Charging equipment self-diagnosis method and system based on multi-source heterogeneous data space-time fusion

Abstract

The invention discloses a charging equipment self-diagnosis method and system based on multi-source heterogeneous data space-time fusion in the technical field of charging equipment diagnosis. The method comprises the following steps: performing data preprocessing on an obtained charging pile data set to obtain a processed data set; according to the processed data set, constructing a charging pile real-time topology network based on a dynamic graph convolutional network, and extracting a multi-dimensional spatial feature matrix; according to the processed data set, extracting a multi-dimensional time feature matrix based on a lightweight long-short term memory network; performing feature fusion based on an attention feature fusion algorithm according to the multi-dimensional spatial feature matrix and the multi-dimensional time feature matrix to obtain a spatial-temporal feature matrix; and according to the spatial-temporal characteristic matrix, realizing multi-scale anomaly detection based on a variational auto-encoder algorithm after adversarial training enhancement and sliding window dynamic statistics. According to the method, the blindness and resource consumption of troubleshooting can be effectively reduced, and real-time data support is provided for equipment update decision and maintenance resource allocation.

Description

Technical Field

[0001] This invention relates to the field of charging equipment diagnostic technology, and in particular to a self-diagnostic method and system for charging equipment based on spatiotemporal fusion of multi-source heterogeneous data. Background Technology

[0002] Faced with the challenges of energy consumption and environmental impact in the transportation sector, new energy vehicles have gained widespread popularity due to their advantages in improving energy efficiency and reducing carbon emissions. According to the latest statistics released on January 17, 2025, by the end of 2024, the number of new energy vehicles in China had reached 31.4 million, an increase of 51.49% compared to 2023, showing a rapid growth trend from 2019 to 2024. Against this backdrop of explosive growth, the charging safety of new energy vehicles has become a major concern as a key constraint. The safety of the charging equipment itself is a significant factor affecting the safety of the charging process for new energy vehicles. Every technical aspect of the charging equipment may have potential safety hazards, prompting online monitoring and self-diagnostic technologies for charging facilities, capable of real-time monitoring and self-diagnosis, to gradually become an industry necessity.

[0003] However, traditional online monitoring systems primarily rely on parameters such as current and voltage for safety assessments, neglecting diverse variables such as the number of charging vehicles, ambient temperature, and the aging of charging infrastructure. This leads to biased monitoring data and insufficient risk prediction. Secondly, the widespread distribution of charging facilities, with significant regional variations in their conditions, makes existing online monitoring systems ill-suited for complex operating conditions, limiting their monitoring capabilities. Furthermore, the single data source and the unique characteristics of each charging facility make fault localization prone to errors, still requiring manual verification, which falls short of the anticipated intelligent monitoring goals. Therefore, establishing a technology capable of accurate online monitoring and self-diagnosis for most charging facilities has become a pressing issue in the context of the rapid expansion of new energy vehicles. Summary of the Invention

[0004] The purpose of this invention is to overcome the problems of single data dimension, poor generalization, easy error in fault location and fuzzy fault tracing in the prior art. It provides a self-diagnosis method and system for charging equipment based on spatiotemporal fusion of multi-source heterogeneous data. Through feature screening of abnormal data series and equipment status mapping, it can achieve accurate screening of faulty charging piles, effectively reduce the blindness and resource consumption of fault diagnosis, and provide real-time data support for equipment update decisions and maintenance resource allocation. The accuracy rate of faulty equipment identification reaches 92.3%.

[0005] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution:

[0006] In a first aspect, the present invention provides a self-diagnosis method for charging devices based on spatiotemporal fusion of multi-source heterogeneous data, characterized in that it includes:

[0007] The obtained charging pile dataset is preprocessed to obtain the processed dataset.

[0008] Based on the processed dataset, a real-time topology network for charging piles is constructed using a dynamic graph convolutional network, and a multi-dimensional spatial feature matrix is ​​extracted.

[0009] Based on the processed dataset, a multidimensional temporal feature matrix is ​​extracted using a lightweight long short-term memory network.

[0010] Based on the multidimensional spatial feature matrix and the multidimensional temporal feature matrix, feature fusion is performed using an attention feature fusion algorithm to obtain a spatiotemporal feature matrix.

[0011] Based on the spatiotemporal feature matrix, multi-scale anomaly detection is achieved using an adversarial training-enhanced variational autoencoder and sliding window dynamic statistics.

[0012] Optionally, the charging pile dataset includes charging pile equipment operation data, system log data, environmental monitoring data, and user behavior data;

[0013] The data preprocessing of the acquired charging pile dataset includes:

[0014] The equipment operation data and system log data of the charging pile are denoised by wavelet transform to obtain denoised equipment operation data and system log data.

[0015] The missing values ​​in the environmental monitoring data were imputed using a weighted average method to obtain the environmental monitoring data after missing value processing.

[0016] The user behavior data, environmental monitoring data after missing value processing, noise-reduced device operation data, and system log data are time-aligned using a uniform time interval to obtain the processed dataset.

