Charging pile data transmission supervision method and system based on data analysis

Abstract

The invention discloses a charging pile data transmission supervision method and system based on data analysis, and the method comprises the steps: segmenting multi-dimensional time series data collected by a charging pile into subsequences, constructing a feature density map and a feature association map, and forming the representation data of a heterogeneous density-association space; the representation data are input into an improved Transform model, a local manifold curvature is calculated to update a self-attention weight, and a heterogeneous space curvature anomaly score is obtained; estimating a generalized Pareto distribution parameter in real time based on the abnormal score, and dynamically calculating a dynamic threshold value of the abnormal score; and comparing the abnormal score with a dynamic threshold value in real time to determine an abnormal level, and automatically triggering a supervision response rule through a Kafka message queue to realize a real-time supervision closed loop of abnormal data. According to the method, the real-time performance and accuracy of abnormal data identification are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data supervision, and in particular to a method and system for supervising charging pile data transmission based on data analysis. Background Art

[0002] With the rapid development of the electric vehicle industry, charging piles have been widely deployed and used as important supporting infrastructure. Charging pile data transmission and monitoring management systems have gradually become a key component of electric vehicle charging operations. Currently, charging pile data monitoring systems primarily collect multi-dimensional operating status data such as voltage, current, power, and temperature during operation. This allows for real-time assessment of the operating conditions of charging facilities, identification of abnormal conditions, and early warning of any such abnormalities, ensuring the safety and reliability of the charging process.

[0003] Current mainstream technologies for monitoring charging pile data transmission primarily utilize threshold-based early warning, time-series anomaly detection, and traditional statistical analysis. Threshold-based early warning methods typically rely on fixed thresholds, triggering an anomaly alarm when collected data exceeds a preset value. Time-series anomaly detection utilizes historical data for trend modeling, identifying data anomalies based on prediction errors. Traditional statistical analysis uses simple statistics, such as mean and variance, to measure whether charging pile operating conditions deviate from normal ranges. These methods are relatively simple to implement and deploy, and have already found application in real-world operational scenarios.

[0004] However, the above existing technologies still have obvious deficiencies in actual operation: (1) The fixed threshold warning method lacks flexibility and is difficult to guarantee accuracy in complex environments, especially incapable of adapting to dynamically changing abnormal data characteristics; (2) Traditional time series anomaly detection methods generally assume that the data presents a clear linear trend, but actual charging pile data exhibits complex characteristics such as nonlinearity, dynamic fluctuations, and multi-dimensional coupling, resulting in low prediction accuracy and anomaly detection accuracy; (3) Existing statistical analysis methods ignore the correlation and dynamic change characteristics between the multidimensional data features of the charging pile data itself, making it impossible to effectively identify local or weak anomalies; (4) Most current regulatory technologies use a passive early warning approach, which has insufficient response speed and cannot meet the needs of rapid disposal of abnormal conditions, making it difficult to form timely and accurate regulatory linkages.

[0005] Therefore, how to provide a charging pile data transmission supervision method and system based on data analysis is a problem that technical personnel in this field urgently need to solve. Summary of the Invention

[0006] One purpose of the present invention is to propose a charging pile data transmission supervision method and system based on data analysis. In response to the problems of insufficient accuracy, slow response speed and lack of effective supervision linkage mechanism in the existing technology in identifying abnormal data of charging piles, a real-time supervision scheme based on heterogeneous density-correlation space data representation, improved Transformer curvature anomaly score calculation and generalized Pareto dynamic threshold judgment is proposed; the present invention has the technical effects of high anomaly identification accuracy, fast response speed and the ability to achieve supervision linkage.

[0007] A charging pile data transmission supervision method based on data analysis according to an embodiment of the present invention includes: The multi-dimensional time series data collected by charging piles is divided into several subsequences, and a feature density map and a feature correlation map are constructed for each subsequence to obtain representation data of the heterogeneous density-correlation space. The representation data of the heterogeneous density-correlation space is input into the improved Transformer model to obtain the curvature anomaly score of the heterogeneous space; The curvature anomaly score in heterogeneous space is used to estimate the generalized Pareto distribution parameters, and the dynamic threshold of the curvature anomaly score is dynamically calculated based on the generalized Pareto distribution parameters. Compare the curvature anomaly score with the dynamic threshold and determine the charging pile data anomaly level based on the comparison result; Automatically trigger preset regulatory response rules based on the abnormal level of charging pile data through the Kafka real-time message queue; Execute regulatory linkage actions in real time through automatically triggered regulatory response rules.

[0008] Optionally, the characterization data for constructing the heterogeneous density-correlation space is specifically: Split multidimensional time series data into continuous subsequences of equal length using a set time window length. Obtain the feature density value of each data point in the corresponding dimension in a point-by-point calculation manner; After calculating the feature density values one by one, a feature density graph is constructed with each data point as a node and the feature density value corresponding to the data point as the node weight. For any two different data dimensions in each subsequence, the feature correlation strength between the data dimensions is calculated pair by pair, and the edge weight value of the feature correlation graph is determined based on the change between the feature correlation strength in the current time window and the previous time window; Construct a feature association graph with each data dimension as a node and the change in the association strength between nodes as the weight of the edge; Update the feature density values of all nodes in the feature density graph and the weights of all edges in the feature association graph to form a real-time updated feature density graph and feature association graph; Directly concatenate the feature density values of the feature density graph nodes and the associated changes of the feature association graph edges to form representation data of the heterogeneous density-association space.

