A method for fault detection of a european wall plug

Abstract

The application provides a fault detection method for a European standard wall penetrating terminal, and belongs to the technical field of fault detection. In the application, a partial discharge characteristic component is input into an artificial intelligence model based on a sparse self-encoder to output a partial discharge characteristic vector and calculate an insulation degradation index. Meanwhile, a contact state observer based on a virtual impedance principle is constructed to identify contact resistance parameters online by using a recursive least square method. A double-layer game optimization model is used to dynamically adjust a detection threshold and a sampling strategy, and a power grid fluctuation robustness evaluation algorithm based on topology dynamics is started to identify energy weak points. Finally, a fault confidence adjustment factor is calculated according to the insulation degradation index, the contact resistance parameter and the system stability to realize a graded response, so that the technical problem that the European standard wall penetrating terminal is difficult to realize the cooperative real-time detection of insulation degradation and contact deterioration in a high-power charging environment is solved.

Description

Technical Field

[0001] This invention belongs to the field of fault detection technology, and more specifically, relates to a fault detection method for European standard through-wall terminals. Background Technology

[0002] As a critical contact component in high-power charging systems, European standard through-wall terminals often experience simultaneous and coupled partial discharge phenomena caused by insulation aging and increased contact resistance due to contact interface oxidation under conditions of 400A high current surges and frequent insertions and removals. Traditional detection methods typically employ offline withstand voltage testing to assess insulation performance or periodically measure contact resistance to monitor contact status. However, these methods rely on downtime maintenance and cannot capture dynamic evolution characteristics during charging. In current high-power charging station operations, the strong electromagnetic noise generated by the power converter covers the partial discharge signal frequency band, while grid surge impacts cause a significant increase in contact resistance parameter identification errors. This makes it difficult for existing single-parameter monitoring schemes to accurately distinguish the coupled state of insulation degradation and contact deterioration in complex electromagnetic environments. In other words, existing technologies face the technical challenge of achieving coordinated real-time detection of insulation degradation and contact deterioration in European standard through-wall terminals under high-power charging conditions. Summary of the Invention

[0003] In view of this, the present invention provides a fault detection method for European standard through-wall terminals, which can solve the technical problem in the prior art that it is difficult to achieve real-time detection of insulation degradation and contact deterioration of European standard through-wall terminals under high-power charging environment.

[0004] This invention is implemented as follows: A fault detection method for European standard through-wall terminals is provided. Multi-channel high-frequency voltage sensors and multi-channel high-frequency current sensors are deployed at the contact interface of the European standard through-wall terminal. Simultaneously, the total loop current reference signal output from the shunt is acquired. The acquired high-frequency voltage and current signals are compressed in real-time using a compressed sensing dimension reduction processor to extract sparse feature coefficients. Adaptive multi-scale wavelet packet transform decomposition is performed on the sparse feature coefficients to extract partial discharge feature components. These components are input into a partial discharge feature decoupling model for nonlinear feature extraction, outputting a partial discharge feature vector. The partial discharge feature vector is then compared with the health baseline feature space using a cosine similarity calculation. Insulation degradation index is calculated, and a contact state observer based on the virtual impedance principle is constructed to monitor the dynamic relationship between voltage drop and current in real time. The recursive least squares method is used to identify contact resistance parameters online. The insulation degradation index and contact resistance parameters are input into a two-layer game optimization model to output the optimized detection threshold and the optimized sampling strategy. A grid fluctuation robustness assessment algorithm based on topology dynamics is launched to identify energy weaknesses under grid surge impact and output system stability. The fault confidence adjustment factor is calculated based on the insulation degradation index, contact resistance parameters and system stability to trigger graded response commands. The detection data at each time point is stored in the fault feature database for continuous training of the partial discharge feature decoupling model.

[0005] The compressed sensing dimensionality reduction processor, based on the principle of signal sparsity, projects the high-dimensional original signal into a low-dimensional space through a measurement matrix and recovers the sparse signal from the observations using an orthogonal matching pursuit algorithm.

[0006] The sparse feature coefficient retention rate is controlled between 15% and 25% of the original data volume.

[0007] The adaptive multi-scale wavelet packet transform decomposition is set to 5 decomposition layers to extract partial discharge characteristic components in the frequency band range of 500kHz to 3MHz.