[0017] Optionally, the noise reduction processing of the charging pile's equipment operation data and system log data using wavelet transform is achieved through the following formula:

[0018]

[0019]

[0020] in, This indicates the equipment operation data of the charging pile. This represents the noise-reduced device operating data. This represents the system log data of the charging pile. This represents the denoised system log data. Indicates a specific time and location. This indicates the determination of the frequency scale. This indicates setting the highest scale for decomposition. Represents a low-frequency wavelet function. Represents a high-frequency wavelet function. Indicates the low-frequency coefficient. Indicates high-frequency coefficients.

[0021] Optionally, based on the processed dataset, a real-time topology network for charging piles is constructed using a dynamic graph convolutional network, and multi-dimensional spatial features are extracted, including:

[0022] Based on the processed dataset, each charging pile is treated as a node, and the dataset corresponding to the charging pile is used as the feature of that node to construct a real-time topology network for the charging piles.

[0023] For any At time 1, perform the following steps to obtain the multidimensional spatial feature matrix. :

[0024] Based on the real-time topology network of the charging piles, the KNN algorithm is used to select the front-end for each node. Find the most similar neighbor nodes and construct an adjacency matrix. ;

[0025] For the adjacency matrix Perform symmetric normalization to obtain the normalized adjacency matrix. ;

[0026] Based on the normalized adjacency matrix and node characteristics Get the features of all neighboring nodes of a node. ;

[0027] Based on the characteristics of all neighboring nodes of the node The features of the nodes are updated using a multilayer perceptron, and the output is... Spatial feature matrix of time nodes .

[0028] Optionally, the similarity between nodes can be calculated using the following formula:

[0029] ,

[0030] in, Represents a node and nodes Node similarity, Indicates the selection of a node The closest 1 node Indicates the use of updating nodes The edge convolution of the features, Indicates the use of updating nodes The edge convolution of the features, Represents a computing node and nodes Similarity;

[0031] The normalized adjacency matrix It can be obtained through the following formula:

[0032] ,

[0033] in, Degree matrix, ;

[0034] Features of all neighbor matrices of the node It can be obtained through the following formula:

[0035] ;

[0036] The Spatial feature matrix of time nodes It can be obtained through the following formula:

[0037] ,

[0038] in, This represents a multilayer perceptron. This represents the PReLU activation function.

[0039] Optionally, the multidimensional temporal feature matrix extracted based on the lightweight long short-term memory network is obtained by the following formula:

[0040]

[0041] in, Indicates the input gate. Indicates candidate memories, Indicates the updated memory. express Lightweight input to long short-term memory networks at all times. express The output of the lightweight long short-term memory network is constantly being optimized. express The output of the lightweight long short-term memory network is constantly being optimized. Indicates to and Perform vector concatenation. This represents the sigmoid activation function. This represents the tanh activation function. This represents the weight matrix of the input gate. The weight matrix representing the candidate memory. This represents the bias value of the input gate. This represents the bias value of the candidate memory. This indicates that element-wise multiplication is performed. This represents a multidimensional time feature matrix.

[0042] Optionally, the feature fusion based on the attention feature fusion algorithm to obtain the spatiotemporal feature matrix includes:

[0043] The multidimensional spatial feature matrix Perform directional projection onto the query space and value space to generate query vectors respectively. Sum value vector ,in, The projection matrix of the query space. The projection matrix representing the value space;

[0044] The multidimensional time feature matrix Perform directional projection onto the bond space and the reserved space to generate bond vectors respectively. and preserved vector ,in, The projection matrix of the bond space is represented. The projection matrix representing the reserved space;

[0045] According to the query vector and key vector The attention score is calculated and normalized to obtain the attention weight matrix. ;

[0046] According to the attention weight matrix Value vector and preserved vector Dual-path feature fusion is performed to obtain the spatiotemporal feature matrix. .

[0047] Optionally, the attention weight matrix It can be obtained through the following formula:

[0048] ,

[0049] in, This indicates the implicit dimensions of memory. This represents the softmax normalization function;

[0050] The spatiotemporal feature matrix The following dual-path fusion formula is used to obtain:

[0051] .

[0052] Optionally, the step of realizing multi-scale anomaly detection based on the spatiotemporal feature matrix, using an adversarial training-enhanced variational autoencoder and sliding window dynamic statistics, includes:

[0053] The spatiotemporal feature matrix is ​​processed using an MLP network. Encode, Mapping to the latent variable space, calculate the mean of the latent variables. and variance ;

[0054] Based on the mean of the latent variables and variance Reparameterized sampling is performed from the latent variable space, and the sampled data is reconstructed using a decoder to obtain the reconstructed spatiotemporal feature matrix. ;

[0055] Through the discriminator Distinguish the reconstructed spatiotemporal feature matrix and the real spatiotemporal feature matrix Furthermore, adversarial training is used to maximize the discriminator's ability to recognize real data, resulting in the objective function. ;

