Rapid detection, diagnosis and debugging system for direct current charging pile

Abstract

The invention, which belongs to the technical field of the direct-current charging pile, provides a direct-current charging pile rapid detection diagnosis debugging system comprising a detection host, a multi-channel data acquisition module, an electric energy quality analysis module, an insulation resistance test module, a grounding resistance test module, a communication protocol analysis module, a temperature monitoring module and a fault simulation module. Parallel data acquisition is realized by adopting 16 independent analog signal acquisition channels and 8 digital signal acquisition channels, a parallel processing unit matched with the number of detection channels is started by utilizing a CUDA flow technology, preliminary fault screening is carried out at a GPU end through a rapid shallow model, and a layered diagnosis mechanism of accurate fault analysis is carried out at a CPU end through a deep model. And in combination with the charging pile fault identification model, real-time monitoring of the operation state of the charging pile and accurate identification and positioning of the fault are realized, and the technical problems of low real-time processing efficiency of multi-channel data and insufficient fault diagnosis accuracy are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of direct current charging piles, and in particular relates to a rapid detection, diagnosis and debugging system for direct current charging piles. BACKGROUND

[0002] With the rapid development of the electric vehicle industry, direct current charging piles as key infrastructure need to be comprehensively detected and diagnosed to ensure safe and reliable operation. The traditional charging pile detection technology mainly adopts a single-channel sequential sampling method to collect data, and a simple threshold judgment method is used for fault identification. The detection equipment is usually configured with independent voltage test modules, current test modules, insulation test modules and other discrete detection units, and there is a lack of effective data fusion mechanism between the modules. In the current charging pile detection application scenario, since the charging pile contains complex power electronic devices, communication control systems, temperature management systems and other multiple subsystems, it is necessary to monitor multi-dimensional parameters such as voltage, current, power, temperature, insulation resistance, ground resistance, communication protocol, etc. The traditional detection method has obvious processing delay and efficiency bottleneck when processing multi-channel data. The traditional technology lacks efficient parallel processing capability and intelligent fault identification algorithm when facing a large amount of multi-source heterogeneous data, resulting in problems such as long detection period, low diagnosis accuracy, high misjudgment rate, etc. Especially in the case of complex fault mode and multi-fault coupling, the performance is not good. That is to say, the existing technology has the technical problems of low real-time processing efficiency of multi-channel data and insufficient fault diagnosis accuracy in the process of detecting and diagnosing direct current charging piles. SUMMARY

[0003] Therefore, the application provides a rapid detection, diagnosis and debugging system for direct current charging piles, which can solve the technical problems of low real-time processing efficiency of multi-channel data and insufficient fault diagnosis accuracy in the process of detecting and diagnosing direct current charging piles.

[0004] The application is implemented in the following manner: the application provides a DC charging pile rapid detection diagnosis debugging system, which comprises a detection host, a multi-channel data acquisition module, an electric energy quality analysis module, an insulation resistance test module, a grounding resistance test module, a communication protocol analysis module, a temperature monitoring module, a fault simulation module and a control chip, and the control chip is provided with a charging pile intelligent diagnosis module, which performs the following steps: controlling the multi-channel data acquisition module to comprehensively scan the parameters of the charging pile, acquiring voltage parameters, current parameters, power parameters, temperature distribution data, insulation resistance values, grounding resistance values and communication protocol states, and establishing a charging pile operation state benchmark database; pre-processing the collected data; using a sliding time window method to analyze the pre-processed data in segments, extracting key feature parameters; identifying abnormal feature points and classifying and marking them according to fault types; inputting the abnormal feature points into a charging pile fault identification model to obtain fault type identification, fault severity level and fault position identification; constructing a fault correlation matrix and optimizing fault propagation path analysis by using a minimum spanning tree algorithm; calculating optimal detection parameter configurations by using a charging pile parameter optimization function; starting a CUDA stream, inputting fault feature data into a rapid shallow layer model to perform preliminary fault scoring, using a Transformer attention mechanism to extract relevant context fragments for data higher than a suspicious threshold, and sending the data into a deep layer model for deep analysis; and generating a charging pile diagnosis report.

[0005] The multi-channel data acquisition module is designed in a modular manner and comprises 16 independent analog signal acquisition channels and 8 digital signal acquisition channels, each analog signal acquisition channel is provided with a 24-bit high-precision analog-to-digital converter and a programmable gain amplifier, the sampling frequency reaches 10 kHz, and the measurement accuracy is better than 0.1%.

[0006] The electric energy quality analysis module is provided with a fast Fourier transform processor, which is used for real-time calculation of harmonic components of voltage parameters and current parameters, the frequency range is covered from DC to 5 kHz, and the harmonic analysis accuracy reaches 0.01%.

[0007] The insulation resistance test module is provided with a DC high-voltage source with adjustable output voltage, the output voltage range is 500V to 5000V, the test range covers 1MΩ to 10GΩ, and a leakage current compensation circuit is built-in to ensure test accuracy.

[0008] The sliding time window is specifically used for segmenting continuous data streams according to a preset time length, each window contains 5 seconds of data, and the windows overlap by 50%, which is used for capturing dynamic change characteristics of the charging pile operation state.

[0009] The key feature parameters are specifically important indicators extracted from each time window that can reflect the operating state of the charging pile, including voltage fluctuation coefficient, current harmonic distortion rate, power factor change rate, temperature rise rate, insulation resistance drop rate, and communication packet loss rate.

[0010] The fault correlation matrix is a two-dimensional matrix that describes the mutual influence between different fault type identifiers, and the matrix element value represents the correlation strength between faults.

[0011] The CUDA stream is an execution unit for GPU parallel computing, used to implement parallel processing of multiple detection channels, with each CUDA stream processing the data of one detection channel independently to avoid data processing blockage.

[0012] The fast shallow model is a lightweight neural network model deployed on the GPU side for preliminary fault screening, which implements fast inference through a simplified network structure, including 2-layer neural network and ReLU activation function.

[0013] The deep model is a complex neural network model deployed on the CPU side for accurate fault analysis, which improves the accuracy and reliability of fault identification through a deep network structure, including 8-layer Transformer encoder and multi-head attention mechanism.

[0014] The specific structure of the charging pile fault identification model is a multi-modal fusion network based on Transformer architecture, including electrical parameter encoder, temperature distribution encoder, communication state encoder, multi-head attention fusion layer, fast shallow model, and deep model, with the number of attention heads dynamically adjusted according to the three parameters of fault type number, data dimension, and model complexity.

[0015] The charging pile parameter optimization function is used to calculate the optimal detection parameter configuration according to the fault type identifier and detection requirements, with input including fault type identifier, fault severity level, environmental temperature value, load characteristic parameter, and historical fault frequency statistics, and output being an optimized parameter combination containing sampling frequency, test voltage, test current, temperature threshold, and communication timeout time.

[0016] The charging pile intelligent diagnosis algorithm function is used to adjust the number of attention heads of the charging pile fault identification model based on fault complexity, data quality, and computing resources, and a comprehensive evaluation value is obtained, which is used to adjust the number of attention heads using different weight adjustment functions when the comprehensive evaluation value belongs to different ranges.

[0017] The electrical fault type is obtained by voltage parameter and current parameter waveform analysis, accounts for 45% of the total fault proportion, and the contribution rate is 0.8; the mechanical fault type is obtained by temperature distribution data and vibration monitoring, accounts for 25% of the total fault proportion, and the contribution rate is 0.6; the communication fault type is obtained by communication protocol state analysis and response time analysis, accounts for 20% of the total fault proportion, and the contribution rate is 0.4.

[0018] The temperature monitoring module is configured with 32 wireless temperature sensors, each sensor is built-in 2.4GHz wireless communication module, the temperature measurement range is-40 DEG C to 150 DEG C, the temperature measurement accuracy is ± 0.5 DEG C, and the fault simulation module is configured with programmable electronic load, and the power range is 0 to 100kW.

[0019] The training of the charging pile fault identification model uses the Adam optimizer for gradient descent training, sets the learning rate to 0.001, the batch size to 32, the training rounds to 200 rounds, adopts the cross-entropy loss function to calculate the classification loss, prevents overfitting through the early stopping mechanism, and uses the learning rate decay strategy to improve the model convergence effect.

[0020] The present application realizes efficient real-time processing and accurate fault identification of multi-dimensional parameters of the charging pile by constructing a multi-channel data acquisition system based on the CUDA stream parallel processing architecture, combining the layered diagnosis mechanism of the fast shallow model and the deep model. The present application adopts a parallel data acquisition mode of 16 independent analog signal acquisition channels and 8 digital signal acquisition channels, realizes synchronous processing of multiple detection channels through CUDA stream technology, avoids time delay caused by traditional sequential sampling, and simultaneously uses a multi-modal fusion network based on the Transformer architecture to uniformly encode and correlate analysis of heterogeneous data such as electrical parameters, temperature distribution, communication state, etc., which significantly improves the fusion efficiency of multi-source data and the accuracy of fault identification. The present application uses a fast shallow model to perform preliminary fault screening on the GPU side, and only starts the deep model for accurate analysis of suspicious data, effectively balancing the relationship between processing speed and diagnosis accuracy, overcoming the efficiency bottleneck and accuracy deficiency problems of traditional methods in large data processing, and realizing real-time and efficient processing of multi-channel data and accurate identification and diagnosis of faults. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 The overall structure schematic diagram of the system provided by the present application is shown.

[0022] Figure 2 The flow chart of the execution steps of the charging pile intelligent diagnosis module in the present application is shown.

[0023] Figure 3 The neural network structure schematic diagram of the charging pile fault identification model is shown.

