuring effective differentiation between abnormal and normal samples.

[0218] The physical constraint loss is determined by the following formula. :

[0219] ;

[0220] in, This is an indicator function that takes a value of 1 when the sample is an anomaly, and 0 otherwise;

[0221] , These are reference values ​​for the rated operating voltage and current of the charging pile. , These are the normalized deviations of voltage and current relative to their rated values, respectively.

[0222] Here, physical constraint loss The system incorporates constraints on the rated operating parameters of the charging system, calculates the loss due to deviations in physical parameters only for abnormal samples, and filters out misjudgments caused by normal fluctuations.

[0223] This application also utilizes the Bayesian probability update concept to dynamically adjust the detection threshold based on historical detection errors. The adaptive optimization Bayesian dynamic threshold adjustment strategy for detection sensitivity includes threshold update formulas and error judgment rules.

[0224] The threshold update formula is as follows:

[0225] ;

[0226] in:

[0227] The initial detection threshold is determined based on the optimal F1 score of the training set; This is the error attenuation coefficient; For the front The cumulative value of the detection error at any given time; The threshold value is used.

[0228] Based on the initial threshold, the threshold is updated exponentially by incorporating the cumulative effect of historical detection errors. Adaptive adjustment of detection sensitivity

[0229] Among them, the detection error index Quantification The formula for the deviation between the time-of-flight detection result and the actual situation is:

[0230] ;

[0231] when When, it indicates the first There is always a possibility of false positives where normal results are mistakenly identified as abnormal, or false negatives where abnormal results are mistakenly identified as normal; when When the time is right, it indicates that the test result is correct.

[0232] As an example, the threshold adjustment logic can be:

[0233] If the historical false positive rate is high, then Automatically lower or raise the detection threshold to reduce false alarms;

[0234] If the historical false negative rate is high, then It automatically raises and lowers the detection threshold to improve sensitivity.

[0235] Through this dynamic adjustment mechanism, the model can adapt to complex operating conditions such as charging station load fluctuations and environmental changes, and maintain stable detection performance.

[0236] This phase includes three parts: feature contribution quantification based on the Shapley Additive Explanations (SHAP) model, fault tracing path modeling, and mapping of tracing results to operation and maintenance decisions. These parts respectively realize the functions of SHAP feature contribution analysis, fault propagation path location, fault source determination, and providing clear handling basis for operation and maintenance personnel.

[0237] Among them, the feature contribution metric based on SHAP uses SHAP values ​​to interpret the model decision process, quantifies the contribution of each input feature to the anomaly detection result, and identifies the key features that dominate the anomaly.

[0238] Specifically, the contribution of each input feature to the anomaly detection result can be quantified using the following formula:

[0239] ;

[0240] in, For fusion features The first in One characteristic, To remove the first The set of remaining features after each feature, To make the first The fused features are obtained by replacing each feature with a baseline value, where the baseline value is the mean value of the features under normal operating conditions. For expectation operator, Indicates the first The average contribution of each feature to the anomaly probability. A positive value indicates that the feature promotes anomaly detection, while a negative value indicates that it inhibits anomaly detection. The larger the absolute value, the stronger the contribution.

[0241] This method can identify the core features that contribute the most to the anomaly. For example, when the SHAP value of the charging gun contact status feature is the largest, it indicates a contact-related fault, providing a preliminary direction for fault tracing.

[0242] The probability of fault propagation is calculated using the following formula:

[0243] ;

[0244] in, For the fault from the device Transmitted to device The probability of;

[0245] For equipment The SHAP contribution value corresponding to the core features reflects the device's... The likelihood of a malfunction occurring;

[0246] Spatial Incidence Matrix medium equipment and The correlation coefficient reflects the electrical coupling strength between the two.

[0247] To be compatible with equipment The set of all related devices, based on the spatial association matrix. Confirmed, correlation coefficient equipment ;

[0248] The denominator is all associated device pairs. The sum of the contribution correlation products is used for normalization to ensure... .

[0249] Among them, the path cumulative probability maximization criterion is used to determine the main propagation path of the fault;

[0250] The steps for determining the main propagation path of the fault using the path cumulative probability maximization criterion include:

[0251] Starting with the device corresponding to the top-1 feature of SHAP contribution and ending with the device whose signal first shows an anomaly in anomaly detection, based on the spatial correlation matrix... Generate all possible device connection paths;

[0252] For each candidate path, the product of the propagation probabilities between adjacent devices along the path is calculated as the cumulative propagation probability of that path. ;

[0253] Selecting cumulative propagation probability The longest connection path is taken as the main fault propagation path, and the starting node of the main fault propagation path is the root cause device of the fault.