[0056] The spatiotemporal feature matrix is ​​displayed through a sliding window. Divided into a time segment The reconstructed spatiotemporal feature matrix Divided into a time segment By dynamically analyzing data through a sliding window, abnormal data sequences can be obtained. ;

[0057] The mean of the latent variables and variance It can be obtained through the following formula:

[0058] ,

[0059] in, This represents a multilayer perceptron;

[0060] The reconstructed spatiotemporal feature matrix It can be obtained through the following formula:

[0061] ,

[0062] in, This represents the sampled data in the latent variable space. To obtain from the standard normal distribution Data obtained from sampling Represents an exponential function;

[0063] The abnormal data sequence It can be obtained through the following formula:

[0064] ;

[0065] The objective function of the variational autoencoder algorithm based on adversarial training enhancement for;

[0066] ,

[0067] in, Indicates the discriminator, Represents the mathematical expectation. Hyperparameters representing the balance between reconstruction and regularization strength, This indicates the calculation of the prior distribution of the latent variables. This represents the approximate posterior distribution of the real data. This indicates the calculation of the KL divergence value.

[0068] Secondly, the present invention provides a self-diagnostic system for charging devices based on spatiotemporal fusion of multi-source heterogeneous data, applicable to the self-diagnostic method for charging devices based on spatiotemporal fusion of multi-source heterogeneous data as described in any one of the first aspects, comprising:

[0069] The preprocessing module is used to preprocess the acquired charging pile dataset to obtain the processed dataset.

[0070] The spatial feature extraction module is used to: construct a real-time topology network for charging piles based on a dynamic graph convolutional network according to the processed dataset, and extract a multi-dimensional spatial feature matrix.

[0071] The time feature extraction module is used to: extract a multi-dimensional time feature matrix based on a lightweight long short-term memory network according to the processed dataset;

[0072] The feature fusion module is used to: perform feature fusion based on the multidimensional spatial feature matrix and the multidimensional temporal feature matrix using an attention feature fusion algorithm to obtain a spatiotemporal feature matrix;

[0073] The multi-scale anomaly detection module is used to: realize multi-scale anomaly detection based on the spatiotemporal feature matrix, the variational autoencoder enhanced by adversarial training, and the dynamic statistics of the sliding window.

[0074] Compared with existing technologies, the beneficial effects achieved by this invention are as follows:

[0075] 1. It integrates real-time data, historical data, environmental data, and user behavior data, which is conducive to more accurate fault location; it uses DGCN to dynamically capture spatial topology evolution and TinyLSTM to extract temporal features in a lightweight manner, and models dynamic spatiotemporal, which solves the problem of "spatiotemporal feature fragmentation" in complex scenarios and improves the sensitivity of anomaly detection.

[0076] 2. By using the AFF algorithm to dynamically balance spatiotemporal weights, the bottleneck of "feature conflict caused by static weights" is overcome, enhancing the robustness and interpretability of joint representation; by using the AT-VAE algorithm to construct a robust temporal reconstruction model, combined with the dynamic statistical mechanism of the sliding window, the error accumulation bottleneck of traditional single-scale feature analysis is overcome, thereby improving the accuracy of multidimensional anomaly detection.

[0077] 3. By using feature screening of abnormal data series and equipment status mapping, the system can accurately screen faulty charging piles, effectively reducing the blindness and resource consumption of fault diagnosis, and providing real-time data support for equipment update decisions and maintenance resource allocation. The accuracy rate of faulty equipment identification reaches 92.3%. Attached Figure Description

[0078] Figure 1 This is a flowchart of a self-diagnosis method for charging devices based on spatiotemporal fusion of multi-source heterogeneous data according to an embodiment of the present invention;

[0079] Figure 2 The flowchart of the TinyLSTM algorithm provided according to an embodiment of the present invention is shown below;

[0080] Figure 3 A simulation diagram of current and voltage fluctuations of a single charging pile on January 25, 2024, according to an embodiment of the present invention;

[0081] Figure 4 This is a process diagram of dynamic threshold detection of anomalies in the anomaly detection process provided by an embodiment of the present invention. Detailed Implementation

[0082] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0083] It should be noted that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0084] Example 1:

[0085] This invention discloses a self-diagnostic method for charging devices based on spatiotemporal fusion of multi-source heterogeneous data, with reference to... Figure 1 Shown, including:

[0086] S1, perform data preprocessing on the obtained charging pile dataset to obtain the processed dataset;

[0087] S2, Based on the processed dataset, construct a real-time topology network for charging piles using a dynamic graph convolutional network, and extract a multi-dimensional spatial feature matrix;

[0088] S3, Based on the processed dataset, extract a multidimensional temporal feature matrix using a lightweight long short-term memory network;

[0089] S4. Based on the multidimensional spatial feature matrix and the multidimensional temporal feature matrix, feature fusion is performed using an attention feature fusion algorithm to obtain the spatiotemporal feature matrix.