[0024] Figure 4 A multi-channel data acquisition parameter monitoring chart for Example 2.

[0025] Figure 5 A temperature monitoring distribution chart for Example 2.

[0026] Figure 6 A CUDA parallel processing time comparison chart for Example 2 DETAILED DESCRIPTION

[0027] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application.

[0028] As Figure 1 shown is a composition schematic diagram of a direct current charging pile rapid detection, diagnosis and debugging system provided by the present application, the system comprises: a detection host, a multi-channel data acquisition module, an electric energy quality analysis module, an insulation resistance test module, a grounding resistance test module, a communication protocol analysis module, a temperature monitoring module, a fault simulation module and a control chip, wherein the detection host is connected with a direct current charging pile through a detection cable group, the multi-channel data acquisition module is arranged inside the detection host and is used for acquiring voltage parameters, current parameters and power parameters of the direct current charging pile; the electric energy quality analysis module is electrically connected with the multi-channel data acquisition module through a signal conditioning circuit and is used for analyzing voltage harmonic components, current harmonic components and power factor parameters; the insulation resistance test module is electrically connected with an output port of the detection host through a high-voltage isolation circuit and is used for testing insulation resistance values of each loop of the charging pile; the grounding resistance test module is electrically connected with a grounding system of the charging pile through a grounding test probe and is used for testing grounding resistance values and grounding continuity states; the communication protocol analysis module is electrically connected with a charging pile controller through a CAN bus interface and is used for analyzing communication protocol states and running state information of the charging pile; the temperature monitoring module is connected with key positions of the charging pile through wireless temperature sensors and is used for monitoring internal temperature distribution data of the charging pile; the fault simulation module is connected with an output end of the charging pile through a programmable load and is used for simulating various fault working conditions; the control chip is arranged inside the detection host, the control chip is electrically connected with the multi-channel data acquisition module, the electric energy quality analysis module, the insulation resistance test module, the grounding resistance test module, the communication protocol analysis module, the temperature monitoring module and the fault simulation module and performs data interaction; the control chip is provided with a charging pile intelligent diagnosis module inside and is used for performing fusion analysis on the collected multi-source data, identifying fault type identifiers and fault position identifiers of the charging pile and finally obtaining a charging pile diagnosis report.

[0029] The multi-channel data acquisition module adopts a modular design and comprises 16 independent analog signal acquisition channels and 8 digital signal acquisition channels, each analog signal acquisition channel is configured with a 24-bit high-precision analog-to-digital converter and a programmable gain amplifier, the sampling frequency reaches 10 kHz, and the measurement accuracy is better than 0.1%; the power quality analysis module is internally provided with a fast Fourier transform processor for real-time calculation of harmonic components of voltage parameters and current parameters, the analysis frequency range covers direct current to 5 kHz, and the harmonic analysis accuracy reaches 0.01%; the insulation resistance test module is configured with a DC high-voltage source with adjustable output voltage, the output voltage range is 500V to 5000V, the test range covers 1MΩ to 10GΩ, and the built-in leakage current compensation circuit ensures test accuracy; the ground resistance test module adopts a three-wire test principle and is configured with an AC signal generator and a phase-sensitive detector, the test frequency is 128Hz, and the test range is 0.01Ω to 2000Ω; the communication protocol analysis module supports CAN2.0A protocol, CAN2.0B protocol, CANopen protocol and J1939 protocol, and is configured with an optical isolation interface to ensure test safety; the temperature monitoring module is configured with 32 wireless temperature sensors, each sensor is internally provided with a 2.4GHz wireless communication module, the temperature measurement range is-40℃ to 150℃, and the temperature measurement accuracy is ±0.5℃; the fault simulation module is configured with a programmable electronic load with a power range of 0 to 100kW, which is used to simulate overload faults, short circuit faults and open circuit faults.

[0030] As shown in Figure 2 The charging pile intelligent diagnosis module is used to perform the following steps: S01, controlling the multi-channel data acquisition module to perform comprehensive parameter scanning on the charging pile, simultaneously acquiring voltage parameters, current parameters, power parameters, temperature distribution data, insulation resistance values, ground resistance values and communication protocol states, and establishing a charging pile operation state benchmark database; S02, preprocessing the acquired voltage parameters, current parameters, power parameters, temperature distribution data, insulation resistance values, ground resistance values and communication protocol states, including data cleaning, filtering and noise reduction, data normalization and abnormal value detection, to ensure that the data quality meets the requirements of subsequent analysis; S03, using a sliding time window method to perform segmented analysis on the preprocessed voltage parameters, current parameters, power parameters, temperature distribution data, insulation resistance values, ground resistance values and communication protocol states, and extracting key feature parameters in each time window, including voltage fluctuation coefficient, current harmonic distortion rate, power factor change rate, temperature rise rate, insulation resistance drop rate and communication packet loss rate; S04, according to the pre-set fault feature threshold, identify the abnormal feature points in the key feature parameters, including voltage fluctuation abnormal points, current harmonic abnormal points, power factor abnormal points, temperature abnormal points, insulation resistance abnormal points, communication abnormal points, and classify and mark the abnormal feature points according to the electrical fault type, mechanical fault type, communication fault type and environmental fault type; S05, input the classified and marked abnormal feature points into the charging pile fault identification model, and get the fault type identification, fault severity level and fault location identification through model reasoning; S06, based on the fault type identification and fault location identification, a fault correlation matrix is constructed, and a minimum spanning tree algorithm is used to optimize the fault propagation path analysis to determine the causal relationship between the main fault source and the secondary fault source; S07, calculate the optimal detection parameter configuration through the charging pile parameter optimization function, the function input includes fault type identification, fault severity level, environmental temperature value, load characteristic parameter and historical fault frequency statistics, and the output is the optimized detection strategy parameter; S08, start a sufficient number of CUDA streams, the number matches the number of detection channels, start two layers of sub-thread grids in each CUDA stream, if the number of detection channels exceeds the number of CUDA streams, use multi-batch processing; S09, each CUDA stream inputs the fault feature data allocated to the fast shallow model of the charging pile fault identification model, and performs preliminary fault scoring to determine whether it is below the suspicious threshold; S10, for the fault feature data with preliminary score below the suspicious threshold, the first layer of sub-thread grids directly returns the result to the CPU side and enters S12; for the fault feature data with preliminary score above the suspicious threshold, use the Transformer attention mechanism to automatically extract the context fragments highly related to the fault feature data from the charging pile historical data; S11, input the extracted context fragments and the fault feature data into the deep model of the charging pile fault identification model, and the second layer of sub-thread grids performs deep analysis on the complete sequence to output accurate fault score; S12, the second layer of sub-thread grids returns the deep scoring result to the CPU side, if the deep score is higher than the judgment threshold, the CPU judges that there is a fault, triggers the diagnosis mechanism, records and archives the fault feature data, and generates fault warning information; otherwise, enter S13; S13, the CPU records and appends the fault feature data not judged as fault to the charging pile historical data, ensures that the complete context can be obtained in subsequent detection, and generates a charging pile diagnosis report according to all processing results.

[0031] The sliding time window is specifically segmenting continuous data streams according to a preset time length, each window containing 5 seconds of data, and the windows overlapping by 50%, for capturing dynamic change characteristics of the operating state of the charging pile; the key feature parameters are specifically important indicators extracted from each time window that can reflect the operating state of the charging pile, including voltage fluctuation coefficient, current harmonic distortion rate, power factor change rate, temperature rise rate, insulation resistance drop rate, and communication packet loss rate, the key feature parameters can effectively identify abnormal operating states of the charging pile; the fault association matrix is specifically a two-dimensional matrix describing the mutual influence degree between different fault type identifiers, the matrix element value representing the association strength between faults, and the fault propagation path and root cause are identified by analyzing the matrix structure.

[0032] The CUDA stream is specifically an execution unit for GPU parallel computing, used to realize parallel processing of multiple detection channels, each CUDA stream independently processing the data of one detection channel to avoid data processing blockage; the fast shallow model is specifically a lightweight neural network model deployed on the GPU end for preliminary fault screening, realizing fast inference through a simplified network structure and reducing unnecessary deep calculation; the deep model is specifically a complex neural network model deployed on the CPU end for accurate fault analysis, improving the accuracy and reliability of fault identification through a deep network structure.

[0033] The context segment is specifically historical information related to the current fault feature data extracted from the charging pile historical data, used to provide background information and trend data for fault analysis. The charging pile historical data is specifically all detection data and diagnosis results accumulated during system operation, including historical fault feature data, fault type identifier, and fault handling result, used to support fault pattern recognition and prediction analysis.

[0034] The specific structure of the charging pile fault identification model is a multi-modal fusion network based on the Transformer architecture, as shown in Figure 3As shown, the charging pile fault identification model includes an electrical parameter encoder, a temperature distribution encoder, a communication state encoder, a multi-head attention fusion layer, a fast shallow model, and a deep model, wherein the number of attention heads of the multi-head attention mechanism is dynamically adjusted according to three parameters of the number of fault types, the data dimension, and the model complexity, and the calculation formula is that the number of attention heads is equal to the square root of the number of fault types multiplied by the ratio of the data dimension to the model complexity; the fast shallow model is used for preliminary fault scoring on the GPU side, and includes 2 layers of neural networks and a ReLU activation function, and the inference time is less than 1 ms; the deep model is used for deep fault analysis on the CPU side, and includes 8 layers of Transformer encoders and a multi-head attention mechanism, and the inference time is 5 ms to 10 ms; the steps of establishing the training data set of the charging pile fault identification model specifically include collecting normal operation data and fault data of charging piles of different brands and models, establishing a multi-dimensional data set containing voltage parameters, current parameters, power parameters, temperature distribution data, insulation resistance values, ground resistance values, communication protocol states, and fault type identifiers, expanding the sample quantity through a data enhancement technique, dividing the training set, the validation set, and the test set by using a cross-validation method, and ensuring the representativeness and generalization ability of the data set; the steps of training the charging pile fault identification model specifically include using an Adam optimizer for gradient descent training, setting the learning rate to 0.001, the batch size to 32, and the number of training rounds to 200 rounds, using a cross-entropy loss function to calculate the classification loss, preventing overfitting through an early stopping mechanism, using a learning rate decay strategy to improve the model convergence effect, and finally obtaining a trained model capable of accurately identifying the fault type identifier.