[0254] The cumulative propagation probability of each path is determined by the following formula. :

[0255] ;

[0256] This application is based on a pre-constructed three-dimensional spatial correlation matrix of charging pile-grid-energy storage system. By combining feature contribution, a fault propagation graph is constructed, and the main path and root cause of fault propagation are located by calculating the fault propagation probability between devices.

[0257] Here, this application focuses on the core objectives of accurate anomaly identification, fault root cause localization, and operation and maintenance decision support. Through a two-level progressive architecture of dual-constrained anomaly detection (DC-AD) and fault tracing decision, based on multi-scale spatiotemporal fusion features, it achieves fault analysis with both high detection accuracy and strong interpretability.

[0258] This phase designs a dual-constraint training mechanism of "data classification loss + physical constraint loss" and introduces a Bayesian dynamic threshold adjustment strategy to achieve adaptive optimization of detection sensitivity while ensuring detection accuracy, effectively reducing false alarm rate and false negative rate.

[0259] The core classifier architecture of this application uses a fully connected network (FCN) as the core for anomaly classification, with multi-scale spatiotemporal fusion features as input. The network structure adopts a progressive design of feature compression, nonlinear mapping, and probability output:

[0260] Feature compression layer: The dimensionality of the fused features is reduced by a fully connected layer of 128→64, while retaining the core discriminative information. The GELU activation function is used to enhance the nonlinear expressive power of the model.

[0261] Classification mapping layer: The mapping from features to the classification space is achieved through a 64→32 fully connected layer, and a Dropout layer is introduced to prevent the model from overfitting;

[0262] Probability output layer: A 32→1 fully connected layer combined with a sigmoid activation function outputs the anomaly probability at each time step. ,in Indicates an anomaly. This indicates that everything is normal.

[0263] The dual-constraint training loss in this application constructs a dual-constraint loss function that integrates data classification accuracy and conformity to physical laws, ensuring that the model fits the data distribution while remaining true to the actual operating logic of the charging system.

[0264] This application determines the time-level anomaly probability for each sample when determining the anomaly detection output. With dynamic threshold Compare and output sample-level detection results:

[0265] If there exists any moment satisfy Then determine the sample For abnormal samples, record the time when the abnormality begins. With the maximum anomaly probability ;

[0266] If all moments All meet Then determine the sample These are normal samples, returned to the data collection stage for model iteration and optimization.

[0267] As an example, the process of mapping the source tracing results to operational decisions can be combined with a fault type tag library and an operational knowledge base to transform the source tracing results into operational decision-making information that can directly guide practice. The output includes:

[0268] (1) Core fault information: root cause device, fault type, and fault confidence level;

[0269] (2) Propagation process information: the main propagation path of the fault, and the time when the anomaly occurred at each propagation node;

[0270] (3) Operation and maintenance handling suggestions: Based on the operation and maintenance knowledge base, output targeted handling solutions, estimated handling time and a list of required tools.

[0271] This invention aims to address the technical bottlenecks of existing electric vehicle charging station fault detection methods, such as high false alarm rates, numerous missed alarms, and poor fault tracing capabilities. To address the challenges of multi-physics coupling in charging systems, complex fault mechanisms, heterogeneous data, and strong spatiotemporal correlations, an innovative framework integrating multimodal data fusion and physical knowledge embedding is proposed. The primary objective of this invention is to efficiently align, clean, and standardize multimodal asynchronous and heterogeneous data (electrical, thermal, and temporal data) related to charging equipment operation through a novel multi-source data preprocessing and feature characterization technique. This constructs unified spatiotemporal data suitable for deep model learning, improving data quality and consistency from the source.

[0272] The core objective of this invention is to propose a physically constrained sparse autoencoder feature learning mechanism. During deep feature extraction, this method explicitly incorporates prior physical laws and operational constraints of the charging system, such as circuit topology, power balance, and heat conduction, as regularization terms. This guides the model to learn deep features with clear physical interpretability, thereby enhancing the model's ability to represent normal operating conditions and its sensitivity to weak anomalous signals, effectively suppressing false alarms caused by fluctuations in normal operating conditions.