[0090] S5. Based on the spatiotemporal feature matrix, multi-scale anomaly detection is achieved using the variational autoencoder enhanced by adversarial training and the dynamic statistics of the sliding window.

[0091] Specifically, in step S1, the charging pile dataset includes charging pile equipment operation data, system log data, environmental monitoring data, and user behavior data; data preprocessing of the acquired charging pile dataset includes:

[0092] S1.1, Wavelet transform is used to denoise the current, voltage, temperature, humidity, vibration, and other data in the equipment operation data and system log data of the charging pile, resulting in denoised equipment operation data and system log data:

[0093]

[0094]

[0095] in, This indicates the equipment operation data of the charging pile. This represents the noise-reduced device operating data. This represents the system log data of the charging pile. This represents the denoised system log data. Indicates a specific time and location. This indicates the determination of the frequency scale. This indicates setting the highest scale for decomposition. Represents a low-frequency wavelet function. Represents a high-frequency wavelet function. Indicates the low-frequency coefficient. Indicates high-frequency coefficients;

[0096] S1.2, The missing values ​​in the weather data portion of the environmental monitoring data are filled using a weighted average method to obtain the environmental monitoring data after missing value processing;

[0097] S1.3, the above-mentioned noise-reduced equipment operating data Denoising-reduced system log data Environmental monitoring data after missing value processing User behavior data is aligned using a uniform time interval to ultimately form a comprehensive dataset. The processed dataset is obtained.

[0098] In step S2, a real-time topology network for the charging pile is constructed based on a Dynamic Graph Convolution Network (DGCN), and multi-dimensional spatial features are extracted, including:

[0099] S2.1 Based on the processed dataset, each charging pile is treated as a node, and the dataset corresponding to the charging pile is used as the feature of that node to construct a real-time charging pile topology network.

[0100] S2.2, Based on the real-time topology network of the charging piles, the KNN algorithm is used to select the front-end for each node. Find the most similar neighbor nodes and construct an adjacency matrix. That is: for Time, node and nodes The node similarity is calculated as follows:

[0101] ,

[0102] in, Represents a node and nodes Node similarity, Indicates the selection of a node The closest 1 node Indicates the use of updating nodes The edge convolution of the features, Indicates the use of updating nodes The edge convolution of the features, Represents a computing node and nodes Similarity;

[0103] S2.3, To address the issue of different adjacency matrix numbers for different nodes, it is necessary to perform symmetric normalization on the adjacency matrix. Perform symmetric normalization to obtain the normalized adjacency matrix. :

[0104] ,

[0105] in, Degree matrix, ;

[0106] S2.4, based on the normalized adjacency matrix and node characteristics Get the features of all neighboring nodes of a node. :

[0107] ;

[0108] S2.5, based on the characteristics of all neighboring nodes of the node. The features of the nodes are updated using a multilayer perceptron, and the spatial feature matrix of the nodes at the next time step is output. :

[0109] ,

[0110] in, This represents a multilayer perceptron. This represents the PReLU activation function.

[0111] This embodiment, based on the spatiotemporal coupling characteristics of the DGCN algorithm, integrates the data at each time step. Mapped to the network topology, each charging pile is treated as a node, ultimately constructing a spatial feature matrix of the charging pile cluster. .

[0112] In step S3, a lightweight long short-term memory network (TinyLSTM) is constructed to establish a time-series model for each charging station, capturing multi-dimensional temporal features. Its core ideas include: merging the input gate and forget gate into a single merged gate, sharing weight parameters; performing low-rank decomposition on the weight matrix to reduce the number of parameters; and compressing the output feature dimension through pooling operations. The TinyLSTM algorithm aims to maintain the core capabilities of LSTM while reducing computational complexity and the number of parameters. It is constructed by reducing the number of parameters, merging the input and forget gates, and using low-rank decomposition to obtain comprehensive data from all time points. Temporal feature matrix .

[0113] by For example, refer to the time frame for the specific changes. Figure 2 As shown, the original LSTM algorithm determines what new information to add to the original memory cell through the input gate, what old information to forget through the forget gate, and finally what information to output to the hidden state through the output gate. This embodiment combines the input gate and the forget gate, sharing the weight parameters. The original forget gate is... Instead, a low-rank decomposition is performed on the weight matrix to reduce the number of parameters; the output feature dimension is compressed through pooling operations; and so on. Taking a specific time point as an example, the complete calculation process of the TinyLSTM algorithm is as follows:

[0114]

[0115] in, Indicates the input gate. Indicates candidate memories, Indicates the updated memory. express Lightweight input to long short-term memory networks at all times. express The output of the lightweight long short-term memory network is constantly being optimized. express The output of the lightweight long short-term memory network is constantly being optimized. Indicates to and Perform vector concatenation. This represents the sigmoid activation function. This represents the tanh activation function. This represents the weight matrix of the input gate. The weight matrix representing the candidate memory. This represents the bias value of the input gate. This represents the bias value of the candidate memory. This indicates that element-wise multiplication is performed. This represents a multidimensional time feature matrix.