[0009] Optionally, the curvature anomaly score of the heterogeneous space is obtained by using the improved Transformer model, specifically: Extract the eigenvectors from the data representing the heterogeneous density-correlation space one by one, and calculate the local manifold curvature value between any two data dimensions based on the node density value of the eigendensity graph; The local manifold curvature value is used as a constraint factor in the multi-head self-attention calculation process of the improved Transformer model and integrated into the weight calculation process of the self-attention mechanism; The self-attention weights corresponding to each attention head in the multi-head self-attention calculation process are updated through the local manifold curvature value to obtain the real-time self-attention weight matrix; Multiply the real-time self-attention weight matrix with the corresponding value matrix one by one to obtain the attention output vector after each attention head constraint. Then, concatenate all the attention head output vectors in sequence to obtain the multi-head self-attention constraint output vector. The multi-head self-attention constraint output vector is reduced in dimensionality by linear projection, and the linearly projected vector is added to the initial input feature vector one by one through the residual connection operation to obtain the intermediate residual representation vector of the feature vector; After layer-by-layer normalization of the intermediate residual representation vectors of the feature vectors, the abstract representation vectors are obtained through a feedforward neural network. After the abstract representation vector undergoes a second residual connection and layer normalization, it is input into the final fully connected output layer, which outputs the curvature anomaly score corresponding to each feature vector.

[0010] Optionally, the improved Transformer model includes an initial feature linear mapping module, an adaptive manifold geometric mapping module, a multi-manifold geometric perception self-attention module, a residual connection and projection module, and an anomaly score output module: The initial feature linear mapping module is used to receive representation data of the heterogeneous density-correlation space and map the input features into an initial query matrix, an initial key matrix and an initial value matrix respectively; The adaptive manifold geometry mapping module is used to receive the initial query matrix and the initial key matrix, and obtain the query matrix and the key matrix after manifold geometry perception; The multi-head manifold geometry perception self-attention module is used to query the matrix, the key matrix, and the initial value matrix, use the local manifold curvature matrix as the attention mechanism constraint factor, update the attention weight matrix of the multi-head self-attention mechanism in real time, and obtain a complete multi-head self-attention constraint output vector; The residual connection and projection module is used to receive the complete multi-head self-attention constrained output vector, reduce the dimension of the multi-head self-attention constrained output vector to a space with the same dimension as the initial input feature vector through a linear projection operation, and obtain the intermediate residual representation vector of the feature vector; The anomaly score output module is used to receive the intermediate residual representation vector of the feature vector, and after layer normalization processing and feedforward neural network operation, it inputs the abstract representation vector into the fully connected output layer through a second residual connection and layer normalization processing, and outputs the curvature anomaly score corresponding to each feature vector in real time.

[0011] Optionally, the curvature anomaly score is compared with a dynamic threshold in real time, and the charging pile data anomaly level is determined based on the comparison result, specifically: Receive the curvature anomaly score values corresponding to all feature vectors in real time in each time window, and obtain the curvature anomaly score dynamic threshold calculated in the current time window in real time; Calculating the difference between the curvature anomaly score value of each eigenvector and the dynamic threshold value one by one; Based on the calculated real-time numerical difference corresponding to each characteristic vector, the abnormal level of charging pile data is graded and determined one by one; After the classification determination operation is completed, the specific number of charging pile data determined to be severe abnormality level, medium abnormality level and slight abnormality level in the current time window is counted, and the characteristic vector identifier, curvature abnormality score value and difference calculation result corresponding to each abnormality level are recorded one by one; Generate a detailed record table of abnormal levels based on the statistical and recorded results; The generated anomaly level detailed record table is stored in the real-time database in real time as the input data source for the next continuous time window real-time anomaly level determination step.

[0012] Optionally, the preset supervisory response rules are automatically triggered through the Kafka real-time message queue based on the abnormal level of the charging pile data, specifically: Perform the abnormal level identification judgment operation on the abnormal data records in the abnormal level detailed record table one by one: When the abnormality level is marked as "serious", the severity level message topic in the pre-configured Kafka message queue in the system is pushed to the operation and maintenance personnel's mobile terminal in a standard JSON message structure; Operation and maintenance personnel automatically generate emergency dispatch work orders based on pre-stored abnormal dispatch templates and send them to the emergency response queue of the operation and maintenance dispatch platform; When the abnormality level is marked as "medium", it is pushed to the abnormality warning report generation module through the medium-level message topic in the pre-configured Kafka message queue in the system; After receiving and parsing Kafka messages in real time, the abnormal warning report generation module automatically generates a medium-level abnormal warning report in real time based on the preset abnormal warning report template and stores it in the abnormal warning report database; When the exception level is marked as "minor", it is pushed to the minor exception continuous tracking module in a standard JSON message structure through the minor level message topic in the pre-configured Kafka message queue in the system; After the minor anomaly continuous tracking module receives and parses Kafka messages in real time, it updates the minor anomaly real-time records of the corresponding devices in the pre-configured minor anomaly tracking database; The minor anomaly continuous tracking module reads the updated minor anomaly tracking database record within the next continuous time window, and continuously tracks the continuous changes of the minor anomaly status of the current device in real time to form a continuous tracking record of the abnormal status; A real-time dynamic cycle process that automatically triggers and implements corresponding regulatory response rules based on the abnormal level of charging pile data.