[0273] A further objective of this invention is to design a multi-scale spatiotemporal attention fusion network. This network can adaptively capture fault evolution patterns and propagation paths at different spatial scales, such as internal components of charging piles, inter-pile relationships, and station-level systems, as well as at different time scales, such as instantaneous, short-term, and long-term. This enables accurate extraction of fault features and efficient fusion of contextual information, thereby addressing the spatiotemporal delay and coupling effects of fault propagation and reducing the false negative rate under complex operating conditions.

[0274] Another important objective of this invention is to establish a dual-constraint driven anomaly detection and fault tracing decision-making method. This method combines data-driven reconstruction error statistical distribution with physical model-driven residual constraints to construct a dynamic adaptive fault threshold, achieving accurate detection of anomalies. Furthermore, by retrospectively analyzing attention weights and feature contributions, it identifies the key time points, spatial locations, and physical quantities that trigger anomalies, providing maintenance personnel with intuitive and actionable fault diagnosis reports, ultimately improving the safety and reliability of charging station operations.

[0275] In some embodiments, the fault detection method of the present invention includes two main models and five core stages. The two main models are: a multi-source data feature enhancement and spatiotemporal fusion model, and a dual-constraint anomaly detection and fault tracing decision model. The five core stages are: multi-source data acquisition and standardization preprocessing, physical-constrained sparse autoencoder (PC-SAE) feature enhancement, multi-scale spatio-temporal attention fusion (MST-AF), dual-constraint anomaly detection (DC-AD), and fault tracing decision. The specific process is shown in the attached figure. Figure 2 As shown.

[0276] Please see Figure 2 , Figure 2With offline construction and online application as the core logic, it fully presents the entire process of electric vehicle charging station fault detection based on multimodal spatiotemporal fusion and physical constraint enhancement. It clearly divides the two major stages of model construction and actual deployment, and intuitively demonstrates the deep integration of physical mechanisms and data-driven characteristics.

[0277] Figure 2 The upper part is the offline model building phase, which is the core implementation link of the technical solution. This phase begins with multi-source data acquisition and preprocessing, outputting high-quality time-series samples and spatial correlation matrices through standardized operations. Subsequently, two core models are built in parallel: one is a multi-source data feature enhancement and spatiotemporal fusion model, which solves the small sample problem through PC-SAE feature enhancement, and then achieves spatiotemporal feature coupling through the MST-AF module; the other is a dual-constraint anomaly detection and fault tracing decision model, which completes anomaly detection and root cause localization based on fused features. The two models are trained collaboratively to minimize both data loss and physical loss, ensuring that the models are both close to the actual data and strictly follow the physical laws of the charging system.

[0278] Figure 2 The second half is the online application phase. The trained model is deployed to the actual operation and maintenance scenario of the charging station, receiving multi-source data in real time from the Supervisory Control and Data Acquisition (SCADA) system and sensors. After rapid preprocessing, this data is input into the model, instantly outputting anomaly detection results, fault root cause location information, and operation and maintenance decision suggestions. If the detection result is normal, the data will be fed back to the offline phase for model iteration and optimization; if an anomaly is detected, the operation and maintenance handling process is directly triggered, forming a closed-loop mechanism of build-application-optimization.

[0279] This diagram comprehensively reveals the technical path of the present invention, highlighting the three core innovations of full-process embedding of physical constraints, spatiotemporal collaborative modeling, and closed-loop feedback optimization. It clearly presents the complete link from data input to operation and maintenance decision-making, providing clear process guidance for engineering implementation.

[0280] For example, please refer to Figure 3 , Figure 3 This paper illustrates the hierarchical and modular architecture design adopted in this application, clearly demonstrating the transformation process from raw data to high-quality fused features. The core of this design embodies the three-level processing logic of data cleaning, feature enhancement, and spatiotemporal fusion.

[0281] The model adopts a vertical architecture of data input layer - core processing layer - feature output layer: The data input layer explicitly receives three types of heterogeneous data: high-frequency electrical parameters, mid-frequency environmental parameters, and low-frequency equipment status, ensuring comprehensive information coverage; The core processing layer is the core of the model, divided into three key modules according to the data processing order, each module embedding physical constraints or advanced algorithms: Multi-source data standardization preprocessing ensures data quality through temporal weight interpolation and physical limit removal, and constructs a 3×3 spatial correlation matrix to quantify the coupling relationship of equipment; The PC-SAE feature enhancement module constructs a three-layer network of encoder-decoder-physical constraint layer, drives training with multi-objective loss function, and achieves sample balance by combining oversampling and undersampling; The MST-AF spatiotemporal fusion module extracts temporal features through multi-scale TCN, aggregates spatial features through GAT, and then achieves cross-modal deep coupling through dual attention heads; The feature output layer finally outputs 128-dimensional fused features, which simultaneously cover temporal fault information and spatial correlation information.