[0116] In step S4, based on the attention feature fusion algorithm (AFF), a spatiotemporal joint representation model of charging piles is constructed to achieve dynamic fusion and enhancement of the spatiotemporal features of charging pile data, resulting in a spatiotemporal feature matrix, including:

[0117] S4.1 uses an attention mechanism to dynamically calculate feature weights and determine the fusion ratio between the two:

[0118] S4.1.1, the multidimensional spatial feature matrix Through a dedicated projection matrix Projected onto the query space, generating a query vector. Learn the topology of the charging pile network; and use the multidimensional time feature matrix. Through a dedicated projection matrix Projected onto the key space, generating a value vector. Extracting dynamic correlation patterns from temporal features;

[0119] S4.1.2, based on the query vector and key vector By calculating attention scores, the correlation strength between spatial nodes and temporal segments is obtained, and softmax normalization is performed to obtain a dynamically adjustable attention weight matrix. :

[0120] ,

[0121] in, This indicates the implicit dimensions of memory. This represents the softmax normalization function;

[0122] S4.2 dynamically fuses the spatial and temporal features at different times:

[0123] S4.2.1, the multidimensional spatial feature matrix Through a dedicated projection matrix Projected onto the value space, generating a value vector. The re-encoded reconstructed spatial features are obtained; the multidimensional temporal feature matrix is ​​then... Through a dedicated projection matrix Projected onto the reserved space, generating a reserved vector. Thus, a time series that resists noise interference and retains the original dynamic fluctuation law is obtained;

[0124] S4.2.2, according to the attention weight matrix Value vector and preserved vector Dual-path feature fusion is performed to obtain the spatiotemporal feature matrix. :

[0125] .

[0126] In step S5, based on the spatiotemporal feature matrix, a robust temporal reconstruction model is constructed using the adversarial training-enhanced variational autoencoder algorithm (AT-VAE), and multi-scale anomaly detection is achieved by combining sliding window dynamic statistics, including:

[0127] S5.1, an MLP network is used as the encoder to process the spatiotemporal feature matrix. The latent variables are compressed into the latent variable space, and the mean of the latent variables is output through a nonlinear transformation by a fully connected layer. and variance Establish probabilistic feature representations:

[0128] ,

[0129] in, This represents a multilayer perceptron;

[0130] S5.2, based on the mean of the latent variables and variance Using reparameterization techniques to sample latent variables A symmetric MLP structure is used as the decoder to reconstruct the sampled data, resulting in the reconstructed spatiotemporal feature matrix. :

[0131] ,

[0132] in, This represents the sampled data in the latent variable space. To obtain from the standard normal distribution Data obtained from sampling Represents an exponential function;

[0133] S5.3, by introducing a discriminator To improve the model's fault tolerance, adversarial training is used to maximize the discriminator's ability to recognize real data. This minimizes the reconstruction error and KL divergence on one hand, and maximizes the reconstructed spatiotemporal feature matrix on the other. and the real spatiotemporal feature matrix Discrimination, hyperparameters By dynamically adjusting the two weights, the objective function can be obtained. :

[0134] ,

[0135] in, Indicates the discriminator, Represents the mathematical expectation. Hyperparameters representing the balance between reconstruction and regularization strength, This indicates the calculation of the prior distribution of the latent variables. This represents the approximate posterior distribution of the real data. This indicates the calculation of the KL divergence value;

[0136] S5.4, using a sliding window to plot the spatiotemporal feature matrix Divided into a time segment The reconstructed spatiotemporal feature matrix Divided into a time segment By dynamically comparing and reconstructing the error distribution segment by segment, an abnormal data sequence that conforms to the extreme value theory is finally obtained. :

[0137] .

[0138] In summary, this embodiment proposes a self-diagnosis method for charging equipment based on spatiotemporal fusion of multi-source heterogeneous data. It integrates real-time data, historical data, environmental data, and user behavior data, facilitating more accurate fault location. By dynamically capturing spatial topology evolution using DGCN and extracting temporal features using TinyLSTM in a lightweight manner, a dynamic spatiotemporal model is created, solving the problem of "spatiotemporal feature fragmentation" in complex scenarios and improving the sensitivity of anomaly detection. Furthermore, the AFF algorithm is used to dynamically balance spatiotemporal weights, overcoming the bottleneck of "feature conflicts caused by static weights" and enhancing the robustness and interpretability of joint representation. The AT-VAE algorithm is used to construct a robust temporal reconstruction model, combined with a sliding window dynamic statistical mechanism, overcoming the error accumulation bottleneck of traditional single-scale feature analysis and achieving improved accuracy in multi-dimensional anomaly detection. Through feature screening of anomaly data series and equipment state mapping, accurate screening of faulty charging piles is achieved, effectively reducing the blindness and resource consumption of fault investigation, providing real-time data support for equipment update decisions and maintenance resource allocation, with a faulty equipment identification accuracy rate of 92.3%.