[0013] Optionally, a charging pile data transmission supervision system based on data analysis includes: The data acquisition module is used to collect multi-dimensional time series data in real time and divide it into several subsequences of the same length according to the set time window length; The heterogeneous density-correlation space data representation module is used to calculate the feature density value of the feature density graph node and the change in the correlation strength of the feature correlation graph edge for each subsequence point by point, and construct the representation data of the heterogeneous density-correlation space; An improved Transformer anomaly score calculation module is used to receive the representation data of the heterogeneous density-correlation space, and use the local manifold curvature value as a constraint factor to update the self-attention weight of the Transformer model in real time, and output the heterogeneous space curvature anomaly score corresponding to each feature vector; A dynamic threshold calculation module is used to receive curvature anomaly score data in real time and calculate the dynamic threshold of curvature anomaly score based on the generalized Pareto distribution log-likelihood function; The abnormality level determination module is used to receive the curvature abnormality score and dynamic threshold in real time, calculate the numerical difference between the curvature abnormality score and the dynamic threshold one by one, and determine the abnormality level of each charging pile device in real time based on the numerical difference, and generate a detailed record table of the abnormality level; The supervision response linkage module is used to receive the detailed record table of the abnormal level in real time, and automatically trigger the corresponding supervision response rules through the Kafka real-time message queue according to the abnormal level identifier.

[0014] The beneficial effects of the present invention are: (1) The present invention realizes the real-time accurate recognition of abnormal charging pile data by constructing a heterogeneous density-correlation space data representation and integrating an improved Transformer model with local manifold curvature constraints, effectively improving the accuracy and reliability of abnormal charging pile data recognition, and significantly improving the limitation of existing technologies that cannot effectively identify local or weak anomalies.

[0015] (2) The present invention realizes rapid response and accurate judgment of abnormal data of charging piles through real-time dynamic threshold calculation and real-time judgment mechanism of abnormal level based on generalized Pareto distribution parameters, significantly improves the real-time performance and accuracy of abnormal state judgment, and shows better adaptability in complex and dynamically changing charging operation scenarios.

[0016] (3) In terms of automatic supervision and linkage of abnormal data of charging piles, the present invention effectively solves the shortcomings of passive supervision response and poor linkage in the existing technology through an automatic supervision response mechanism based on Kafka real-time message queue, breaks through the bottleneck of lack of real-time and proactive control of the supervision system, realizes the initiative and self-organization of the abnormal supervision process, and effectively improves the actual application effect in the field of charging pile data transmission supervision. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings: Figure 1 This is a flow chart of a charging pile data transmission supervision method based on data analysis proposed by the present invention. DETAILED DESCRIPTION

[0018] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0019] refer to Figure 1 , a charging pile data transmission supervision method based on data analysis, comprising: The multidimensional time series data collected by charging piles is divided into several subsequences. A feature density graph and a feature association graph are constructed for each subsequence. The density values of nodes in the feature density graph and the weights of edges in the feature association graph are calculated and updated in real time to obtain representation data in a heterogeneous density-association space. The representation data of the heterogeneous density-correlation space is input into the improved Transformer model, and the local manifold curvature of each feature in the representation data is calculated in real time. The local manifold curvature is used to calculate and update the self-attention weight of the model to obtain the curvature anomaly score of the heterogeneous space; Based on the curvature anomaly score of heterogeneous space, the generalized Pareto distribution parameters are estimated in real time, and the dynamic threshold of the curvature anomaly score is dynamically calculated according to the real-time estimated generalized Pareto distribution parameters; Comparing the curvature anomaly score with the dynamic threshold in real time, and determining the charging pile data anomaly level based on the comparison result; Based on the abnormal level of charging pile data, the preset regulatory response rules are automatically triggered through the Kafka real-time message queue. The response rules include automatically pushing the abnormality severity level to the operation and maintenance personnel terminal in real time and starting the automatic scheduling program, automatically generating an early warning report for the abnormality medium level and recording it in the database, and automatically recording and continuously tracking the abnormality minor level. The above-mentioned automatically triggered regulatory response rules are used to execute regulatory linkage operations in real time, forming a real-time regulatory closed-loop system for charging pile data transmission.

[0020] In this embodiment, the multidimensional time series data collected by the charging pile is divided into several subsequences, and a feature density graph and a feature association graph are constructed for each subsequence. The density value of the node in the feature density graph and the weight of the edge in the feature association graph are calculated and updated to obtain the representation data of the heterogeneous density-association space. Specifically, The multi-dimensional time series data of voltage, current, power, and temperature collected by the charging pile in real time are divided into several continuous subsequences of the same length based on the set time window length; For each data dimension in each subsequence, the feature density value of each data point in the corresponding dimension is obtained by point-by-point calculation: ; in, For the The characteristic density value of the data point, is the feature vector of the current data point, is the feature vector of other data points in the same subsequence, is the number of data points in the subsequence, is a fixed density radius parameter determined based on the distribution of historical data; After calculating the feature density values one by one, a feature density graph is constructed with each data point as a node and the feature density value corresponding to the data point as the node weight. For any two different data dimensions in each subsequence, the feature correlation strength between the data dimensions is calculated pair by pair, and the edge weight value of the feature correlation graph is determined based on the change between the feature correlation strength in the current time window and the previous time window: ; in, Indicates the feature dimension in the current time window and feature dimensions The change in the strength of the correlation between is the feature dimension in the current time window and feature dimensions The Pearson correlation coefficient between is the Pearson correlation coefficient between the same dimensions in the previous time window; After calculating the correlation changes of all data dimension pairs, a feature correlation graph is constructed with each data dimension as a node and the correlation strength changes between nodes as edge weights. Update the feature density values of all nodes in the feature density graph and the weights of all edges in the feature association graph in real time to form a real-time updated feature density graph and feature association graph; By directly concatenating the feature density values of the feature density graph nodes and the associated changes of the feature association graph edges, representation data of the heterogeneous density-association space is formed.