[0282] This structural diagram intuitively presents the core design of the model to address the three major technical challenges of heterogeneity of multi-source data, insignificant features of small samples, and fragmented spatiotemporal correlations. It highlights the synergistic mechanism of physical constraints and data-driven approaches, and provides a detailed structural reference for the engineering implementation of the model.

[0283] For example, please refer to Figure 4 , Figure 4 This paper demonstrates the design concept of parallel coupling + closed-loop decision-making adopted in this application, clearly showing the entire process from fusion feature input to operation and maintenance decision output, and the core reflects the three-level decision logic of anomaly detection - feature attribution - root cause localization.

[0284] The model also follows an architecture of data input layer - core processing layer - decision output layer: the data input layer receives 128-dimensional fused features from the preceding model to ensure high discriminative power of the input features; the core processing layer contains two parallel modules: the dual-constraint anomaly detection module outputs anomaly probability through a fully connected network, and combines a dual-constraint mechanism of cross-entropy loss + physical constraint loss with a Bayesian dynamic threshold to achieve high-precision anomaly determination; the fault tracing decision module quantifies feature contribution through SHAP value, calculates fault propagation probability by combining spatial correlation matrix, and locates the main propagation path and fault root cause; the decision output layer clearly outputs anomaly or normal label, anomaly probability, and fault source information, and designs a node for returning data acquisition, forming a closed-loop optimization with the preceding model.

[0285] This structure diagram highlights the model's core advantages of high detection accuracy, low false alarm rate, and strong source attribution capability. It clearly presents the coupling relationship of key technologies such as dual-constraint loss, dynamic threshold, and SHAP attribution, providing clear structural guidance for model deployment and debugging.

[0286] For example, please refer to Figure 5 , Figure 5 The aim is to intuitively demonstrate the technical advantages of the PC-SAE feature enhancement method proposed in this invention compared with the traditional sparse autoencoder SAE in small sample fault feature extraction and sample balancing. The experimental data is based on mixed operating condition samples of actual charging stations, covering typical fault types such as overvoltage, poor contact, and abnormal temperature, as well as normal operating condition data.

[0287] Figure 5 The image above is a feature visualization scatter plot. The horizontal and vertical axes represent the dimensions of the two core feature classes after dimensionality reduction. Gray dots represent the distribution of normal sample features, and red triangles represent the distribution of fault sample features. In the feature scatter plot of the traditional SAE method, there is a large overlap between fault sample features and normal sample features. Feature points of small fault samples are easily submerged by normal sample feature clusters, making it difficult to form clear distinguishing boundaries. This indicates that it fails to effectively separate fault features in heterogeneous data. In contrast, the feature scatter plot of the PC-SAE method shows a clear clustering separation effect between normal and fault sample features. Even small fault sample features with a very low proportion can form independent clusters and maintain a significant degree of distinction from normal sample features. This proves that by embedding the physical constraints of power balance in the charging system and the sample balancing strategy, PC-SAE can enhance the identification of core fault features and effectively solve the problem of insignificant small fault sample features.

[0288] Figure 5 The figure below is a bar chart comparing the reconstruction errors of fault samples before and after feature enhancement. The horizontal axis represents the sample number, and the vertical axis represents the reconstruction error value. In the reconstruction error bar chart of the traditional SAE method, the error values ​​of fault samples and normal samples overlap, and the reconstruction error of some fault samples is even lower than that of normal samples, making it difficult to effectively distinguish between fault and normal conditions using the error threshold. In contrast, in the reconstruction error bar chart of the PC-SAE method, the reconstruction error of normal samples remains at a low level, while the reconstruction error of fault samples is significantly higher than that of normal samples, and the error distribution is concentrated, with a clear and distinguishable error boundary between normal and fault samples. This comparison verifies the effectiveness of the physical constraint embedding and sample balancing strategy, and PC-SAE can provide highly recognizable core feature support for subsequent spatiotemporal fusion and anomaly detection.