[0139] Example 2:

[0140] Based on the same inventive concept as Embodiment 1, this embodiment of the invention discloses a self-diagnostic system for charging devices based on spatiotemporal fusion of multi-source heterogeneous data. It is applicable to any self-diagnostic method for charging devices based on spatiotemporal fusion of multi-source heterogeneous data in Embodiment 1, and includes:

[0141] The preprocessing module is used to preprocess the acquired charging pile dataset to obtain the processed dataset.

[0142] The spatial feature extraction module is used to: construct a real-time topology network for charging piles based on a dynamic graph convolutional network according to the processed dataset, and extract a multi-dimensional spatial feature matrix.

[0143] The time feature extraction module is used to: extract a multi-dimensional time feature matrix based on a lightweight long short-term memory network according to the processed dataset;

[0144] The feature fusion module is used to: perform feature fusion based on the multidimensional spatial feature matrix and the multidimensional temporal feature matrix using an attention feature fusion algorithm to obtain a spatiotemporal feature matrix;

[0145] The multi-scale anomaly detection module is used to: realize multi-scale anomaly detection based on the spatiotemporal feature matrix, the variational autoencoder enhanced by adversarial training, and the dynamic statistics of the sliding window.

[0146] In this embodiment, the charging pile dataset uses charging data from 24 locations in a city, including six months of data from December 1, 2023 to May 31, 2024, and includes a total of 97 electric vehicle charging users. The dataset not only includes charging data from charging piles, but also hourly traffic density data, environmental monitoring data, and user behavior data from traffic points near each charging pile.

[0147] For charging data, this embodiment uses wavelet transform for denoising, without handling null values; for the weather data portion of the environmental monitoring data, a wavelet transform is used... The method is used to fill in missing values, where, represent Missing values ​​at time points and represent Time and The value at any given moment.

[0148] The charging pile data only includes the start and end times of each charging session and the total charging energy consumption. In order to unify the time intervals of all datasets, it is necessary to convert the data of each charging session into the charging duration and average power per hour.

[0149] Similarly, environmental monitoring data only includes 26 meteorological elements such as the local daily high and low temperatures and rainfall. It is necessary to summarize the local temperature and temporal fluctuation patterns, and based on the daily high and low temperatures and seasonal conditions, to develop a cosine model to obtain the hourly local temperature data.

[0150] ,

[0151] in, It is the average daily temperature of that day, i.e. , Indicates the current hour. , It is the number of hours with the highest temperature.

[0152] Based on the rated voltage of 230V, average charging efficiency fluctuation range of 60%-88%, temperature, and traffic density around the charging piles in this embodiment, the current and voltage of each charging pile are estimated. Taking January 25, 2024 as an example, the current and voltage fluctuation of a single charging pile is as follows: Figure 3 As shown.

[0153] The preprocessed charging pile charging data, traffic density data, environmental monitoring data, and user behavior data are merged into a comprehensive dataset by row according to the time intervals of year, month, day, and hour. This comprehensive dataset is then divided into a training dataset and a test dataset in an 8:2 ratio.

[0154] In the process of constructing a real-time topology network for charging piles based on the DGCN algorithm, this embodiment compares multiple... The final result of the value selection is to choose the first 5 for each node. Find the most similar neighbors, establish connections, construct an adjacency matrix, and use symmetric normalization to solve the problem of different adjacency matrix numbers for different nodes.

[0155] In this embodiment, the TinyLSTM algorithm is used to reduce computational complexity and the number of parameters while maintaining the core capabilities of LSTM. It combines the input gate and the forget gate and is constructed using a low-rank decomposition method with a rank of 16 to obtain the temporal feature matrix of the comprehensive data at all time points.

[0156] The AFF algorithm dynamically fuses the multidimensional spatial features extracted by the DGCN algorithm and the multidimensional temporal features extracted by the TinyLSTM algorithm. To avoid overfitting and accelerate the convergence speed of the training process, this embodiment uses the Adam optimizer to reduce the overfitting probability. The initial learning rate is 1e-5, and the learning rate is automatically adjusted according to the subsequent training process. A total of 50 training rounds are conducted, and the specific training results are shown in Table 1 below.

[0157] Table 1: Comparison of Training Results of AFF Algorithm

[0158] frequency Total Loss Refactor Loss KL Loss Fighting Loss 0 0.9578 0.2925 0.0087 0.6653 10 0.4233 0.2782 0.3319 0.1451 20 0.2280 0.2259 0.4843 0.0017 30 0.1978 0.1940 0.4869 0.0121 40 0.1947 0.1835 0.3170 0.0103

[0159] This embodiment employs a dynamic threshold strategy to fuse long- and short-term monitoring windows, significantly improving the system's resistance to transient noise interference. The final anomaly detection process is as follows: Figure 4 As shown, based on the reconstruction error distribution characteristics within the sliding window, a dual-threshold judgment system is constructed using the 99th percentile (long-term window) and the 99.5th percentile (short-term window), respectively. When the real-time error exceeds the window threshold, it is marked as an anomaly, with the anomaly point located at the peak of the reconstruction error. Under normal circumstances, the time series fluctuates smoothly; however, when an anomaly occurs, the error amplitude suddenly increases. Window smoothing filtering technology effectively distinguishes between steady-state fluctuations and anomalous pulse signals.