[0021] In this embodiment, the representation data of the heterogeneous density-correlation space is input into the improved Transformer model, the local manifold curvature of each feature in the representation data is calculated in real time, and the self-attention weight of the model is calculated and updated using the local manifold curvature to obtain the curvature anomaly score of the heterogeneous space, specifically: Taking the representation data of the heterogeneous density-correlation space as input, the eigenvectors in the representation data are extracted one by one, and the local manifold curvature value between any two data dimensions is calculated based on the node density value of the feature density graph: ; in, Indicates the current time window Neidi , No. The local manifold curvature value of the local feature space between the data dimensions, Indicates the current time window Neidi In the data dimension The characteristic density value corresponding to the data point is Indicates the current time window Neidi In the data dimension The characteristic density value corresponding to the data point is 、 Represents the current time window Neidi , No. The mean of the feature density values of all data points in the data dimension, Indicates the current time window The total number of data points contained in the inner subsequence; The calculated local manifold curvature value is used as a constraint factor in the multi-head self-attention calculation process of the improved Transformer model and fused into the weight calculation process of the self-attention mechanism to constrain the model attention distribution. The fusion formula is: ; in, is the query matrix after the current eigenvector is linearly transformed, is the key matrix after the linear transformation of the current eigenvector, is the value matrix of the current eigenvector after linear transformation, is the fusion weight parameter of the local manifold curvature constraint, is a fixed dimension parameter of the model’s internal feature embedding space; According to the fusion formula, the self-attention weights corresponding to each attention head of the multi-head self-attention mechanism are updated in real time through the local manifold curvature value, and the real-time self-attention weight matrix of each attention head after being constrained by the local manifold curvature is obtained; Multiply the real-time self-attention weight matrix of each attention head after the manifold curvature constraint by the corresponding value matrix one by one to obtain the attention output vector after each attention head constraint. Then, concatenate all the attention head output vectors in sequence to obtain the complete multi-head self-attention constraint output vector. The complete multi-head self-attention constrained output vector is reduced to the same dimension as the initial feature vector through linear projection operation, and then the linearly projected vector is added to the initial input feature vector one by one through residual connection operation to obtain the intermediate residual representation vector of the feature vector; After performing layer-by-layer normalization on the intermediate residual representation vectors of the feature vectors, the abstract representation vectors of the feature vectors are obtained through a feedforward neural network. After the abstract representation vector undergoes a second residual connection and layer normalization process, it is input into the final fully connected output layer, and the fully connected layer is used to output the curvature anomaly score corresponding to each feature vector in real time.

[0022] In this embodiment, the improved Transformer model includes an initial feature linear mapping module, an adaptive manifold geometric mapping module, a multi-manifold geometric perception self-attention module, a residual connection and projection module, and an anomaly score output module: The initial feature linear mapping module is used to receive the representation data of the heterogeneous density-correlation space and use the linear projection weight matrix to map the input features into an initial query matrix, an initial key matrix and an initial value matrix respectively; The adaptive manifold geometry mapping module is used to receive the initial query matrix and the initial key matrix, and use the local manifold curvature matrix between the data dimensions in the current time window as a geometric constraint factor to perform a mapping transformation on the initial query matrix and the initial key matrix through a nonlinear adaptive mapping method to obtain the query matrix and key matrix after manifold geometry perception; The nonlinear adaptive mapping method is: ; ; in, 、 are the initial query matrix and the initial key matrix respectively, 、 They are respectively the query matrix and key matrix after manifold geometry perception, 、 are the weight matrices used for nonlinear mapping, is the nonlinear adaptive manifold mapping strength parameter, is the local manifold curvature matrix calculated in real time; The multi-head manifold geometry perception self-attention module is used to receive the query matrix, key matrix and initial value matrix after manifold geometry perception, use the local manifold curvature matrix as the attention mechanism constraint factor, update the attention weight matrix of the multi-head self-attention mechanism in real time, and perform matrix multiplication operation with the initial value matrix to obtain the attention output vector of each attention head, and then sequentially connect the attention output vectors of each attention head to obtain a complete multi-head self-attention constraint output vector; The attention mechanism constraints are: ; in, is the query matrix after manifold geometry perception, is the bond matrix after manifold geometry perception, is the initial value matrix, is the fusion weight parameter of the local manifold curvature constraint, is a fixed dimension parameter of the model’s internal feature embedding space; The residual connection and projection module is used to receive the complete multi-head self-attention constrained output vector, reduce the dimension of the multi-head self-attention constrained output vector to a space with the same dimension as the initial input feature vector through a linear projection operation, and then add the reduced dimension vector to the initial input feature vector one by one through a residual connection operation to obtain a residual representation vector of the feature vector; The anomaly score output module is used to receive the intermediate residual representation vector of the feature vector, obtain the abstract representation vector of the feature vector after layer normalization processing and feedforward neural network operation, and then input the abstract representation vector into the fully connected output layer through a second residual connection and layer normalization processing, and output the curvature anomaly score corresponding to each feature vector in real time.