[0008] The partial discharge feature decoupling model is a deep neural network based on a sparse autoencoder, comprising an input layer, a first sparse coding layer, a second sparse coding layer, a third sparse coding layer, a first decoding layer, a second decoding layer, and an output layer.

[0009] The first sparse coding layer contains 64 neurons and is subject to L1 regularization constraints, the second sparse coding layer contains 32 neurons, and the third sparse coding layer contains 16 neurons.

[0010] The sparsity in the sparse autoencoder is achieved through L1 regularization constraints, which force the activation values ​​of most neurons to approach zero, retaining only the activation responses of a few key neurons.

[0011] The health reference feature space is composed of partial discharge feature vectors of a large number of healthy European standard through-wall terminal samples. The maximum value of the cosine similarity between the partial discharge feature vector corresponding to the signal to be detected and all vectors in the health reference feature space is taken as the insulation degradation index.

[0012] The virtual impedance principle involves constructing a contact state observer to estimate the equivalent impedance value of the contact resistance in real time, and the difference between the actual impedance and the virtual impedance reflects the degree of deviation of the contact state.

[0013] The recursive least squares method estimates the contact resistance parameters in real time through recursive updates, and sets the forgetting factor to 0.98 to achieve exponential weighting of historical data.

[0014] The two-layer game optimization model includes an upper-layer game model and a lower-layer game model. The upper-layer game model optimizes the detection threshold with the goal of minimizing the failure missed detection rate, while the lower-layer game model optimizes the sampling strategy with the goal of maximizing the response speed.

[0015] The upper-level game model and the lower-level game model are associated through a coupling term, which is the product of the total fault detection delay and the detection sensitivity.

[0016] The topology dynamics-based grid fluctuation robustness assessment algorithm abstracts each branch node inside the charging module as a vertex of a weighted graph and identifies energy weak points under grid surge impact through spectral clustering analysis.

[0017] The power grid fluctuation robustness assessment algorithm based on topology dynamics dynamically adjusts the impedance matching strategy of the corresponding branch according to the location of the energy weak point identified by spectral clustering when a surge impact is detected.

[0018] The fault confidence adjustment factor is calculated based on the insulation degradation index, contact resistance parameter and system stability with weights of 0.5, 0.3 and 0.2 respectively. When the fault confidence adjustment factor is between 0.75 and 1.00, a contactor disconnection command is triggered.

[0019] The fault feature database uses a time-series data structure to organize records in chronological order. After accumulating 1,000 sets of new data, an incremental training update is triggered to update the partial discharge feature decoupling model and the health baseline feature space.

[0020] This invention constructs a multi-physics collaborative monitoring architecture based on compressed sensing dimensionality reduction, combining sparse feature extraction of partial discharge signals with online identification of contact resistance parameters. Based on adaptive multi-scale wavelet packet transform decomposition to isolate strong electromagnetic noise, a deep neural network using a sparse autoencoder achieves nonlinear decoupling between partial discharge characteristics and noise components. Simultaneously, a contact state observer based on the virtual impedance principle tracks dynamic changes in contact resistance using recursive least squares, overcoming the technical deficiency of traditional offline detection methods in capturing the coupled evolution of insulation degradation and contact deterioration under operating conditions. A two-layer game optimization model dynamically adjusts the detection threshold and sampling strategy based on the correlation between insulation degradation indicators and contact resistance parameters. Furthermore, a grid fluctuation robustness assessment algorithm based on topological dynamics identifies energy weaknesses under surge impact through spectral clustering and implements adaptive impedance matching, providing system-level electromagnetic environment purification for the stable extraction of weak partial discharge signals. In summary, this invention solves the technical problem mentioned in the background art of the difficulty in achieving collaborative real-time detection of insulation degradation and contact deterioration of European standard through-wall terminals under high-power charging environments. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention.

[0022] Figure 2 A schematic diagram of the input module circuit structure for a 400kW dual-gun charging pile.

[0023] Figure 3 This is a frequency domain distribution diagram of the characteristic components of partial discharge under healthy conditions.

[0024] Figure 4 A comparison of the peak energy values ​​of characteristic components of partial discharge at different degradation stages.