[0160] In this embodiment, data without anomalies are used as positive samples and abnormal data are used as negative samples. The final detection results are shown in Table 2 below.

[0161] Table 2: Anomaly Detection Results

[0162] Real / Predicted Positive Class negative class Positive Class 1827 210 negative class 60 1408

[0163] The AFF algorithm proposed in this embodiment shows significant advantages in key indicators: the detection accuracy reaches 92.3% and the recall rate reaches 89.7%, indicating that the AFF algorithm can effectively capture key abnormal events. At the same time, the false alarm rate is controlled at 4.1% and the F1 score is 93.1%, which verifies the balance of detection performance. It achieves the best balance between the cost of misjudgment and the risk of missed detection while ensuring high detection accuracy.

[0164] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.

[0165] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0166] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0167] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1The function specified in one or more boxes.

[0168] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0169] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A self-diagnostic method for charging devices based on spatiotemporal fusion of multi-source heterogeneous data, characterized in that, include: The obtained charging pile dataset is preprocessed to obtain the processed dataset. Based on the processed dataset, a real-time topology network for charging piles is constructed using a dynamic graph convolutional network, and a multi-dimensional spatial feature matrix is ​​extracted. Based on the processed dataset, a multidimensional temporal feature matrix is ​​extracted using a lightweight long short-term memory network. Based on the multidimensional spatial feature matrix and the multidimensional temporal feature matrix, feature fusion is performed using an attention feature fusion algorithm to obtain a spatiotemporal feature matrix. Based on the spatiotemporal feature matrix, multi-scale anomaly detection is achieved using an adversarial training-enhanced variational autoencoder and sliding window dynamic statistics.

2. The self-diagnosis method for charging equipment based on spatiotemporal fusion of multi-source heterogeneous data according to claim 1, characterized in that, The charging pile dataset includes charging pile equipment operation data, system log data, environmental monitoring data, and user behavior data; The data preprocessing of the acquired charging pile dataset includes: The equipment operation data and system log data of the charging pile are denoised by wavelet transform to obtain denoised equipment operation data and system log data. The missing values ​​in the environmental monitoring data were imputed using a weighted average method to obtain the environmental monitoring data after missing value processing. The user behavior data, environmental monitoring data after missing value processing, noise-reduced device operation data, and system log data are time-aligned using a uniform time interval to obtain the processed dataset.

3. The self-diagnosis method for charging equipment based on spatiotemporal fusion of multi-source heterogeneous data according to claim 2, characterized in that, The noise reduction processing of the charging pile's equipment operation data and system log data using wavelet transform is achieved through the following formula: in, This indicates the equipment operation data of the charging pile. This represents the noise-reduced device operating data. This represents the system log data of the charging pile. This represents the denoised system log data. Indicates a specific time and location. This indicates the determination of the frequency scale. This indicates setting the highest scale for decomposition. Represents a low-frequency wavelet function. Represents a high-frequency wavelet function. Indicates the low-frequency coefficient. Indicates high-frequency coefficients.

4. The self-diagnosis method for charging equipment based on spatiotemporal fusion of multi-source heterogeneous data according to claim 1, characterized in that, Based on the processed dataset, a real-time topology network for charging piles is constructed using a dynamic graph convolutional network, and multi-dimensional spatial features are extracted, including: Based on the processed dataset, each charging pile is treated as a node, and the dataset corresponding to the charging pile is used as the feature of that node to construct a real-time topology network for the charging piles. For any At time 1, perform the following steps to obtain the multidimensional spatial feature matrix. : Based on the real-time topology network of the charging piles, the KNN algorithm is used to select the front-end for each node. Find the most similar neighbor nodes and construct an adjacency matrix. ; For the adjacency matrix Perform symmetric normalization to obtain the normalized adjacency matrix. ; Based on the normalized adjacency matrix and node characteristics Get the features of all neighboring nodes of a node. ; Based on the characteristics of all neighboring nodes of the node The features of the nodes are updated using a multilayer perceptron, and the output is... Spatial feature matrix of time nodes .

5. The self-diagnosis method for charging equipment based on spatiotemporal fusion of multi-source heterogeneous data according to claim 4, characterized in that, The similarity between nodes is calculated using the following formula: ， in, Represents a node and nodes Node similarity, Indicates the selection of a node The closest 1 node Indicates the use of updating nodes The edge convolution of the features, Indicates the use of updating nodes The edge convolution of the features, Represents a computing node and nodes Similarity; The normalized adjacency matrix It can be obtained through the following formula: ， in, Degree matrix, ; Features of all neighbor matrices of the node It can be obtained through the following formula: ； The Spatial feature matrix of time nodes It can be obtained through the following formula: ， in, This represents a multilayer perceptron. This represents the PReLU activation function.