[0023] In this embodiment, the curvature anomaly score based on the heterogeneous space estimates the generalized Pareto distribution parameters in real time, and dynamically calculates the dynamic threshold of the curvature anomaly score based on the real-time estimated generalized Pareto distribution parameters, specifically: In each time window, the curvature anomaly score data corresponding to all feature vectors is obtained in real time. After being sorted in descending order of value, a subset of anomaly score data exceeding the high quantile threshold is selected and recorded based on a preset high quantile threshold; the preset high quantile threshold is the 95th percentile of the historical curvature anomaly score data, which is determined by offline statistical analysis of historical data and pre-stored in the system before actual operation; Taking the recorded anomaly score data subset as input, the log-likelihood function is constructed, and then the scale parameter is solved in real time through an iterative optimization algorithm. With shape parameters The best estimate of : ; in, is the log-likelihood function of the real-time estimation of the generalized Pareto distribution parameters in the current time window, The scale parameter estimated in real time is used to describe the degree of discreteness of the tail distribution of the data. The shape parameter estimated in real time is used to describe the tail attenuation trend of the data tail distribution. Represents the first Rating values, Represents the preset high-percentile threshold of abnormality score, Indicates the total number of data contained in the current anomaly scoring data subset; The scale parameters estimated in real time and shape parameters Based on the total amount of abnormal score data recorded in real time, the value of the dynamic threshold of the curvature abnormal score in the current time window is dynamically determined: ; in, The specific value of the dynamic threshold of curvature anomaly score calculated in real time in the current time window. is the preset dynamic significance level parameter and represents the probability control threshold for being judged as abnormal. Indicates the total number of all curvature anomaly score data in the current time window; After the dynamic threshold is calculated, the curvature anomaly scores corresponding to all feature vectors in the current time window are numerically compared with the dynamic threshold one by one, and the curvature anomaly scores and corresponding feature vectors that exceed the dynamic threshold are marked and recorded one by one as the specific abnormal data determined in real time in the current time window; The number, specific values and corresponding feature vectors of specific abnormal data determined in real time within the current time window are completely stored in the real-time database, and the stored abnormal data results are used as data input for real-time estimation of the parameters of the next continuous time window.

[0024] In this embodiment, the curvature anomaly score is compared with the dynamic threshold in real time, and the charging pile data anomaly level is determined based on the comparison result, specifically: Receive the curvature anomaly score values corresponding to all feature vectors in real time in each time window, and obtain the curvature anomaly score dynamic threshold calculated in the current time window in real time; Calculate the difference between the curvature anomaly score value of each feature vector and the dynamic threshold value one by one: ; in, Indicates the feature vectors in the current time window The numerical difference between the curvature abnormality score and the dynamic threshold, Indicates the feature vectors in the current time window The curvature abnormality score value obtained by real-time calculation, Indicates the current time window The dynamic threshold value of curvature abnormality score obtained by real-time calculation; Based on the calculated real-time numerical difference corresponding to each eigenvector, the abnormal level of charging pile data is graded and judged one by one: when When the value is greater than 0.1 times the dynamic threshold, the corresponding charging pile data is judged to be a serious abnormality level and an abnormality level flag of "serious" is generated; when When the value is greater than 0 but less than or equal to 0.1 times the dynamic threshold, the corresponding charging pile data is determined to be of medium abnormality level, and an abnormality level indicator of "medium" is generated; when When it is less than or equal to 0, the corresponding charging pile data is determined to be at a slight abnormality level, and an abnormality level indicator of "minor" is generated; After the classification judgment operation is completed, the specific number of charging pile data judged as severe abnormality level, medium abnormality level and slight abnormality level in the current time window is counted in real time, and the characteristic vector identifier, curvature abnormality score value and difference calculation result corresponding to each abnormality level are recorded one by one; Based on the statistical and recorded results, a detailed record table of abnormality levels is generated in real time, including the feature vector identifier, abnormality level identifier, curvature abnormality score value, difference calculation results, and the corresponding real-time calculated dynamic threshold for each charging pile data; The generated abnormality level detailed record table is stored in the real-time database in real time, and the detailed record table is used as the input data source for the next continuous time window real-time abnormality level determination step to realize continuous, real-time, closed-loop dynamic cycle calculation of charging pile data abnormality level determination.

[0025] In this embodiment, the preset supervisory response rules are automatically triggered through the Kafka real-time message queue based on the abnormal level of the charging pile data, specifically: Read the characteristic vector identifier, abnormality level identifier, curvature abnormality score value, difference calculation result and dynamic threshold value of each charging pile device in the abnormality level detailed record table in real time within each time window; Perform the abnormal level identification judgment operation on each abnormal data record in the abnormal level detailed record table one by one: When the abnormality level is marked as "serious", the system will use the severity level message topic in the pre-configured Kafka message queue in the system to encapsulate the characteristic vector identifier of the current charging pile device, the abnormality level identifier, the curvature abnormality score value, the difference calculation result, the dynamic threshold value, the abnormality judgment timestamp and the unique device identifier in real time in a standard JSON message structure, and push the encapsulated JSON message to the operation and maintenance personnel's mobile terminal in real time; The real-time automatic dispatch program pre-configured in the operator's mobile terminal receives the above-mentioned Kafka message in real time and parses the message content in real time. Based on the pre-stored abnormal dispatch template, it automatically generates an emergency dispatch work order containing the device identification, device location, abnormal data details, real-time curvature abnormality score and corresponding difference value, and automatically sends the emergency dispatch work order to the emergency disposal queue of the operation and maintenance dispatch platform in real time through the automatic dispatch interface pre-installed in the operator's mobile terminal; When the abnormality level is marked as "medium", the system will use the medium-level message topic in the pre-configured Kafka message queue in the system to encapsulate the characteristic vector identifier, abnormality level identifier, curvature abnormality score value, difference calculation result, dynamic threshold value, abnormality judgment timestamp and device unique identifier of the current charging pile device in real time in a standard JSON message structure, and push the encapsulated JSON message to the abnormality warning report generation module in real time; After receiving and parsing the above Kafka messages in real time, the abnormal warning report generation module automatically generates a medium-level abnormal warning report based on the preset abnormal warning report template, which includes the device unique identifier, the device region, the abnormality level, the specific abnormal data details, the curvature abnormality score and difference information calculated in real time, and stores the generated warning report in its entirety in the abnormal warning report database in real time; When the abnormality level is marked as "minor", the system will use the pre-configured Kafka message queue in the minor message topic to encapsulate the characteristic vector identifier, abnormality level identifier, curvature abnormality score value, difference calculation result, dynamic threshold value, abnormality judgment timestamp and device unique identifier of the current charging pile device in real time using the standard JSON message structure, and push the encapsulated JSON message to the minor abnormality continuous tracking module in real time; After receiving and parsing the above-mentioned Kafka messages in real time, the minor anomaly continuous tracking module combines the parsed unique identifier of the charging pile device with the curvature anomaly score, difference result, dynamic threshold value, and anomaly judgment timestamp within the current time window to form a real-time record item, and updates the minor anomaly real-time record of the corresponding device in the pre-configured minor anomaly tracking database in real time; The minor anomaly continuous tracking module reads the updated minor anomaly tracking database records in real time within the next continuous time window to continuously track the continuous changes in the minor anomaly status of the current device and feeds back the change information to the database in real time to form a continuous tracking record of the abnormal status; A real-time dynamic cycle process that automatically triggers and implements corresponding regulatory response rules based on the abnormal level of charging pile data.