[0025] Figure 5 This is a dynamic evolution trajectory diagram of the fault confidence adjustment factor during the testing process. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0027] like Figure 1 The diagram shows a flowchart of a fault detection method for European standard through-wall terminals provided by the present invention. This method includes the following steps:

[0028] S1. Install a multi-channel high-frequency voltage sensor and a multi-channel high-frequency current sensor at the contact interface of the European standard through-wall terminal. Set the sampling frequency of the multi-channel high-frequency voltage sensor and the multi-channel high-frequency current sensor to 2MHz. At the same time, collect the mV level voltage signal output by the 400A shunt as the reference signal for the total loop current.

[0029] S2. The high-frequency voltage and current signals collected by the multi-channel high-frequency voltage sensor and the multi-channel high-frequency current sensor are compressed in real time by the compression sensing dimensionality reduction processor. The sparse feature coefficients are extracted and transmitted to the charging main control module. The retention rate of the sparse feature coefficients after dimensionality reduction is controlled at 15% to 25% of the original data volume.

[0030] S3. Perform adaptive multi-scale wavelet packet transform decomposition on the sparse feature coefficients after dimensionality reduction. The number of decomposition layers of the adaptive multi-scale wavelet packet transform decomposition is set to 5 layers to extract the partial discharge feature components in the frequency band range of 500kHz to 3MHz.

[0031] S4. Input the partial discharge characteristic components into the partial discharge characteristic decoupling model for nonlinear feature extraction. The partial discharge characteristic decoupling model outputs the partial discharge characteristic vector. Calculate the cosine similarity between the partial discharge characteristic vector and the health benchmark feature space to obtain the insulation degradation index.

[0032] S5. Construct a contact state observer. The contact state observer monitors the dynamic relationship between the voltage drop at the European standard through-wall terminal and the current corresponding to the total loop current reference signal in real time based on the virtual impedance principle. The contact resistance parameters are identified online using the recursive least squares method.

[0033] S6. Input the insulation degradation index and contact resistance parameter into the two-layer game optimization model. The two-layer game optimization model includes an upper-layer game model and a lower-layer game model. The upper-layer game model optimizes the detection threshold with the goal of minimizing the fault missed detection rate, and the lower-layer game model optimizes the sampling strategy with the goal of maximizing the response speed. The two-layer game optimization model outputs the optimized detection threshold and the optimized sampling strategy.

[0034] S7. Start the grid fluctuation robustness assessment algorithm based on topology dynamics. The grid fluctuation robustness assessment algorithm based on topology dynamics abstracts each branch node inside the charging module into the vertices of a weighted graph. It identifies the energy weak points under grid surge impact through spectral clustering analysis. The grid fluctuation robustness assessment algorithm based on topology dynamics outputs the system stability.

[0035] S8. Calculate the fault confidence adjustment factor based on the insulation degradation index, contact resistance parameters and system stability. When the fault confidence adjustment factor is within [0.75, 1.00], trigger the 400A contactor disconnect command. When the fault confidence adjustment factor is within [0.50, 0.75), trigger the corresponding branch switch disconnect command. When the fault confidence adjustment factor is <0.50, only record abnormal data.

[0036] S9. Store the insulation degradation index, contact resistance parameter and fault confidence adjustment factor at each time point into the fault feature database of the charging main control module for continuous training of the partial discharge feature decoupling model and dynamic updating of the health benchmark feature space.

[0037] The compressed sensing dimensionality reduction processor is used to perform dimensionality compression of high-frequency voltage and high-frequency current signals at the sampling end. Based on the principle of signal sparsity, the high-dimensional original signal is projected into a low-dimensional space through the measurement matrix, retaining only the key feature components that can reconstruct the original signal. The input includes the high-frequency voltage signal output by the multi-channel high-frequency voltage sensor and the high-frequency current signal output by the multi-channel high-frequency current sensor at a sampling rate of 2MHz. The output is the dimensionality-reduced sparse feature coefficient matrix. The compressed sensing dimensionality reduction processor uses the orthogonal matching pursuit algorithm to recover the sparse signal from the observations. By iteratively selecting the atom with the highest correlation to the residual, the signal is gradually reconstructed. The amount of sparse feature coefficient data after dimensionality reduction is reduced to 15% to 25% of the original data, effectively alleviating the real-time processing pressure and storage bandwidth impact of the charging main control module.