[0035] The charging pile parameter optimization function is used for calculating the optimal detection parameter configuration according to the fault type identifier and the detection requirements, the input includes the fault type identifier, the fault severity level, the environmental temperature value, the load characteristic parameter, and the historical fault frequency statistics, and the output is an optimal parameter combination containing the sampling frequency, the test voltage, the test current, the temperature threshold, and the communication timeout time.

[0036] The charging pile intelligent diagnosis algorithm function is used to adjust the number of attention heads of the charging pile fault identification model. The function is based on three data calculations of fault complexity, data quality and computing resources to obtain a comprehensive evaluation value. When the comprehensive evaluation value belongs to different ranges, different weight adjustment functions are used to adjust the number of attention heads of the charging pile fault identification model. When the comprehensive evaluation value is less than 0.3, a linear growth function is used, the number of attention heads is set to 4 to 8, and it is suitable for simple fault scenarios; when the comprehensive evaluation value is between 0.3 and 0.6, a quadratic function is used for adjustment, the number of attention heads is set to 8 to 16, and it is suitable for medium complexity fault scenarios; when the comprehensive evaluation value is between 0.6 and 0.8, an exponential function is used for adjustment, the number of attention heads is set to 16 to 32, and it is suitable for complex fault scenarios; when the comprehensive evaluation value is greater than 0.8, a logarithmic function is used for adjustment, the number of attention heads is set to 32 to 64, and it is suitable for extremely complex fault scenarios.

[0037] The electrical fault type is obtained by voltage parameter and current parameter waveform analysis, accounting for 45% of the total fault proportion, directly affecting the charging efficiency, and the contribution rate of the electrical fault type is 0.8; the mechanical fault type is obtained by temperature distribution data and vibration monitoring, accounting for 25% of the total fault proportion, affecting the service life of the equipment, and the contribution rate of the mechanical fault type is 0.6; the communication fault type is obtained by communication protocol state analysis and response time analysis, accounting for 20% of the total fault proportion, affecting the user experience, and the contribution rate of the communication fault type is 0.4; the environmental fault type is obtained by temperature distribution data and environmental monitoring, accounting for 10% of the total fault proportion, affecting the stability of the equipment, and the contribution rate of the environmental fault type is 0.2. The contribution rates of different fault types are different, and the contribution rates directly affect the accuracy of fault diagnosis and the allocation strategy of detection resources. Proportion imbalance will lead to detection precision decline and resource waste. By dynamically adjusting the detection weight and sampling frequency of each type of fault, an adaptive fault diagnosis mechanism is established to reduce the negative impact of proportion imbalance on system performance.

[0038] The specific implementation of the above steps is described in detail below.

[0039] The specific implementation of step S01 is to send instructions to the multi-channel data acquisition module through the control chip to start the comprehensive parameter scanning program. First, system initialization is performed, including hardware self-test, channel calibration, reference voltage setting, to ensure that all acquisition channels are in normal working condition. Then, according to the preset sampling sequence, each detection module is activated in turn for data acquisition. Voltage parameter acquisition obtains the primary side voltage signal through a high-precision voltage transformer, which is input into an analog-to-digital converter after impedance matching and amplitude adjustment by a signal conditioning circuit. Current parameter acquisition obtains the current signal through a Rogowski coil sensor, avoiding the saturation problem of traditional current transformers. Power parameter is obtained by real-time calculation of the product of voltage and current, and power factor correction is performed at the same time. Temperature distribution data is collected through a wireless sensor network, and each sensor reports temperature data to the data aggregation node at regular intervals. Insulation resistance value is calculated by applying a test voltage and measuring leakage current, and the test process uses automatic range switching technology. Grounding resistance value is obtained by injecting an AC test signal and measuring loop impedance. The communication protocol state is monitored through the CAN bus interface to listen to the data frames of the charging pile controller, and the protocol format and data content are analyzed. The purpose of establishing the charging pile operating state benchmark database is to provide a reference standard for normal working conditions for subsequent fault diagnosis.

[0040] The specific implementation of step S02 is to preprocess the collected raw data to ensure that the data quality meets the requirements of subsequent analysis. The data cleaning process first identifies and eliminates obvious outliers, uses the 3σ criterion to judge whether the data point is outside the normal range, and marks the data points outside the range as outliers and performs interpolation repair. The filter denoising process uses a Butterworth low-pass filter, with a cutoff frequency set to 2 times the signal frequency, and a filter order of 4 to effectively suppress high-frequency noise interference. Data normalization processing unifies parameters of different dimensions to a value range of 0 to 1, using the maximum and minimum value normalization method to avoid the influence of numerical range differences on subsequent algorithms. The outlier detection method is based on statistical distribution, calculating the mean and standard deviation of each parameter, and marking data points deviating from the normal distribution by more than 3 standard deviations as outliers. When the proportion of outliers exceeds 5%, a data quality warning is triggered. The preprocessing process also includes time synchronization correction to ensure consistency of data from different acquisition channels on the time axis, and linear interpolation method is used to align the time stamps.

[0041] The specific implementation of step S03 is to use the sliding time window method to analyze the preprocessed data in segments and extract key feature parameters that can reflect the operating status of the charging pile. The sliding time window is set to 5 seconds long, with a 50% overlap between windows, i.e., a new analysis window is generated every 2.5 seconds, ensuring that the continuous change process of the system state can be captured. Within each time window, the fluctuation coefficient of the voltage parameter is calculated, and the ratio of the standard deviation to the mean is used to measure the voltage stability. Under normal circumstances, the fluctuation coefficient should be less than 2%. The current harmonic distortion rate is calculated by analyzing the frequency domain characteristics of the current waveform using fast Fourier transform, and the ratio of the amplitude of each harmonic to the amplitude of the fundamental wave is calculated. The total harmonic distortion rate should be less than 5% under normal circumstances. The power factor change rate is calculated by comparing the power factor difference between adjacent time windows, reflecting the trend of load characteristics. The temperature rise rate is obtained by linear regression analysis of the change slope of the temperature data, which should be less than 1℃ / min under normal circumstances. The insulation resistance drop rate is calculated by comparing the difference between the current value and the historical reference value, reflecting the degree of insulation performance degradation. The communication packet loss rate is obtained by calculating the ratio of the number of communication frames that have not received a response to the total number of sent frames within the time window, which should be less than 1% under normal circumstances.

[0042] The specific implementation of step S04 is to identify abnormal feature points in key feature parameters according to pre-set fault feature threshold values, and to classify and label fault types. The determination threshold of voltage fluctuation abnormal points is set to a fluctuation coefficient greater than 5%, indicating that there is a problem with power supply quality or internal voltage regulation circuit failure. The determination threshold of current harmonic abnormal points is set to a total harmonic distortion rate greater than 10%, indicating that the power device is working abnormally or the load characteristics have changed. The determination threshold of power factor abnormal points is set to a power factor less than 0.85 or a change rate greater than 10% / min, indicating that the reactive power compensation device is faulty or the load is unbalanced. The determination threshold of temperature abnormal points is set to a temperature rise rate greater than 5℃ / min or an absolute temperature exceeding 85℃, indicating that the heat dissipation system is faulty or overloaded. The determination threshold of insulation resistance abnormal points is set to an insulation resistance value less than 1MΩ or a drop rate greater than 20% / day, indicating that the insulation material is aging or damp. The determination threshold of communication abnormal points is set to a packet loss rate greater than 5% or a response time exceeding 500ms, indicating that the communication link is faulty or the controller is abnormal. Abnormal feature points are classified and labeled according to the fault source, including electrical fault types such as voltage fluctuation abnormality, current harmonic abnormality, and power factor abnormality, mechanical fault types such as temperature abnormality and vibration abnormality, communication fault types such as communication abnormality and protocol parsing error, and environmental fault types such as temperature abnormality and humidity abnormality.

[0043] The specific implementation of step S05 is to input the classified and labeled abnormal feature points into the charging pile fault identification model, and obtain the fault type identification, fault severity level and fault location identification through deep learning model inference. The fault identification model adopts a multi-modal fusion network based on the Transformer architecture, which can simultaneously process different types of input data such as electrical parameters, temperature data, and communication status. The model input layer vectorizes and encodes the abnormal feature points, the electrical parameters are encoded into a 128-dimensional feature vector through a multi-layer perceptron, the temperature data are encoded into a 64-dimensional feature vector through a one-dimensional convolutional network, and the communication status is encoded into a 32-dimensional feature vector through an embedding layer. The multi-head attention mechanism calculates the attention weights according to the correlation between different modal data, and integrates multi-source information for comprehensive judgment. The model output layer outputs the probability distribution of the fault type through the softmax activation function, and selects the class with the maximum probability as the fault type identification. The fault severity level is divided into four levels: slight, general, severe, and critical, and is quantitatively evaluated according to the number and deviation of abnormal feature points. The fault location identification is determined by analyzing the abnormality of different detection points, and a weighted voting mechanism is used to integrate the results of multiple detection channels.