[0026] In this embodiment, a charging pile data transmission supervision system based on data analysis includes: The data acquisition module is used to collect multi-dimensional time series data of the voltage, current, power and temperature of the charging pile in real time, and divide the multi-dimensional time series data into several subsequences of the same length according to the set time window length; A heterogeneous density-association space data representation module is used to calculate the characteristic density value of the characteristic density map node and the correlation strength change of the characteristic correlation map edge for each subsequence point by point, and to update the characteristic density value of the characteristic density map node and the weight of the characteristic correlation map edge in real time, and to concatenate the updated characteristic density value of the characteristic density map node and the correlation change of the characteristic correlation map edge to form the representation data of the heterogeneous density-association space; An improved Transformer anomaly score calculation module is used to receive the representation data of the heterogeneous density-correlation space, calculate the local manifold curvature value between the data dimensions in real time, and use the local manifold curvature value as a constraint factor to update the self-attention weight of the Transformer model in real time, and output the heterogeneous space curvature anomaly score corresponding to each feature vector; A dynamic threshold calculation module is used to receive curvature anomaly score data in real time, estimate the scale parameter and shape parameter of the generalized Pareto distribution in real time based on the log-likelihood function of the generalized Pareto distribution, and calculate the dynamic threshold of the curvature anomaly score in real time based on the real-time estimated scale parameter and shape parameter; The abnormality level determination module is used to receive the curvature abnormality score and dynamic threshold in real time, calculate the numerical difference between the curvature abnormality score and the dynamic threshold one by one, determine the abnormality level of each charging pile device in real time based on the numerical difference, and generate a detailed abnormality level record table containing the feature vector identifier, abnormality level identifier, curvature abnormality score value and difference calculation result; The supervision response linkage module is used to receive the detailed record table of the abnormality level in real time, and automatically trigger the corresponding supervision response rules through the Kafka real-time message queue according to the abnormality level identifier, including automatically pushing serious abnormalities to the mobile terminal of the operation and maintenance personnel in real time and automatically starting the scheduling program, automatically generating early warning reports for moderate abnormalities in real time and storing them in the abnormality early warning report database, and updating minor abnormalities in real time and continuously tracking and storing them in the minor abnormality tracking database, so as to realize a real-time closed-loop system for charging pile data transmission supervision.

[0027] Example 1: To verify the feasibility of the present invention in practice, the present invention is applied to a new energy vehicle charging facility supervision scenario in a certain city. Real-time data transmission and abnormal status supervision are performed on a large number of charging piles deployed in the city to improve the accuracy of charging pile data anomaly identification and the supervision response speed. In this actual scenario, traditional supervision methods usually use fixed threshold alarms or basic statistical analysis, which are insufficient in identifying data anomalies, especially for multi-dimensional dynamically coupled data features such as voltage, current, power and temperature. This results in a low abnormal data recognition rate and severe response delays. In addition, traditional methods do not have an effective automatic supervision linkage mechanism and are unable to quickly handle serious anomalies, affecting the safety and stability of charging operations.

[0028] In actual operation, this embodiment first uses the data acquisition module to acquire multidimensional time series data of voltage, current, power, and temperature from the charging pile in real time. It then automatically divides this data into continuous subsequences based on a preset time window length. The system then calculates the feature density of each data point in the subsequence and further calculates the real-time variation in the strength of associations between different data dimensions, forming a heterogeneous density-correlation spatial data representation consisting of a feature density map and a feature correlation map. Specific parameters are set as a density calculation radius of 0.3 and a time window length of 5 seconds.

[0029] After obtaining the above-mentioned heterogeneous data representation, the system inputs the data into the improved Transformer anomaly score calculation module in real time. Through nonlinear adaptive mapping, the local manifold curvature matrix is explicitly introduced as a constraint factor, and the weight parameters of the multi-head self-attention mechanism within the model are updated. The model sets the number of attention heads to 8, the feature embedding dimension to 128, and the nonlinear mapping parameter λ to 0.5. After processing by the Transformer module, the curvature anomaly score value is output in real time.

[0030] Subsequently, this embodiment inputs the curvature anomaly score data in real time through the dynamic threshold calculation module, and estimates the generalized Pareto distribution parameters through the real-time optimization algorithm based on the log-likelihood function of the generalized Pareto distribution. Specifically, the initial value of the scale parameter σ(t) is 0.2, and the initial value of the shape parameter ξ(t) is 0.05. The real-time optimization algorithm is iteratively calculated, and the number of optimization iterations is 20. Stable and accurate generalized Pareto distribution parameters are obtained in real time in each time window, and the dynamic threshold of the curvature anomaly score is calculated in real time based on these parameters, and the dynamic significance level parameter α is set to 0.05.