[0289] Please see Figure 6 , Figure 6 By using time-series detection curves, the advantages of the proposed DC-AD method over traditional threshold methods and single data-driven model MLP in terms of anomaly detection accuracy, response timeliness, and false alarm / missed alarm control are verified. The experiment is based on continuous real-time operation data of charging stations, covering real fault events such as overcurrent, abnormal temperature, and poor contact.

[0290] Figure 6The above figure shows the comparison curves of time-series anomaly detection. The horizontal axis represents the time series, and the vertical axis represents the anomaly probability value, where 0 represents normal and 1 represents anomaly. The black solid line represents the sequence of real fault labels, with 1 indicating the time period of the fault occurrence. The gray dashed line represents the detection results of the traditional MLP model, and the black solid line represents the detection results of the DC-AD method. In the detection curve of the traditional MLP model, there are phenomena of missed fault detection and false alarms during normal periods, and the response to abnormal signals has a significant delay, making it difficult to capture transient fault characteristics. In contrast, the detection curve of the DC-AD method is highly consistent with the real fault labels, and all real faults can be accurately detected without missed detection. The false alarm phenomenon is significantly reduced, and the anomaly response delay is greatly shortened. It can effectively capture transient fault characteristics such as instantaneous current spikes and short-term power fluctuations, proving the synergistic effect of the dual-constraint loss function and the Bayesian dynamic threshold adjustment strategy, which significantly improves the real-time performance and accuracy of detection.

[0291] Figure 6 The following chart is a bar chart comparing the core performance indicators of the detection method. The horizontal axis represents the detection method, and the vertical axis represents the performance indicator values. The comparison results show that the traditional threshold method and the single data-driven model have significant shortcomings in terms of fault detection rate, false alarm control, and response speed. In contrast, the DC-AD method achieves a significant improvement in fault detection rate, a substantial reduction in false alarm rate, and a significantly optimized response speed compared to traditional methods. This comparison result intuitively demonstrates that this invention, through the synergistic optimization of data classification loss and physical constraint loss, combined with a dynamic threshold adaptive adjustment mechanism, effectively overcomes the shortcomings of traditional methods such as low detection rate, high false alarm rate, and slow response, and can achieve high-precision anomaly detection under complex working conditions.

[0292] This application proposes a fault detection method for electric vehicle charging stations. Addressing the technical shortcomings of related technologies in multi-source heterogeneous data processing, small-sample fault identification, physical consistency assurance, and fault tracing, this application constructs an innovative framework that deeply integrates physical mechanisms and data-driven approaches. Through multi-source data acquisition and standardized preprocessing, it achieves unified formatting of high-frequency electrical parameters, mid-frequency environmental parameters, and low-frequency equipment status data. Furthermore, it proposes a physically constrained sparse autoencoder network to embed the power balance law of the charging system into the feature learning process, combining it with a sample balancing strategy to enhance small-sample fault features. Finally, through a multi-scale spatiotemporal attention fusion module, it accurately captures the temporal evolution and spatial characteristics of faults. Propagation characteristics; Finally, a dual-constraint anomaly detection and fault tracing decision-making mechanism is constructed. The detection model is optimized by combining data classification loss and physical constraint loss. The root cause of the fault is located by combining the feature contribution of the Shapley Additive Explanations (SHAP) model with the spatial propagation probability. This effectively overcomes the problems of low detection rate, high false alarm rate, slow response and weak tracing of traditional methods. The proposed method has a high fault detection rate, low false alarm rate, fast response and high tracing accuracy. It has high precision, strong robustness and physical interpretability, providing reliable technical support for the safe and stable operation of charging stations and significantly reducing operation and maintenance costs and safety risks.

[0293] Finally, it should be noted that although this specification contains many specific implementation details, these should not be construed as limiting the scope of any invention or the scope of the claims, but rather are primarily used to describe the features of specific embodiments of a particular invention. Certain features described in the various embodiments of this specification may also be implemented in combination in a single embodiment. On the other hand, various features described in a single embodiment may also be implemented separately in various embodiments or in any suitable sub-combination. Furthermore, while features may function in certain combinations as described above and even initially claimed in this way, one or more features from a claimed combination may be removed from that combination in some cases, and a claimed combination may refer to a sub-combination or a variation of a sub-combination.