[0044] The specific implementation of step S06 is to construct a fault correlation matrix based on the fault type identification and the fault location identification, optimize the fault propagation path analysis using the minimum spanning tree algorithm, and determine the causal relationship between the primary fault source and the secondary fault source. The fault correlation matrix is a two-dimensional matrix that describes the mutual influence between different fault types, and the matrix element value represents the correlation strength between faults, with a value range of 0 to 1. The matrix construction process first calculates the conditional probability as the initial value of the correlation strength according to the historical fault data statistics of the simultaneous occurrence frequency of different fault types. Then, the physical model is introduced to correct the correlation strength, considering the propagation mechanism and influence range of the fault in the electrical system. The minimum spanning tree algorithm converts the fault correlation matrix into an undirected weighted graph, taking the fault type as the node and the correlation strength as the edge weight, and constructs the minimum spanning tree through the Kruskal algorithm or Prim algorithm. The root node of the spanning tree represents the primary fault source, the leaf node represents the secondary fault source, and the branch structure of the tree represents the fault propagation path. By analyzing the topological structure of the spanning tree, the key fault nodes and propagation bottlenecks are identified, providing a basis for fault handling priority sorting.

[0045] The specific implementation of step S07 is to calculate the optimal detection parameter configuration through the charging pile parameter optimization function, and realize the adaptive adjustment of the detection strategy. The optimization function uses a multi-objective optimization algorithm, and simultaneously considers multiple optimization objectives such as detection accuracy, detection efficiency, and resource consumption. The input parameters include fault type identification, fault severity level, environmental temperature value, load characteristic parameter, and historical fault frequency statistics. The fault type identification determines the detection focus and parameter weight. The electrical fault needs to increase the voltage and current sampling frequency, and the mechanical fault needs to increase the temperature monitoring density. The fault severity level affects the detection accuracy requirement. The serious fault needs higher precision measurement parameters. The environmental temperature value affects the device working characteristics and test accuracy. The high temperature environment needs to reduce the test voltage to avoid insulation breakdown. The load characteristic parameter affects the test strategy selection. The resistive load and inductive load need different test methods. The historical fault frequency statistics provide fault prediction information. The high-frequency fault area needs to be monitored. The optimization function outputs the optimal configuration combination of parameters such as sampling frequency, test voltage, test current, temperature threshold, and communication timeout time, and performs global optimization through genetic algorithm or particle swarm algorithm.

[0046] The specific implementation of step S08 is to start a sufficient number of CUDA streams to realize GPU parallel computing. Each CUDA stream corresponds to a data processing task of a detection channel. CUDA stream is the execution unit of GPU parallel computing, which can realize parallel processing of multiple detection channels, avoid data processing blockage and resource competition. The number of CUDA streams is determined according to the number of detection channels, and is usually set to an integer multiple of the number of detection channels to improve GPU utilization. Two layers of sub-thread grids are started in each CUDA stream. The first layer of sub-thread grid is responsible for the inference calculation of the fast shallow model, and the second layer of sub-thread grid is responsible for the accurate analysis of the deep model. If the number of detection channels exceeds the limit of the number of CUDA streams of the GPU, a multi-batch processing strategy is adopted, and the detection channels are processed in groups. After each group is processed, the CUDA stream resource is released for use by the next group. The scheduling of CUDA stream adopts an asynchronous execution mode, and different streams can be executed in parallel to improve the overall computing efficiency. The memory management adopts the unified memory technology, and the CPU and the GPU share the memory space, reducing the data transmission overhead.

[0047] The specific implementation of step S09 is that each CUDA stream will input the fault feature data allocated to the fast shallow model of the charging pile fault identification model for preliminary fault scoring and judgment. The fast shallow model is a lightweight neural network deployed on the GPU side, which includes 2 layers of fully connected layers and ReLU activation functions, the number of network parameters is about 10,000, and the inference time is less than 1 ms. The model input is a 128-dimensional fault feature vector, which is transformed into a 64-dimensional hidden layer representation by the first layer of fully connected layer, and then output a 1-dimensional fault score by the second layer of fully connected layer. The ReLU activation function provides nonlinear transformation ability and enhances the expression ability of the model. The preliminary fault score ranges from 0 to 1, indicating the possibility of fault existence. The suspicious threshold is set to 0.3, and the feature data below this threshold is considered to be in a normal state and does not need to be further analyzed in depth. The role of the fast shallow model is to quickly screen a large number of input data, filter out obviously normal data, reduce the computational burden of the subsequent deep model, and improve the overall processing efficiency.

[0048] The specific implementation of step S10 is that the fault feature data with a preliminary score below the suspicious threshold is directly returned to the CPU side, and the data above the suspicious threshold is extracted using the Transformer attention mechanism to extract relevant context fragments. The data below the suspicious threshold is determined to be in a normal state, and the first layer of sub-thread grid directly transmits the result back to the CPU side and enters the subsequent record archiving process. The data above the suspicious threshold needs to be further analyzed in depth, and the Transformer attention mechanism automatically extracts the historical information fragments highly related to the current fault feature data from the charging pile historical data. The attention mechanism calculates the similarity between the current fault feature vector and each time segment in the historical data, and selects the top K similar segments as the context information. The similarity calculation uses cosine similarity or Euclidean distance measurement, and the value of K is usually set to 5-10. The context fragments include historical fault feature data, processing results, environmental parameters and other information, providing rich background knowledge for the deep model. The calculation of attention weight considers the time decay factor, and the newer historical data has a higher weight, and the time decay coefficient is set to 0.95.

[0049] The specific implementation of step S11 is to send the extracted context segment together with the fault feature data into the deep model of the charging pile fault recognition model. The second layer of sub-thread grid performs deep analysis on the complete sequence to output accurate fault scores. The deep model includes 8 layers of Transformer encoder and multi-head attention mechanism, and the number of network parameters is about 1 million, and the inference time is 5-10 ms. The model input is the splicing sequence of fault feature data and context segment, and the total length is about 1000 tokens. Each layer of Transformer encoder includes a multi-head self-attention module and a feedforward neural network module, which improves the training stability through residual connection and layer normalization technology. The number of heads of the multi-head attention mechanism is dynamically adjusted according to the data complexity, and is usually set to 8-16 heads. The deep model captures complex fault patterns and trends by modeling the dependency between different positions in the sequence. The model output is an accurate fault score, with a value range of 0-1, reflecting the confidence of the existence of the fault. The deep model uses mixed precision training technology, and some calculations use 16-bit floating point numbers to improve the calculation speed, and key calculations use 32-bit floating point numbers to ensure accuracy.

[0050] The specific implementation of step S12 is that the second layer of sub-thread grid returns the deep score result to the CPU side, and the CPU determines whether to trigger the fault diagnosis mechanism according to the decision threshold. The decision threshold is set to 0.7, and the deep score higher than this threshold is considered to have an exact fault, which needs to trigger the subsequent diagnosis and processing process. After receiving the deep score result, the CPU first performs threshold comparison. If the score is higher than the decision threshold, it is determined that there is a fault, and the diagnosis mechanism is immediately started. The diagnosis mechanism includes fault type identification, severity assessment, location positioning, impact analysis and other functional modules. The fault feature data is recorded and archived in the fault database, including timestamp, fault feature vector, score result, context information and complete information. At the same time, fault warning information is generated and sent to the operation and maintenance personnel. The warning information includes fault type, occurrence time, severity, recommended treatment measures and other contents. The sending methods of the warning information include short message notification, email notification, mobile application push and other channels to ensure timely response. If the deep score is lower than the decision threshold, it is considered to be a normal state or a slight abnormality, and enters the normal data archiving process.

[0051] The specific implementation of step S13 is that the CPU records the fault feature data not determined as a fault and appends it to the charging pile historical data, ensures that complete context information can be obtained during subsequent detection, and generates a charging pile diagnosis report according to all processing results. Although the data not determined as a fault does not need to be processed immediately, it still has important historical value and can provide data support for subsequent fault prediction and trend analysis. These data are recorded into the historical database in a time series storage format, facilitating quick retrieval and analysis. The storage of historical data adopts data compression technology to reduce storage space occupation, with a compression rate of about 80%. The database is regularly cleaned and maintained, and historical data exceeding the retention period, usually set to one year, is deleted. The generation of the diagnosis report is based on all processing results within the current detection period, including normal data statistics, abnormal event records, fault diagnosis results, and device health status evaluation.

[0052] Need to be explained in detail, the charging pile fault recognition model adopts a multi-modal fusion network based on the Transformer architecture, and the overall structure includes five main parts: input encoding layer, multi-modal fusion layer, attention calculation layer, feature extraction layer, and classification decision layer. The input encoding layer is responsible for converting different types of raw data into a unified vector representation. The electrical parameter encoder adopts a multi-layer perceptron structure, including 3 fully connected layers with 256, 128, and 64 neurons respectively, and uses the ReLU function as the activation function to process continuous numerical data such as voltage, current, and power. The temperature distribution encoder adopts a one-dimensional convolutional neural network structure, including 2 convolutional layers with convolution kernel sizes of 3 and 5 and channel numbers of 32 and 64 respectively, and uses maximum pooling for spatial distribution feature extraction of temperature data. The communication state encoder adopts an embedding layer structure to map discrete communication states to continuous vector representations with an embedding dimension of 32.