[0031] The system then receives the above-mentioned curvature anomaly score and the corresponding dynamic threshold in real time through the anomaly level judgment module, calculates the difference between the score and the threshold one by one, and clearly divides the levels of severe anomaly, moderate anomaly and slight anomaly. The specific judgment threshold difference ratio parameter is set to 0.1 times the dynamic threshold, and a detailed record table of anomaly levels is generated in real time.

[0032] Finally, the supervision response linkage module of this embodiment reads the detailed record table of abnormality levels in real time, and performs real-time supervision linkage operations on the three types of abnormalities: severe, moderate, and minor, through the pre-deployed Kafka real-time message queue: severe abnormalities are pushed to the mobile terminal of the operation and maintenance personnel in real time, triggering the automatic scheduling program; moderate abnormalities generate early warning reports in real time and record them in the database; minor abnormalities are recorded in real time, and the abnormal status is continuously tracked.

[0033] In order to clearly demonstrate the actual improvement brought about by the present invention, the specific data of 5 charging pile devices are specially selected to compare the performance of the traditional threshold alarm method and the method of the present invention in terms of indicators such as abnormality recognition accuracy, response time and missed reporting rate. The data are shown in Table 1.

[0034] Table 1 Comparison of charging pile data anomaly recognition performance

[0035] It can be clearly seen from the data shown in Table 1 that the method of the present invention performs significantly better than the traditional threshold alarm method in terms of anomaly identification accuracy, avoiding the obvious omissions in anomaly identification by the traditional method. For example, for the CP-001 device, the traditional method failed to identify the anomaly, while the method of the present invention accurately identified and responded. At the same time, the present invention significantly shortens the response time, with the average response time stabilized at less than 20 milliseconds, while the response time of the traditional method exceeds 100 milliseconds, significantly improving the real-time and reliability of the system's anomaly handling. In addition, the dynamic threshold calculation method of the present invention can effectively adapt to changes in data characteristics, effectively reduce false alarms and omissions, and significantly enhance the adaptability of the supervision system to different operating conditions.

[0036] This embodiment fully verifies the effectiveness and superiority of the charging pile data transmission supervision method and system proposed in the present invention in addressing the problems of multi-dimensional data complexity and insufficient anomaly recognition accuracy in actual application scenarios. In particular, it achieves significant improvements in anomaly recognition accuracy and real-time supervision response compared with existing traditional methods, and has clear industrial applicability and promotion value.

[0037] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A charging pile data transmission supervision method based on data analysis, characterized in that: include: The multi-dimensional time series data collected by charging piles is divided into several subsequences, and a feature density map and a feature correlation map are constructed for each subsequence to obtain representation data of the heterogeneous density-correlation space. The representation data of the heterogeneous density-correlation space is input into the improved Transformer model to obtain the curvature anomaly score of the heterogeneous space; The curvature anomaly score in heterogeneous space is used to estimate the generalized Pareto distribution parameters, and the dynamic threshold of the curvature anomaly score is dynamically calculated based on the generalized Pareto distribution parameters. Compare the curvature anomaly score with the dynamic threshold and determine the charging pile data anomaly level based on the comparison result; Automatically trigger preset regulatory response rules based on the abnormal level of charging pile data through the Kafka real-time message queue; Execute regulatory linkage actions in real time through automatically triggered regulatory response rules.

2. A charging pile data transmission supervision method based on data analysis according to claim 1, characterized in that: The characterization data for constructing the heterogeneous density-correlation space is specifically: Split multidimensional time series data into continuous subsequences of equal length using a set time window length. Obtain the feature density value of each data point in the corresponding dimension in a point-by-point calculation manner; After calculating the feature density values one by one, a feature density graph is constructed with each data point as a node and the feature density value corresponding to the data point as the node weight. For any two different data dimensions in each subsequence, the feature correlation strength between the data dimensions is calculated pair by pair, and the edge weight value of the feature correlation graph is determined based on the change between the feature correlation strength in the current time window and the previous time window; Construct a feature association graph with each data dimension as a node and the change in the association strength between nodes as the weight of the edge; Update the feature density values of all nodes in the feature density graph and the weights of all edges in the feature association graph to form a real-time updated feature density graph and feature association graph; Directly concatenate the feature density values of the feature density graph nodes and the associated changes of the feature association graph edges to form representation data of the heterogeneous density-association space.

3. The method for supervising charging pile data transmission based on data analysis according to claim 1, characterized in that: The representation data of the heterogeneous density-correlation space is input into the improved Transformer model to obtain the curvature anomaly score of the heterogeneous space, specifically: Extract the eigenvectors from the data representing the heterogeneous density-correlation space one by one, and calculate the local manifold curvature value between any two data dimensions based on the node density value of the eigendensity graph; The local manifold curvature value is used as a constraint factor in the multi-head self-attention calculation process of the improved Transformer model and integrated into the weight calculation process of the self-attention mechanism; The self-attention weights corresponding to each attention head in the multi-head self-attention calculation process are updated through the local manifold curvature value to obtain the real-time self-attention weight matrix; Multiply the real-time self-attention weight matrix with the corresponding value matrix one by one to obtain the attention output vector after each attention head constraint. Then, concatenate all the attention head output vectors in sequence to obtain the multi-head self-attention constraint output vector. The multi-head self-attention constraint output vector is reduced in dimensionality by linear projection, and the linearly projected vector is added to the initial input feature vector one by one through the residual connection operation to obtain the intermediate residual representation vector of the feature vector; After layer-by-layer normalization of the intermediate residual representation vectors of the feature vectors, the abstract representation vectors are obtained through a feedforward neural network. After the abstract representation vector undergoes a second residual connection and layer normalization, it is input into the final fully connected output layer, which outputs the curvature anomaly score corresponding to each feature vector.