6. The self-diagnosis method for charging equipment based on spatiotemporal fusion of multi-source heterogeneous data according to claim 1, characterized in that, The multidimensional temporal feature matrix extracted based on a lightweight long short-term memory network is obtained by the following formula: in, Indicates the input gate. Indicates candidate memories, Indicates the updated memory. express Lightweight input to long short-term memory networks at all times. express The output of the lightweight long short-term memory network is constantly being optimized. express The output of the lightweight long short-term memory network is constantly being optimized. Indicates to and Perform vector concatenation. This represents the sigmoid activation function. This represents the tanh activation function. This represents the weight matrix of the input gate. The weight matrix representing the candidate memory. This represents the bias value of the input gate. This represents the bias value of the candidate memory. This indicates that element-wise multiplication is performed. This represents a multidimensional time feature matrix.

7. The self-diagnosis method for charging equipment based on spatiotemporal fusion of multi-source heterogeneous data according to claim 1, characterized in that, The attention-based feature fusion algorithm is used to perform feature fusion to obtain a spatiotemporal feature matrix, including: The multidimensional spatial feature matrix Perform directional projection onto the query space and value space to generate query vectors respectively. Sum value vector ,in, The projection matrix of the query space. The projection matrix representing the value space; The multidimensional time feature matrix Perform directional projection onto the bond space and the reserved space to generate bond vectors respectively. and preserved vector ,in, The projection matrix of the bond space is represented. The projection matrix representing the reserved space; According to the query vector and key vector The attention score is calculated and normalized to obtain the attention weight matrix. ; According to the attention weight matrix Value vector and preserved vector Dual-path feature fusion is performed to obtain the spatiotemporal feature matrix. .

8. The self-diagnosis method for charging equipment based on spatiotemporal fusion of multi-source heterogeneous data according to claim 7, characterized in that, The attention weight matrix It can be obtained through the following formula: ， in, This indicates the implicit dimensions of memory. This represents the softmax normalization function; The spatiotemporal feature matrix The following dual-path fusion formula is used to obtain: 。 9. The self-diagnosis method for charging equipment based on spatiotemporal fusion of multi-source heterogeneous data according to claim 1, characterized in that, The step of achieving multi-scale anomaly detection based on the spatiotemporal feature matrix, using an adversarial training-enhanced variational autoencoder and sliding window dynamic statistics, includes: The spatiotemporal feature matrix is ​​processed using an MLP network. Encode, Mapping to the latent variable space, calculate the mean of the latent variables. and variance ; Based on the mean of the latent variables and variance Reparameterized sampling is performed from the latent variable space, and the sampled data is reconstructed using a decoder to obtain the reconstructed spatiotemporal feature matrix. ; Through the discriminator Distinguish the reconstructed spatiotemporal feature matrix and the real spatiotemporal feature matrix Furthermore, adversarial training is used to maximize the discriminator's ability to recognize real data, resulting in the objective function. ; The spatiotemporal feature matrix is ​​displayed through a sliding window. Divided into a time segment The reconstructed spatiotemporal feature matrix Divided into a time segment By dynamically analyzing data through a sliding window, abnormal data sequences can be obtained. ; The mean of the latent variables and variance It can be obtained through the following formula: ， in, Represents a multilayer perceptron; The reconstructed spatiotemporal feature matrix It can be obtained through the following formula: ， in, This represents the sampled data in the latent variable space. To obtain from the standard normal distribution Data obtained from sampling Represents an exponential function; The abnormal data sequence It can be obtained through the following formula: ； The objective function of the variational autoencoder algorithm based on adversarial training enhancement for; ， in, Indicates the discriminator, Represents the mathematical expectation. Hyperparameters representing the balance between reconstruction and regularization strength, This indicates the calculation of the prior distribution of the latent variables. This represents the approximate posterior distribution of the real data. This indicates the calculation of the KL divergence value.

10. A self-diagnostic system for charging equipment based on spatiotemporal fusion of multi-source heterogeneous data, characterized in that, include: The preprocessing module is used to preprocess the acquired charging pile dataset to obtain the processed dataset. The spatial feature extraction module is used to: construct a real-time topology network for charging piles based on a dynamic graph convolutional network according to the processed dataset, and extract a multi-dimensional spatial feature matrix. The time feature extraction module is used to: extract a multi-dimensional time feature matrix based on a lightweight long short-term memory network according to the processed dataset; The feature fusion module is used to: perform feature fusion based on the multidimensional spatial feature matrix and the multidimensional temporal feature matrix using an attention feature fusion algorithm to obtain a spatiotemporal feature matrix; The multi-scale anomaly detection module is used to: realize multi-scale anomaly detection based on the spatiotemporal feature matrix, the variational autoencoder enhanced by adversarial training, and the dynamic statistics of the sliding window.

Images