[0053] The multi-modal fusion layer adopts an attention mechanism to effectively fuse different modal data and calculate the contribution weight of each modal data to the final decision. The calculation of attention weights is based on the query, key-value, and numerical triple mechanism, and the correlation between different modalities is calculated through the scaled dot-product attention formula. The number of multi-head attention heads is dynamically adjusted according to three parameters: the number of fault types, the data dimension, and the model complexity. When the number of fault types is small and the data dimension is low, 4 to 8 attention heads are used, and when the number of fault types is large and the data dimension is high, 16 to 32 attention heads are used. The attention calculation layer includes a self-attention module and a cross-attention module. The self-attention module calculates the correlation between different features within the same modality, and the cross-attention module calculates the interaction between different modalities.

[0054] The feature extraction layer adopts a multi-layer Transformer encoder structure, each encoder contains a multi-head self-attention sublayer and a feedforward neural network sublayer, and the training stability is ensured through residual connection and layer normalization technology. The number of encoder layers is dynamically adjusted according to the complexity of the data, 4 layers of encoder are used for simple fault scenarios, and 8 to 12 layers of encoder are used for complex fault scenarios. The middle layer dimension of the feedforward neural network is set to 4 times the input dimension, and the GELU function is used as the activation function to improve the model expression ability. The classification decision layer contains a fully connected layer and a softmax activation function, which outputs the probability distribution of each fault type, and the output dimension of the fully connected layer is equal to the number of fault types.

[0055] The fast shallow model, as a preliminary filter on the GPU side, uses a simplified network structure to achieve millisecond-level inference speed. The model contains 2 layers of fully connected layers, the first layer has 64 neurons, and the second layer has 32 neurons, and finally a single neuron output layer produces a fault suspiciousness score between 0 and 1. The activation function uses the ReLU function, and the loss function uses the binary cross-entropy loss. The deep model, as an accurate analyzer on the CPU side, uses a complete Transformer architecture to achieve high-precision fault identification. The model has about 1 million parameters, including 8 layers of Transformer encoders, each layer containing 512-dimensional hidden states and 8 attention heads.

[0056] The training data set establishment process includes four main steps of data collection, data labeling, data enhancement, and data division. In the data collection stage, running data is collected from charging pile equipment of different brands and models, covering domestic mainstream manufacturers such as Telcom, Starstar Charge, and Pureneng New Energy. The data types include normal operation data and various fault data, normal operation data is obtained by long-term monitoring of healthy equipment, and fault data is obtained by fault simulation and actual fault records. In the data labeling stage, experienced equipment engineers manually label the collected data, including fault type, fault severity, fault location, fault cause, and other information. The labeling quality is guaranteed through multiple labeling consistency verification, with a consistency threshold of 90% or higher.

[0057] Optionally, in the data enhancement stage, the number of training samples is expanded through various technical means, including noise injection, time warping, amplitude scaling, and frequency domain transformation. Noise injection adds Gaussian white noise to the original signal to simulate interference in actual measurement, with noise intensity controlled within 5% of the signal amplitude. Time warping performs nonlinear time transformation on the time series to simulate sampling frequency changes, with warping amplitude controlled within 10% of the original time length. Amplitude scaling randomly scales the signal amplitude to simulate sensor drift, with scaling factors ranging from 0.9 to 1.1. Frequency domain transformation converts time domain signals to frequency domain signals through fast Fourier transform, increasing frequency domain feature samples.

[0058] Optionally, the data division stage divides the data set into training set, validation set and test set using stratified sampling method, with the proportions being 70%, 15% and 15% respectively. Stratified sampling ensures that the distribution ratio of each fault type in different data sets remains consistent, avoiding the influence of data skew on model performance. The training set is used for model parameter learning, the validation set is used for hyperparameter tuning and model selection, and the test set is used for final performance evaluation. The representativeness of the data set is verified through statistical analysis, including the sample number distribution of each fault type, data quality indicators, time span coverage, etc. The generalization ability is evaluated by cross-validation method, and 5-fold cross-validation is used to ensure the stable performance of the model on different data subsets.

[0059] Optionally, the model training process uses the Adam optimizer for gradient descent training, with the learning rate set to 0.001, the momentum parameter β1 set to 0.9, and β2 set to 0.999. The batch size is set to 32, the number of training rounds is set to 200 rounds, and the early stopping mechanism is used to prevent overfitting according to the performance of the validation set, with the early stopping patience value set to 20 rounds. The loss function uses weighted cross-entropy loss to handle the class imbalance problem, and the weights are calculated according to the inverse of the number of samples in each class. The learning rate decay strategy uses the cosine annealing algorithm, and the initial learning rate is gradually decayed to a minimum value during the training process, with the minimum learning rate set to 1% of the initial learning rate. Regularization techniques include Dropout and weight decay, with the Dropout probability set to 0.1 and the weight decay coefficient set to .

[0060] Further details are needed to explain that the CPU and GPU cooperative computing architecture achieves efficient fault diagnosis computation through heterogeneous parallel processing. CPU, as the main controller, is responsible for task scheduling, data preprocessing, deep model inference and result comprehensive analysis, and has strong logical control ability and complex branch processing ability. GPU, as a coprocessor, is responsible for parallel inference computation of fast shallow models, and uses its thousands of computing cores to realize large-scale parallel data processing.

[0061] The cooperative working mechanism adopts a pipeline processing mode. CPU first performs data preprocessing and feature extraction, and distributes the processed data to multiple CUDA streams in GPU for parallel computation. Each CUDA stream independently processes the data of one detection channel and performs preliminary fault scoring through a fast shallow model. After GPU computation, the results are returned to CPU, which decides whether further deep model analysis is needed based on the scoring results. For data that needs deep analysis, CPU calls the Transformer attention mechanism to extract historical context information, and then performs accurate fault identification through a deep model.

[0062] Memory management adopts unified memory technology, CPU and GPU share the same physical memory space, avoiding frequent data copy overhead. Asynchronous computing scheduling enables CPU to handle other tasks while GPU performs computing, improving overall computing efficiency. Load balancing algorithm dynamically allocates computing resources according to the data complexity of each detection channel, ensuring maximum GPU utilization.

[0063] The synergistic computing effect is reflected in the improvement of processing speed. The parallel processing architecture greatly improves the processing speed of multi-channel data compared to traditional serial processing. In terms of computing precision, the layered processing strategy ensures high-precision deep analysis while reducing computational burden through rapid screening. In terms of resource utilization optimization, CPU focuses on complex logic processing, and GPU focuses on parallel computing, each playing to its strengths to achieve optimal resource allocation.

[0064] Optionally, the fault type classification is classified and the contribution rate is calculated by statistical analysis and weight allocation algorithm to identify the charging pile fault type, and a data-driven fault diagnosis priority system is established. The fault type classification process first establishes a fault sample statistical database, collects all fault event records in the operation process of the charging pile, including fault occurrence time, fault phenomenon description, fault cause analysis, treatment method record, repair time statistics and other complete information.

[0065] The electrical fault type identification is realized through voltage parameter and current parameter waveform analysis, and the calculation process includes voltage fluctuation coefficient analysis, current harmonic distortion rate calculation and power factor deviation evaluation. The voltage fluctuation coefficient calculation formula is , wherein represents the voltage value of the i-th sampling point, represents the average voltage value, and N represents the total number of sampling points. The current harmonic distortion rate calculation uses fast Fourier transform to obtain frequency domain components, and the total harmonic distortion rate calculation formula is , wherein represents the effective value of the n-th harmonic current, represents the effective value of the fundamental wave current. The power factor change rate is calculated by comparing adjacent time windows, and the formula is .

[0066] Mechanical fault type identification is realized through temperature distribution data and vibration monitoring. Temperature anomaly judgment adopts temperature rise rate and absolute temperature double threshold. The temperature rise rate calculation formula is , wherein represents the temperature value at time t, represents the time interval. Vibration monitoring is performed by frequency domain analysis of acceleration sensor data to extract vibration frequency spectrum features for mechanical fault identification.

[0067] Communication fault type identification is realized through communication protocol state analysis and response time analysis. The packet loss rate calculation formula is , wherein represents the number of sent frames, represents the number of received frames. Communication delay statistics use round-trip time measurement to calculate average delay, maximum delay, and delay jitter parameters.

[0068] Environmental fault type identification is realized through temperature distribution data and environmental monitoring. Environmental temperature anomalies are determined by comparing with historical data, and humidity anomalies are evaluated by dew point temperature calculation.

[0069] Fault type proportion calculation is based on large sample statistical analysis. A total of 10,000 charging pile fault data samples covering different brands, models, and use environments are collected. The number of electrical fault samples is 4,500, accounting for 45% of the total fault proportion. The calculation formula is . The number of mechanical fault samples is 2,500, accounting for 25% of the total fault proportion. The number of communication fault samples is 2,000, accounting for 20% of the total fault proportion. The number of environmental fault samples is 1,000, accounting for 10% of the total fault proportion.

[0070] Fault type contribution rate calculation uses a weighted scoring method, considering fault frequency, repair difficulty, impact range, and economic loss. The electrical fault type contribution rate calculation process is as follows: fault frequency weight , repair difficulty weight , impact range weight , and economic loss weight . The electrical fault dimension scores are frequency score , repair difficulty score , impact range score , and economic loss score . The electrical fault contribution rate calculation formula is , rounded to 0.8.

[0071] The mechanical fault type contribution rate calculation process is as follows: frequency score , repair difficulty score , impact range score , and economic loss score , resulting in . The communication fault type contribution rate calculation process is as follows: frequency score , repair difficulty score , impact range score , and economic loss score , resulting in , rounded to 0.4. The environmental fault type contribution rate calculation process is as follows: frequency score , repair difficulty score , impact range score , economic loss score , calculated , rounded to 0.2.