4. The method for supervising charging pile data transmission based on data analysis according to claim 3 is characterized in that: The improved Transformer model includes an initial feature linear mapping module, an adaptive manifold geometric mapping module, a multi-head manifold geometric perception self-attention module, a residual connection and projection module, and an anomaly score output module: The initial feature linear mapping module is used to receive representation data of the heterogeneous density-correlation space and map the input features into an initial query matrix, an initial key matrix and an initial value matrix respectively; The adaptive manifold geometry mapping module is used to receive the initial query matrix and the initial key matrix, and obtain the query matrix and the key matrix after manifold geometry perception; The multi-head manifold geometry perception self-attention module is used to query the matrix, the key matrix, and the initial value matrix, use the local manifold curvature matrix as the attention mechanism constraint factor, update the attention weight matrix of the multi-head self-attention mechanism in real time, and obtain a complete multi-head self-attention constraint output vector; The residual connection and projection module is used to receive the complete multi-head self-attention constrained output vector, reduce the dimension of the multi-head self-attention constrained output vector to a space with the same dimension as the initial input feature vector through a linear projection operation, and obtain the intermediate residual representation vector of the feature vector; The anomaly score output module is used to receive the intermediate residual representation vector of the feature vector, and after layer normalization processing and feedforward neural network operation, it inputs the abstract representation vector into the fully connected output layer through a second residual connection and layer normalization processing, and outputs the curvature anomaly score corresponding to each feature vector in real time.

5. The method for supervising charging pile data transmission based on data analysis according to claim 1, characterized in that: The curvature anomaly score is compared with the dynamic threshold, and the charging pile data anomaly level is determined based on the comparison result, specifically: Receive the curvature anomaly score values corresponding to all feature vectors in real time in each time window, and obtain the curvature anomaly score dynamic threshold calculated in the current time window in real time; Calculating the difference between the curvature anomaly score value of each eigenvector and the dynamic threshold value one by one; Based on the calculated real-time numerical difference corresponding to each characteristic vector, the abnormal level of charging pile data is graded and determined one by one; After the classification determination operation is completed, the specific number of charging pile data determined to be severe abnormality level, medium abnormality level and slight abnormality level in the current time window is counted, and the characteristic vector identifier, curvature abnormality score value and difference calculation result corresponding to each abnormality level are recorded one by one; Generate a detailed record table of abnormal levels based on the statistical and recorded results; The generated anomaly level detailed record table is stored in the real-time database in real time as the input data source for the next continuous time window real-time anomaly level determination step.

6. The method for supervising charging pile data transmission based on data analysis according to claim 1, characterized in that: The preset supervisory response rules are automatically triggered through the Kafka real-time message queue based on the abnormal level of charging pile data, specifically: Perform the abnormal level identification judgment operation on the abnormal data records in the abnormal level detailed record table one by one: When the abnormality level is marked as "serious", the severity level message topic in the pre-configured Kafka message queue in the system is pushed to the operation and maintenance personnel's mobile terminal in a standard JSON message structure; Operation and maintenance personnel automatically generate emergency dispatch work orders based on pre-stored abnormal dispatch templates and send them to the emergency response queue of the operation and maintenance dispatch platform; When the abnormality level is marked as "medium", it is pushed to the abnormality warning report generation module through the medium-level message topic in the pre-configured Kafka message queue in the system; After receiving and parsing Kafka messages in real time, the abnormal warning report generation module automatically generates a medium-level abnormal warning report in real time based on the preset abnormal warning report template and stores it in the abnormal warning report database; When the exception level is marked as "minor", it is pushed to the minor exception continuous tracking module in a standard JSON message structure through the minor level message topic in the pre-configured Kafka message queue in the system; After the minor anomaly continuous tracking module receives and parses Kafka messages in real time, it updates the minor anomaly real-time records of the corresponding devices in the pre-configured minor anomaly tracking database; The minor anomaly continuous tracking module reads the updated minor anomaly tracking database record within the next continuous time window, and continuously tracks the continuous changes of the minor anomaly status of the current device in real time to form a continuous tracking record of the abnormal status; A real-time dynamic cycle process that automatically triggers and implements corresponding regulatory response rules based on the abnormal level of charging pile data.

7. A charging pile data transmission supervision system based on data analysis, executing a charging pile data transmission supervision method based on data analysis according to any one of claims 1 to 7, characterized in that: include: The data acquisition module is used to collect multi-dimensional time series data in real time and divide it into several subsequences of the same length according to the set time window length; The heterogeneous density-correlation space data representation module is used to calculate the feature density value of the feature density graph node and the change in the correlation strength of the feature correlation graph edge for each subsequence point by point, and construct the representation data of the heterogeneous density-correlation space; An improved Transformer anomaly score calculation module is used to receive the representation data of the heterogeneous density-correlation space, and use the local manifold curvature value as a constraint factor to update the self-attention weight of the Transformer model in real time, and output the heterogeneous space curvature anomaly score corresponding to each feature vector; A dynamic threshold calculation module is used to receive curvature anomaly score data in real time and calculate the dynamic threshold of curvature anomaly score based on the generalized Pareto distribution log-likelihood function; The abnormality level determination module is used to receive the curvature abnormality score and dynamic threshold in real time, calculate the numerical difference between the curvature abnormality score and the dynamic threshold one by one, and determine the abnormality level of each charging pile device in real time based on the numerical difference, and generate a detailed record table of the abnormality level; The supervision response linkage module is used to receive the detailed record table of the abnormal level in real time, and automatically trigger the corresponding supervision response rules through the Kafka real-time message queue according to the abnormal level identifier.

Images