[0038] The adaptive multi-scale wavelet packet transform decomposition is a time-frequency domain signal decomposition technique. It decomposes sparse characteristic coefficients into sub-band components of different frequency bands by using wavelet basis functions of different scales. The adaptive characteristic is reflected in the dynamic selection of the optimal wavelet basis and the number of decomposition layers based on the local characteristics of the sparse characteristic coefficients. For the high-frequency pulse noise generated by the high-power switching action of the charging module, a 5-layer decomposition structure is set to decompose the sparse characteristic coefficients into 32 sub-bands. The sub-band with a frequency range of 500kHz to 3MHz is extracted as the partial discharge characteristic component. The frequency band covers the main energy distribution range of the early partial discharge signal caused by the aging of the insulation of the European standard through-wall terminal, while avoiding the main frequency of the 400kW power converter and its harmonic interference frequency band, thus achieving effective separation of weak partial discharge signals.

[0039] The specific structure of the partial discharge feature decoupling model is a deep neural network based on a sparse autoencoder, including an input layer, a first sparse coding layer, a second sparse coding layer, a third sparse coding layer, a first decoding layer, a second decoding layer, and an output layer. The input layer receives a 32-dimensional wavelet packet decomposition coefficient vector. The first sparse coding layer contains 64 neurons and is subject to L1 regularization constraints. The second sparse coding layer contains 32 neurons, and the third sparse coding layer contains 16 neurons, forming a bottleneck structure. The first and second decoding layers use a symmetrical structure to progressively reconstruct to 32 dimensions. The output layer outputs an 8-dimensional partial discharge feature vector. The training dataset establishment steps of the partial discharge feature decoupling model specifically include applying a 400V voltage to European standard through-wall terminal samples with different insulation degradation levels in a laboratory environment and gradually increasing it until breakdown. At 80% of the voltage, partial discharge signals throughout the entire process are collected as positive samples, while signals from healthy European standard through-wall terminals at the same voltage are collected as negative samples. Adaptive multi-scale wavelet packet transform decomposition is performed on all samples to obtain 32-dimensional feature vectors, constructing a training dataset containing 5000 positive samples and 5000 negative samples. The training steps of the partial discharge feature decoupling model specifically include training with the Adam optimizer at an initial learning rate of 0.001. The loss function consists of a reconstruction error term and a sparse regularization term, with the coefficient of the sparse regularization term set to 0.01. During training, the cosine similarity between the partial discharge feature vector on the validation set and the healthy baseline feature space is calculated every 50 epochs. Training stops when the change in cosine similarity is less than 0.001 for 10 consecutive epochs.

[0040] The sparsity in the sparse autoencoder is achieved through L1 regularization constraints. L1 regularization adds a penalty term to the loss function, representing the sum of the absolute values ​​of neuron activations. This forces most neurons to have activations close to zero, retaining only the activation responses of a few key neurons. This sparsity closely matches the sparsity characteristics of partial discharge signals in the time-frequency domain. Partial discharge events caused by insulation degradation at European standard through-wall terminals exhibit pulse-like intermittent characteristics on the time axis, and their energy distribution is concentrated only in the frequency band between 500kHz and 3MHz in the frequency domain. By forcing the partial discharge characteristic decoupling model to learn this sparse pattern through L1 regularization constraints, the essential characteristics of weak partial discharge pulses can be effectively decoupled under strong electromagnetic noise backgrounds, avoiding the partial discharge characteristic decoupling model... The sparsity in the partial discharge feature decoupling model structure is specifically implemented in the L1 regularization constraint applied to the first sparse coding layer to prevent overfitting of noise components. The number of neurons in the first sparse coding layer is set to 64, and the sparsity is controlled by the coefficient of the sparse regularization term of 0.01. After training, the average activation rate of the first sparse coding layer is controlled between 5% and 10%, ensuring that only a few neurons respond to the partial discharge feature components. The correlation between the sparsity characteristic and the entire fault detection scheme is that the redundancy of the feature space is reduced by sparse representation, so that the subsequent cosine similarity calculation can more accurately distinguish between the insulation degradation state and the healthy state. At the same time, the low-dimensional partial discharge feature vector output by the third sparse coding layer reduces the computational complexity of the two-layer game optimization model.