[0072] The fault diagnosis resource allocation strategy is optimized based on the product of the contribution rate and the proportion, the electrical fault resource allocation weight is , the mechanical fault resource allocation weight is , the communication fault resource allocation weight is , the environmental fault resource allocation weight is . The sampling frequency distribution adopts weight proportion adjustment, the electrical fault detection sampling frequency is set to 120% of the reference frequency, the mechanical fault detection sampling frequency is set to 100% of the reference frequency, the communication fault detection sampling frequency is set to 80% of the reference frequency, and the environmental fault detection sampling frequency is set to 60% of the reference frequency.

[0073] It should be noted that the key technical ideas of the present application mainly reflect in two aspects of hierarchical fault identification and GPU parallel computing.

[0074] The hierarchical fault identification technology adopts a two-level processing architecture of fast shallow model and deep model, which greatly improves the processing efficiency while ensuring the diagnosis accuracy. Traditional methods usually use a single model to process all data, resulting in waste of computing resources and long response time. The present application preliminarily screens a large amount of input data through a fast shallow model, and only analyzes suspicious data in depth, effectively reducing unnecessary computing overhead. This layered processing strategy enables the system to complete the processing of most data within milliseconds, while accurately diagnosing abnormal conditions that really need attention.

[0075] The GPU parallel computing technology realizes efficient parallel processing of multi-channel data through CUDA stream and multi-layer sub-thread grid, breaking through the performance bottleneck of traditional CPU serial processing. Compared with the traditional sequential processing method, the parallel computing architecture can process the data of multiple detection channels at the same time, and the processing speed is improved by tens of times. Especially in the multi-pile simultaneous detection scene of large-scale charging station, the advantage of parallel computing technology is more obvious, which can significantly shorten the overall detection time and improve the utilization rate of equipment.

[0076] The synergistic effect of these two key technical ideas produces significant systematic advantages. GPU parallel computing provides strong computing support for multi-modal data fusion and deep model reasoning, making real-time application of complex algorithms possible.

[0077] Specifically, the principle of the present application is that the present application can solve the technical problems of low real-time processing efficiency and insufficient fault diagnosis accuracy of multi-channel data, and the fundamental principle lies in the deep integration of GPU parallel computing architecture and intelligent layered diagnosis mechanism. First, a sufficient number of CUDA streams are started to match the number of detection channels, and each CUDA stream independently processes the data of one detection channel, realizing real parallel data processing and eliminating the time bottleneck of traditional sequential processing. Secondly, the present application designs a layered architecture of fast shallow model and deep model, the fast shallow model is deployed on the GPU end for preliminary fault scoring, and through the simplified network structure, the fast inference of milliseconds is realized, effectively filtering out a large amount of normal data, only the data whose preliminary score is higher than the suspicious threshold will enter the deep model for accurate analysis, this layered processing strategy greatly reduces unnecessary deep calculation and improves the overall processing efficiency. Thirdly, the present application adopts a multi-modal fusion network based on the Transformer architecture, which automatically extracts the historical context fragments highly related to the current fault feature data through the multi-head attention mechanism, realizes accurate recognition of complex fault patterns, and the number of attention heads is dynamically adjusted according to the number of fault types, data dimensions and model complexity, ensuring the adaptability of the model in different complexity scenarios. Finally, the sliding time window method is used to analyze the segmented data, extract key feature parameters and construct a fault correlation matrix, and the minimum spanning tree algorithm is used to optimize the fault propagation path analysis, realizing accurate positioning of the fault root cause. This multi-level and multi-dimensional technical architecture design makes the present application ensure high processing efficiency while significantly improving the accuracy and reliability of fault diagnosis.

[0078] A specific embodiment 1 of the present application is provided below, and the specific implementation of each step in the embodiment 1 is described in detail as follows.

[0079] In this embodiment, the detection host uses Advantech industrial computer model IPC-610L, which is configured with Intel Core i7-8700 processor, 16GB memory, 500GB solid state disk, supports ATX motherboard or up to 15 slots PICMG passive backplane, and provides rich interfaces for connecting each module. The multi-channel data acquisition module uses Altai PCIe2313 series, which includes 16 independent analog signal acquisition channels, each channel is configured with 24-bit high-precision analog-to-digital converter model AD7779, and programmable gain amplifier uses Linglierte LTC6915, the sampling frequency reaches 10kHz, and the measurement accuracy is better than 0.1%.

[0080] The power quality analysis module adopts the GDE8000 series of Guodian Xigao, with a built-in fast Fourier transform processor model TMS320C6678, supporting real-time calculation of harmonic components of voltage parameters and current parameters, with an analysis frequency range covering DC to 5 kHz, and a harmonic analysis accuracy of 0.01%. The insulation resistance test module adopts the HY2671 series of Huayi Power, with an adjustable output voltage DC high voltage source, an output voltage range of 500V to 5000V, a test range covering 1MΩ to 10GΩ, and a built-in leakage current compensation circuit to ensure test accuracy.

[0081] The ground resistance test module adopts a clamp-type ground resistance tester model Shengli VC4105A, with a three-wire test principle, an AC signal generator frequency of 128Hz, a phase-sensitive detector using lock-in amplification technology, and a test range of 0.01Ω to 2000Ω. The communication protocol analysis module adopts the CANalyst-II of Ciyuan Electronics, supporting CAN2.0A protocol, CAN2.0B protocol, CANopen protocol, and J1939 protocol, with an optical isolation interface model 6N137 to ensure test safety.

[0082] The temperature monitoring module is configured with 32 wireless temperature sensors, using the RY-WS03 model of Shanghai Ruyu Electronics, with a built-in 2.4GHz wireless communication module model nRF24L01 in each sensor, a temperature measurement range of -40℃ to 150℃, a temperature measurement accuracy of ±0.5℃, and a data transmission distance of up to 100m. The fault simulation module is configured with a programmable electronic load, using the IT8800 series of Aideks, with a power range of 0 to 100kW, supporting four working modes of constant current, constant voltage, constant power, and constant resistance, for simulating overload faults, short circuit faults, and open circuit faults. The control chip uses the HiSilicon 9000E processor, integrating the ARM Cortex-A77 architecture, with a main frequency of up to 3.13GHz, supporting GPU parallel computing and AI acceleration functions.

[0083] The specific implementation of the charging pile intelligent diagnosis module execution step is described below.

[0084] In this embodiment, the specific implementation of step S01 is the same as described above, and will not be described in detail here.

[0085] The specific implementation of step S02 is to perform preprocessing operations on the collected raw data, including data cleaning, filtering and noise reduction, data normalization, and outlier detection. The 3σ criterion is used for outlier detection during data cleaning, with the determination condition being , where is the measured value of the ith data point, is the mean value of the data sequence, is the standard deviation of the data sequence. The filter denoising process uses a 4th order Butterworth low-pass filter, whose transfer function is , where, is the filter gain, is the filter coefficient, is the Laplace transform complex variable. The data normalization process uses the maximum-minimum value normalization method, and the normalization formula is , where, is the normalized value of the ith data point, is the original value of the ith data point, is the minimum value of the data sequence, is the maximum value of the data sequence.

[0086] The specific implementation of step S03 is to use the sliding time window method to analyze the preprocessed data in segments and extract key feature parameters. The voltage fluctuation coefficient calculation formula is , where, is the voltage fluctuation coefficient, is the voltage value of the ith sampling point, is the average voltage value within the time window, is the total number of sampling points within the time window. The current harmonic distortion rate calculation formula is , where, is the total harmonic distortion rate of the current, is the effective value of the nth harmonic current, is the effective value of the fundamental current, is the highest harmonic number for analysis, usually 50. The power factor change rate calculation formula is , where, is the power factor change rate, is the power factor of the current time window, is the power factor of the next time window. The temperature rise rate calculation formula is , where, is the temperature rise rate, is the temperature value at time t, is the time interval, taking 5 seconds. The communication packet loss rate calculation formula is , where, is the communication packet loss rate, is the number of sent frames, is the number of received frames. The insulation resistance drop rate calculation formula is , where, is the insulation resistance drop rate, is the reference insulation resistance value, is the current insulation resistance value.

[0087] The specific implementation of step S04 is to identify abnormal feature points in the key feature parameters according to the pre-set fault feature threshold. The abnormality determination condition adopts a multivariate threshold comparison method, and the comprehensive determination function is , wherein is an abnormality determination index, is the total number of feature parameters, is the weight coefficient of the jth feature parameter, is the measured value of the jth feature parameter, is the threshold value of the jth feature parameter, is the abnormality determination function of the jth feature parameter, and is defined as . When , it is determined to be abnormal, wherein is an abnormality determination threshold, and the value range is 0.6-0.8.

[0088] The specific implementation of step S05 is the same as the foregoing, and will not be described in detail here. The attention head number calculation formula is , wherein is the number of attention heads, is the number of fault types, is the data dimension, is a model complexity parameter, and the value range is 1-10. The comprehensive evaluation value calculation formula is , wherein is the comprehensive evaluation value, are weight coefficients of fault complexity, data quality, and computing resource, respectively, and the values are 0.4, 0.3, and 0.3, respectively, is a fault complexity score, which is calculated based on the number of fault types and the feature dimension, is a data quality score, which is evaluated based on data integrity and noise level, is a computing resource score, which is calculated based on CPU and GPU utilization, and the value range is 0-1.