[0294] Similarly, although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and components in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0295] Thus, specific embodiments of the subject matter have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims may be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings are not necessarily shown in a specific order or sequence to achieve the desired result. In some implementations, multitasking and parallel processing may be advantageous.

[0296] The above description is merely a preferred embodiment of this disclosure and is not intended to limit this disclosure. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A fault detection method for electric vehicle charging stations, characterized in that, The method includes: Acquire various types of data, including equipment operating status data, environmental impact data, and anomaly label information; The various data are standardized and preprocessed to obtain the processed data. The processed data are input into a pre-established multi-source data feature enhancement and spatiotemporal fusion model to obtain intermediate data after feature enhancement and spatiotemporal fusion. The intermediate data is input into a pre-established dual-constraint anomaly detection and fault tracing decision model to determine the main fault propagation path, and the root cause device of the fault is determined based on the main fault propagation path. Based on the pre-established type tag library and operation and maintenance knowledge base, the fault root cause device, and the fault main propagation path, the operation and maintenance decision-making and handling information is determined; The multi-source data feature enhancement and spatiotemporal fusion model is established through the following steps: We designed a physically constrained sparse autoencoder and a PC-SAE loss function that embed the physical laws of the charging system, and added a combination strategy of oversampling and undersampling to output a balanced sample set based on multiple training samples. The system employs temporal feature extraction via a multi-scale temporal convolutional network and spatial feature aggregation via a graph attention network, combined with cross-modal attention, to output fused features based on the spatiotemporal coupling relationship of the captured data of each sample in the balanced sample set. The PC-SAE loss function is: ； in, These are the loss weighting coefficients; The reconstruction loss is determined by the following formula. : ； in, Reconstructed data output from the PC-SAE network. For the sample size, The time window length, For feature dimension, These are standardized time-series samples; The sparsity penalty loss is determined by the following formula. : ； ； ； Among them, target sparsity For hidden layer 2 The output of each neuron It is the Sigmoid activation function. For hidden layer 2 The average activation probability of each neuron, the hidden layer 2 is formed in the PC-SAE network structure, and is used to extract core fault features by applying sparse constraints; The constrained generated samples satisfy the power physics formula: ； Where 1.05 is the line loss coefficient; The physical loss is determined by the following formula. : ； in, These are the reconstructed values ​​of power, voltage, and current output from PC-SAE, respectively.

2. The method according to claim 1, characterized in that, Standardization preprocessing is performed on the various types of data, including: The various types of data are cleaned to obtain cleaned data. By combining statistical criteria and physical constraints, outlier data is removed from various cleaned datasets, resulting in a variety of reliable data. Spatiotemporal standardization is performed on the various reliable data to obtain the processed data.

3. The method according to claim 2, characterized in that, The steps for cleaning the various types of data include: The filling method is performed using time-series weighted interpolation. The time is calculated using the following formula. missing values : ； in, For time decay weight, This is a decay coefficient, ensuring that data from adjacent time points contribute more to missing values. , Missing time points Normal data from consecutive moments; The steps for performing spatiotemporal standardization on the various types of trusted data include: The various trusted data are divided into segments according to a fixed time window to obtain a variety of trusted data in a unified format for time series data; Construct a three-dimensional spatial correlation matrix of charging piles, power grid, and energy storage system to determine the electrical coupling relationships between devices; The electrical coupling relationship between devices is determined by the following formula: ； in, For covariance, Standard deviation Range of values The larger the absolute value, the better the equipment. and The stronger the correlation between the parameters.

4. The method according to claim 1, characterized in that, The physical constraint sparse autoencoder of the charging system physical laws includes an encoder, a decoder and a physical constraint layer.

5. The method according to claim 4, characterized in that, Temporal feature extraction for multi-scale TCN is performed using the following formula: ； in, , This is a 1D convolution operation with a kernel size of 3. For the expansion coefficient, select To capture the time granularity characteristics of 1ms current spikes, 5ms voltage fluctuations, and 20ms power changes; For bias terms, The activation function is used to enhance the model's nonlinear fitting ability; multi-scale temporal features. ; The formula for calculating attention weights is as follows: ； ； in, For attention vectors, This is the weight matrix. For nodes The set of neighboring nodes, For activation functions; The formula for outputting spatial features is as follows: ； in, It is the Sigmoid activation function. ; Among them, time series features Spatial features The weighted fusion formula is: ； In this context, temporal features are used for queries, and spatial features are used for keys and values. , , Using spatial features as the query and temporal features as the key and value, i.e. , , ; The key dimension is half the feature dimension. For feature splicing operations; To output fused features.