[0089] The specific implementation of step S06 is to construct a fault correlation matrix based on the fault type identification and the fault location identification, and to optimize the fault propagation path analysis by using a minimum spanning tree algorithm. The fault correlation matrix is represented as , wherein is the fault correlation matrix, is the correlation strength between fault type i and fault type j, is the total number of fault types. The correlation strength calculation formula is , wherein is the conditional probability of fault j occurring under the condition of fault i occurring, which is obtained by statistical analysis of historical fault data, is the physical correlation coefficient between fault i and fault j, which is calculated by analyzing the circuit topology, and its value ranges from 0 to 1, is the time delay from fault i to fault j, with the unit of seconds, which is determined by analyzing the fault propagation path, is the time decay constant, with a value of 3600 seconds, is the weight coefficient and satisfies , and the values are 0.5, 0.3, and 0.2, respectively. The Prim algorithm is used for the minimum spanning tree construction, and the edge weight is defined as , which ensures that the edge weight between faults with greater correlation strength is smaller.

[0090] The specific implementation of step S07 is to calculate the optimal detection parameter configuration through the charging pile parameter optimization function. The optimization objective function is , where is the optimization objective function value, is the optimization objective quantity, is the weight coefficient of the kth optimization objective, is the kth optimization objective function, is the detection parameter to be optimized, is the total number of parameters. The constraint condition is , where and are the minimum and maximum values of the ith parameter, respectively. The optimization algorithm uses the genetic algorithm, and the fitness function is , where is the fitness value.

[0091] The specific implementation of steps S08-S09 is the same as described above, and will not be described in detail here.

[0092] The specific implementation of step S10 is to classify the preliminary scores and extract relevant context segments. The attention weight calculation formula is , where is the attention weight of the current fault feature vector and the jth historical segment, is the attention energy function value, is the total number of historical segments. The attention energy function is , where is the learnable parameter matrix, is the hidden state vector of the jth historical segment, is the current query vector, is the bias vector. The time decay factor is , where is the time decay factor of the jth historical segment, is the decay coefficient with a value of 0.95, is the current time, is the time of the jth historical segment.

[0093] The specific implementation of step S11 is the same as the foregoing, and will not be described in detail here.

[0094] The specific implementation of step S12 is to compare the deep score result with a judgment threshold. The fault judgment condition is , wherein is the fault score output by the deep model, is the fault judgment threshold, and is 0.7. The confidence calculation formula is , wherein is the fault judgment confidence, is the score standard deviation, is the score mean.

[0095] The specific implementation of step S13 is the same as the foregoing, and will not be described in detail here.

[0096] It should be noted that the fault type proportion calculation adopts the statistical frequency method, and the electrical fault proportion is , wherein is the electrical fault proportion, is the number of electrical fault samples, is the total number of fault samples. The fault type contribution rate calculation adopts the weighted scoring method, and the electrical fault contribution rate is , wherein is the electrical fault contribution rate, are respectively the weight coefficients of the fault frequency, the repair difficulty, the influence range, and the economic loss, and are respectively 0.3, 0.25, 0.25, and 0.2, is the score corresponding to the dimension, which is determined through expert evaluation, is the fault frequency score, which is obtained based on historical statistical data after standardization, is the repair difficulty score, which is evaluated based on the average repair time and the technical complexity, is the influence range score, which is calculated based on the number of devices affected by the fault and the number of users, is the economic loss score, which is evaluated based on the direct repair cost and the indirect loss, and is respectively 0.9, 0.8, 0.85, and 0.75. The resource allocation weight calculation formula is , wherein is the resource allocation weight of a fault type, is the occurrence proportion of the fault type, is the contribution rate of the fault type.

[0097] It should be noted that the voltage fluctuation coefficient formula is Based on the statistical variance principle, the voltage stability is quantified by calculating the degree of voltage deviation from the mean value. Compared with the traditional peak detection method, this formula can more comprehensively reflect the overall characteristics of voltage fluctuations, effectively identify intermittent voltage disturbances, and improve the sensitivity and accuracy of electrical fault detection.

[0098] Abnormality determination function Based on the relative deviation quantification principle, the severity of abnormality is evaluated by calculating the relative degree of measured values exceeding the threshold. Compared with the simple binary decision method, this function provides continuous abnormality degree quantification, can distinguish abnormal states of different severity, and provides more detailed judgment basis for fault diagnosis, significantly improving the accuracy and reliability of abnormality detection.

[0099] Current harmonic distortion rate formula Based on the Fourier frequency domain analysis principle, the current quality is evaluated by quantifying the ratio of high-order harmonics to fundamental wave. Compared with simple amplitude analysis, this formula can identify the nonlinear working state of power devices and provide accurate quantitative indicators for power module fault diagnosis, significantly improving the health state evaluation ability of electrical systems.

[0100] Insulation resistance decline rate formula Based on the relative change quantification principle, the insulation performance degradation is evaluated by calculating the decline amplitude of current insulation resistance relative to the reference value. Compared with the absolute threshold determination method, this formula can adapt to the reference differences of different devices and provide a more accurate insulation fault warning mechanism to effectively prevent insulation breakdown accidents.

[0101] Fault correlation matrix formula Considering the statistical correlation, physical correlation and time correlation in three dimensions, compared with the traditional single correlation analysis method, this formula constructs a more complete fault propagation network, which can accurately identify the root cause and propagation path of the fault, provides a scientific basis for fault handling priority determination, and effectively reduces the fault handling time and maintenance cost.

[0102] Parameter optimization objective function Using multi-objective weighted optimization principle, the importance between different optimization objectives is balanced by weight coefficients. Compared with the fixed parameter configuration scheme, this function realizes the dynamic adaptive adjustment of detection parameters, optimizes the detection strategy according to the actual fault characteristics and environmental conditions, significantly improves the detection efficiency and resource utilization, and reduces the false positive rate and false negative rate.

[0103] Attention weight calculation formula Based on the principle of soft attention mechanism, the attention weight is allocated by calculating the correlation between the current feature and the historical information. Compared with the traditional fixed window history analysis method, this mechanism can automatically select the most relevant historical information for context modeling, improving the accuracy and robustness of fault identification, especially when dealing with complex fault patterns.

[0104] Attention head number formula Based on the principle of adaptive computing complexity allocation, the number of attention heads is dynamically determined by considering the number of fault types, data dimensions and model complexity. Compared with the fixed head configuration method, this formula realizes intelligent allocation of computing resources, optimizes computing efficiency while ensuring model expression ability, and significantly improves the adaptive performance of the fault identification system.

[0105] In order to better understand and implement the present application, the following provides an embodiment 2 of a specific application scenario of the present application: The technical team first builds a specific system platform according to the hardware device model parameters in embodiment 1. After the system starts, the technical team performs a comprehensive parameter scan on the first 60kW charging pile. The multi-channel data acquisition module simultaneously acquires the voltage, current and power parameters of the charging pile. The obtained basic data shows that the output voltage is 750V, the output current is 80A, and the power factor is 0.98. The temperature monitoring module monitors the temperature distribution of 12 key positions inside the charging pile in real time. The data shows that the power module temperature is 65℃, the transformer temperature is 58℃, and the radiator temperature is 52℃. The insulation resistance test result shows that the positive electrode-to-ground insulation resistance is Ω, and the negative electrode-to-ground insulation resistance is Ω. The grounding resistance test result shows that the grounding resistance of the charging pile is 0.8Ω. The communication protocol analysis module shows that the CAN bus communication is normal, and the data packet transmission success rate is 99.8%. Figure 4 The multi-channel data acquisition module shows the real-time monitoring results of the voltage, current and power parameters of the 60kW charging pile within 30 seconds. The blue curve in the figure represents the voltage parameter, which shows that the output voltage fluctuates around the 750V reference value with a fluctuation amplitude of about ±5V, which meets the description of the voltage fluctuation coefficient 0.012 in the embodiment. The red curve represents the current parameter, which shows that the output current changes around the 80A reference value with a change amplitude of about ±8A. The green curve represents the power parameter, which is calculated by the product of voltage and current, and the power value fluctuates around 60kW. Figure 5The temperature distribution of 32 wireless temperature sensors in normal operation state and fault state is shown. The green column chart represents the normal operation temperature, and the red column chart represents the fault state temperature. The horizontal coordinate is the sensor position (every 4 positions show a label), and the vertical coordinate is the temperature value. The orange dotted line in the figure represents the temperature warning threshold of 70°C, and the red dotted line represents the temperature danger threshold of 85°C. It can be seen that the temperature of key parts such as power modules, transformers, and heat sinks in the fault state increases significantly, especially the temperature of the power module increases from 65°C to 78°C, which exceeds the warning threshold.

[0106] In the data preprocessing stage, the system processes the collected raw data, detects outliers using the 3σ criterion, and eliminates 0.3% of the abnormal data points. The filter denoising process uses a 4th order Butterworth low-pass filter with a cutoff frequency of 2.5kHz, effectively filtering out high-frequency noise. The data normalization process uses the maximum and minimum value normalization method to map all types of parameters to the interval of 0 to 1.

[0107] The system uses the sliding time window method to analyze the preprocessed data in segments, with a time window length of 5s and a window overlap rate of 50%. Key feature parameters are extracted from each time window, with a voltage fluctuation coefficient of 0.012, a current harmonic distortion rate of 3.2%, a power factor change rate of 0.008, a temperature rise rate of 0.5°C / min, an insulation resistance drop rate of 0.002, and a communication packet loss rate of 0.2%.

[0108] According to the preset fault feature threshold, the system makes an abnormality determination on the key feature parameters. The voltage fluctuation coefficient threshold is set to 0.02, the current harmonic distortion rate threshold is set to 5%, the power factor change rate threshold is set to 0.01, the temperature rise rate threshold is set to 2°C / min, the insulation resistance drop rate threshold is set to 0.1, and the communication packet loss rate threshold is set to 1%. The abnormality determination index is calculated by the multivariate threshold comparison method, with weight coefficients of 0.25, 0.20, 0.15, 0.20, 0.15, and 0.05. The abnormality determination index is 0.15, which is lower than the abnormality determination threshold of 0.6, indicating that the charging pile is in normal operation state.

[0109] The technical team continues to detect the second 120kW charging pile, which has been running for 3 years, and there are potential hidden troubles. The test results show that the output voltage is 745V, there is a slight fluctuation, and the voltage fluctuation coefficient is 0.025, which exceeds the set threshold. The current harmonic distortion rate is 7.8%, which exceeds the set threshold, indicating that the power device is in a nonlinear working state. The power factor is 0.92, which is lower than the standard value of 0.98. Temperature monitoring data shows that the power module temperature is 78℃, which exceeds the normal working temperature range. The insulation resistance test shows that the positive electrode to ground insulation resistance is Ω, which is much lower than the reference value. Communication protocol analysis shows that the data packet transmission success rate is 96.5%, there is an intermittent communication failure. The fault feature data is shown in Table 1.

[0110] Table 1 Charging pile fault feature parameter statistics table

[0111] The system inputs the classified abnormal feature points into the charging pile fault recognition model, which is based on the Transformer architecture and contains electrical parameter encoder, temperature distribution encoder and communication state encoder. The number of attention heads of the multi-head attention mechanism is dynamically adjusted according to the fault complexity, which is set to 12 attention heads in the current fault scenario. The number of fault types is set to 4, the data dimension is 128, and the model complexity parameter is set to 8.

[0112] The system starts 8 CUDA streams for parallel processing, each CUDA stream processes the data of 2 detection channels, Figure 6 The CUDA stream processing time comparison analysis of GPU parallel processing architecture is given. The fast shallow model first performs preliminary fault scoring, and the suspicious threshold is set to 0.4 and the judgment threshold is set to 0.7. For fault feature data with preliminary score higher than the suspicious threshold, the system uses the Transformer attention mechanism to extract relevant context fragments from the charging pile historical data, and the extracted historical fragments contain similar fault feature data in the past 30 days.

[0113] The deep model analyzes the complete sequence in depth, and the output fault score is 0.78, which is higher than the judgment threshold 0.7, and the system determines that there is a fault. The fault type identification result shows that the fault is an electrical fault, the fault severity level is medium, and the fault location is identified as the power module and the communication module. The confidence calculation result is 0.85, indicating that the diagnosis result is reliable.

[0114] The system constructs a fault correlation matrix to analyze the correlation between different fault types. The correlation strength between electrical faults and communication faults is 0.65, indicating that power device abnormalities may cause communication interference. The fault propagation path is analyzed by the minimum spanning tree algorithm, and it is determined that the power module fault is the main fault source and the communication fault is the secondary fault source. The fault correlation strength analysis results are shown in Table 2.

[0115] Table 2 Fault type correlation strength matrix

[0116] The system calculates the optimal detection parameter configuration through the parameter optimization function, and the input parameters include fault type identification, fault severity level, environmental temperature 32℃, load characteristic parameters, and historical fault frequency statistics. The optimization results show that the sampling frequency is adjusted to 12kHz, the test voltage is adjusted to 1500V, the test current is adjusted to 100A, the temperature threshold is adjusted to 75℃, and the communication timeout time is adjusted to 500ms.

[0117] The technical team detects the third 180kW charging pile, which is a newly installed device, and all parameters are within the normal range. The detection results show that the output voltage is 750V, the voltage fluctuation coefficient is 0.008, the current harmonic distortion rate is 2.1%, the power factor is 0.99, the temperature distribution is uniform, the insulation resistance value is normal, and the communication protocol state is good. The system comprehensive evaluation value calculation result is 0.25, the number of attention heads adjusted by the linear growth function is 6, which is suitable for simple fault scenarios.

[0118] After completing the detection of 40 charging piles, the system generates a comprehensive diagnosis report. The detection results show that 35 charging piles are in normal operation, 3 charging piles have minor faults, and 2 charging piles have moderate faults. The fault type distribution statistics show that electrical faults account for 60%, mechanical faults account for 25%, communication faults account for 10%, and environmental faults account for 5%. The system suggests that the charging piles with faults should be maintained, the aging power devices should be replaced, the heat dissipation system should be optimized, and the communication protocol should be upgraded. The fault distribution statistics are shown in Table 3.

[0119] Table 3 Charging pile fault type distribution statistics

[0120] The technical progress brought by the present application relative to traditional detection methods mainly embodies in the following aspects. The traditional detection method mainly relies on single parameter threshold judgment and cannot comprehensively reflect the overall health status of the charging pile, while the present application adopts multi-source data fusion technology, comprehensively analyzes multi-dimensional parameters such as voltage, current, power, temperature, insulation and communication, captures dynamic change characteristics through a sliding time window method, and realizes the change from static detection to dynamic monitoring. The traditional method mainly relies on expert experience and preset rules in fault identification, and the accuracy and adaptability are limited, the present application constructs an intelligent fault identification model based on the Transformer architecture, automatically learns the complex correlation between fault features through the multi-head attention mechanism, and realizes the change from experience-driven to data-driven. The traditional detection system usually adopts a serial processing method, and the detection efficiency is low, the present application adopts a GPU parallel computing architecture, realizes multi-channel parallel processing through a CUDA stream, and combines a layered processing strategy of a fast shallow model and a deep model, which significantly improves the detection efficiency while ensuring the detection accuracy. The traditional method mainly focuses on single-point faults in fault analysis, and it is difficult to identify the correlation between faults, the present application can accurately identify the fault propagation path and the root cause through fault correlation matrix modeling and minimum spanning tree algorithm analysis, and provide a scientific decision basis for fault handling. The traditional detection parameter configuration is usually fixed and cannot adapt to the detection needs of different fault scenes, the present application realizes the dynamic adaptive adjustment of the detection parameters through a parameter optimization function, optimizes the detection strategy according to the fault type, severity, environmental conditions and other factors, and improves the pertinence and effectiveness of the detection.

[0121] It should be noted that the variables involved in the present application are explained in detail as shown in Tables 4 and 5.

[0122] Table 4 Variable explanation table (first part)

[0123] Table 5 Variable explanation table (second part)

[0124] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A rapid testing, diagnosis, and debugging system for DC charging piles, characterized in that, It includes a detection host, a multi-channel data acquisition module, a power quality analysis module, an insulation resistance testing module, a grounding resistance testing module, a communication protocol parsing module, a temperature monitoring module, a fault simulation module, and a control chip. The control chip is equipped with a charging pile intelligent diagnostic module, which performs the following steps: controlling the multi-channel data acquisition module to perform a comprehensive parameter scan of the charging pile, acquiring voltage parameters, current parameters, power parameters, temperature distribution data, insulation resistance value, grounding resistance value, and communication protocol status, and establishing a charging pile operating status benchmark database; Preprocess the collected data; The sliding time window method is used to perform segmented analysis on the preprocessed data and extract key feature parameters. Identify abnormal feature points and classify and label them according to fault type; Inputting abnormal feature points into the charging pile fault identification model yields fault type identifier, fault severity level, and fault location identifier. Construct a fault correlation matrix and use the minimum spanning tree algorithm to optimize fault propagation path analysis; calculate the optimal detection parameter configuration through the charging pile parameter optimization function; start the CUDA stream, input the fault feature data into the fast shallow model for preliminary fault scoring, use the Transformer attention mechanism to extract relevant context fragments for data above the suspicious threshold, and send them into the deep model for in-depth analysis; generate a charging pile diagnostic report.

2. The system according to claim 1, characterized in that, The multi-channel data acquisition module adopts a modular design, including 16 independent analog signal acquisition channels and 8 digital signal acquisition channels. Each analog signal acquisition channel is equipped with a 24-bit high-precision analog-to-digital converter and a programmable gain amplifier, with a sampling frequency of up to 10kHz.

3. The system according to claim 2, characterized in that, The power quality analysis module has a built-in fast Fourier transform processor for real-time calculation of the harmonic components of voltage and current parameters, with an analysis frequency range covering DC to 5kHz.

4. The system according to claim 3, characterized in that, The insulation resistance test module is equipped with an adjustable DC high voltage source with an output voltage range of 500V to 5000V and a test range covering 1MΩ to 10GΩ.

5. The system according to claim 4, characterized in that, The sliding time window specifically divides the continuous data stream into segments according to a preset time length. Each window contains 5 seconds of data, and the windows overlap by 50%, which is used to capture the dynamic changes in the operating status of the charging pile.

6. The system according to claim 5, characterized in that, The key feature parameters are specifically important indicators that reflect the operating status of the charging pile, extracted from each time window, including voltage fluctuation coefficient, current harmonic distortion rate, power factor change rate, temperature rise rate, insulation resistance decrease rate, and communication packet loss rate.

7. The system according to claim 6, characterized in that, The fault correlation matrix is ​​specifically a two-dimensional matrix that describes the degree of mutual influence between different fault type identifiers, and the matrix element values ​​represent the correlation strength between faults.

8. The system according to claim 7, characterized in that, The CUDA stream is specifically an execution unit for GPU parallel computing, used to implement parallel processing of multiple detection channels. Each CUDA stream independently processes the data of one detection channel, avoiding data processing blockage.

9. The system according to claim 8, characterized in that, The fast shallow model is specifically a lightweight neural network model deployed on a GPU for initial fault screening. It achieves fast inference through a simplified network structure, including a two-layer neural network and a ReLU activation function.

10. The system according to claim 9, characterized in that, The deep model, specifically a complex neural network model, is deployed on the CPU for accurate fault analysis. It improves the accuracy and reliability of fault identification through a deep network structure, including an 8-layer Transformer encoder and a multi-head attention mechanism.

Images