6. The method according to claim 1, characterized in that, The following steps are used to establish a dual-constraint anomaly detection and fault tracing decision-making model: Establish a fully connected network to determine the probability of anomalies at each time step; A dual-constraint loss function is constructed that integrates data classification accuracy and conformity to physical laws. By constraining the deviation of physical parameters of abnormal samples, the dual-constraint anomaly detection and fault tracing decision model can identify normal fluctuations. A Bayesian dynamic threshold adjustment strategy is established. Through this adjustment strategy, a dynamic threshold is determined so that the dual-constraint anomaly detection and fault tracing decision model can adapt to the complex working conditions of the charging station. The anomaly probability of each sample at each time step is compared with the dynamic threshold to determine the output sample-level detection result; Based on the Shapley additive interpretation model, the contribution of each input feature to the anomaly detection result is determined in order to identify the key features that dominate the anomaly; A three-dimensional spatial correlation matrix of the charging pile-grid-energy storage system is constructed based on the standardized preprocessing of the aforementioned data. By combining feature contribution, a fault propagation graph is constructed, and the main propagation path and root cause device of the fault are determined by calculating the fault propagation probability between devices.

7. The method according to claim 6, characterized in that, The double-constraint training loss function is: ； in, These are the loss weighting coefficients; The cross-entropy loss is calculated using the following formula. : ； in, For the sample At any moment The true label is 1 for abnormal and 0 for normal. This represents the anomaly probability output by the model. The physical constraint loss is determined by the following formula. : ； in, This is an indicator function that takes a value of 1 when the sample is an anomaly, and 0 otherwise; , These are reference values ​​for the rated operating voltage and current of the charging pile. , These are the normalized deviations of voltage and current relative to their rated values, respectively.

8. The method according to claim 7, characterized in that, The Bayesian dynamic threshold adjustment strategy includes the threshold update formula and error judgment rules; The threshold update formula is as follows: ； in: The initial detection threshold is determined based on the optimal F1 score of the training set; This is the error attenuation coefficient; For the front The cumulative value of the detection error at any given time; For the threshold; Among them, the detection error index Quantification The formula for the deviation between the time-of-flight detection result and the actual situation is: ； when When, it indicates the first There is always a possibility of false positives where normal results are mistakenly identified as abnormal, or false negatives where abnormal results are mistakenly identified as normal; when When the time is right, it indicates that the test result is correct.

9. The method according to claim 8, characterized in that, The contribution of each input feature to the anomaly detection result is quantified using the following formula: ； in, For fusion features The first in One characteristic, To remove the first The set of remaining features after each feature, To make the first The fused features are obtained by replacing each feature with a baseline value, where the baseline value is the mean value of the features under normal operating conditions. For expectation operator, Indicates the first The average contribution of each feature to the anomaly probability, a positive value indicates that the feature promotes anomaly detection, a negative value indicates that it inhibits anomaly detection, and the larger the absolute value, the stronger the contribution. The probability of fault propagation is calculated using the following formula: ； in, For the fault from the device Transmitted to device The probability of; For equipment The SHAP contribution value corresponding to the core features reflects the device's... The likelihood of a malfunction occurring; Spatial Incidence Matrix medium equipment and The correlation coefficient reflects the electrical coupling strength between the two. To be compatible with equipment The set of all related devices, based on the spatial association matrix. Confirmed, correlation coefficient equipment ; The denominator is all associated device pairs. The sum of the contribution correlation products is used for normalization to ensure... ; Among them, the path cumulative probability maximization criterion is used to determine the main propagation path of the fault; The steps for determining the main propagation path of the fault using the path cumulative probability maximization criterion include: Starting with the device corresponding to the top-1 feature of SHAP contribution and ending with the device whose signal first shows an anomaly in anomaly detection, based on the spatial correlation matrix... Generate all possible device connection paths; For each candidate path, the product of the propagation probabilities between adjacent devices along the path is calculated as the cumulative propagation probability of that path. ; Selecting cumulative propagation probability The longest connection path is taken as the main fault propagation path, and the starting node of the main fault propagation path is the root cause device of the fault. The cumulative propagation probability of each path is determined by the following formula. : 。

Images