[0041] The health baseline feature space is a reference set composed of partial discharge feature vectors from a large number of healthy European standard through-wall terminal samples. The cosine similarity between the partial discharge feature vector corresponding to the signal to be detected and all vectors in the health baseline feature space is calculated, and the maximum cosine similarity value is taken as the insulation degradation index. The cosine similarity calculation formula is as follows: Normalize each component of the partial discharge feature vector to be detected by dividing it by the standard deviation of that component in the training dataset. Normalize each component of each vector in the health baseline feature space by dividing it by the corresponding standard deviation. Calculate the dot product of the two normalized vectors and then divide it by the product of the magnitudes of the two normalized vectors to obtain a cosine similarity value ranging from -1 to 1. When the cosine similarity value is greater than 0.85, it is considered a healthy state; a cosine similarity value between 0.65 and 0.85 is considered early insulation degradation; and a cosine similarity value less than 0.65 is considered severe insulation degradation.

[0042] The virtual impedance principle is to estimate the equivalent impedance value of the contact resistance in real time by constructing a contact state observer. Based on Ohm's law, a theoretical relationship between the voltage drop and the current at the European standard through-wall terminal is established. The ratio of the actual measured voltage drop to the current is used as the actual impedance, and the theoretical impedance value without contact degradation is used as the virtual impedance. The difference between the actual impedance and the virtual impedance reflects the degree of deviation of the contact state. The contact state observer obtains the voltage signal at both ends of the European standard through-wall terminal and the current signal through the total loop current reference signal through a multi-channel high-frequency voltage sensor. The ratio of the real-time voltage drop to the current is calculated to obtain the actual contact impedance. The contact resistance parameters are identified online using the recursive least squares method.

[0043] The recursive least squares method is an adaptive parameter identification algorithm that estimates system parameters in real time through recursive updates. In contact resistance parameter identification, the contact resistance parameter is modeled as a parameter that changes slowly over time. At each sampling moment, the prediction error is calculated using the current voltage drop and current measurement value. The estimated contact resistance parameter value is updated based on the prediction error. The recursive formula includes iterative updates of the gain matrix and covariance matrix. The gain matrix controls the step size of parameter adjustment, and the covariance matrix reflects the uncertainty of the contact resistance parameter estimation. By setting a forgetting factor of 0.98, the historical data is exponentially weighted, making the recursive least squares method pay more attention to recent data to track the dynamic changes of the contact resistance parameter. When the cumulative deviation of the estimated contact resistance parameter value relative to the virtual impedance of 100 consecutive sampling points exceeds 20%, the contact degradation trend is determined to be significant.

[0044] The two-layer game optimization model includes an upper-layer game model and a lower-layer game model. The upper-layer game model aims to minimize the fault false detection rate by adjusting the insulation degradation index threshold and the contact resistance deviation threshold to reduce the risk of false alarms. The lower-layer game model aims to maximize the response speed by optimizing the sampling frequency and data processing cycle to shorten the fault detection delay. The upper-layer and lower-layer game models are linked by a coupling term, which is the product of the total fault detection delay and the detection sensitivity. The objective function of the upper-layer game model is expressed as follows: the false alarm rate normalization value is obtained by dividing the actual number of false alarm samples by the total number of fault samples; the threshold deviation normalization value is obtained by dividing the absolute value of the difference between the insulation degradation index threshold and the standard threshold of 0.65 by 0.2; and the deviation threshold deviation normalization value is obtained by dividing the absolute value of the difference between the contact resistance deviation threshold and the standard deviation of 0.2 by 0.1. The weighted sum of the false alarm rate normalization value, the threshold deviation normalization value, and the deviation threshold deviation normalization value, multiplied by the coupling term, is the objective function of the upper-layer game model. The numbers, with weights of 0.6, 0.3, and 0.1 respectively, are constrained by the insulation degradation index threshold range being between 0.50 and 0.80 and the contact resistance deviation threshold range being between 0.10 and 0.30. The objective function of the lower-level game model is expressed as follows: the processing speed normalization value is obtained by dividing the standard data processing cycle of 50ms by the actual data processing cycle, and the sampling efficiency normalization value is obtained by dividing the sampling frequency by the standard sampling frequency of 2MHz. The geometric mean of the processing speed normalization value and the sampling efficiency normalization value is divided by the coupling term as the objective function of the lower-level game model. The constraint is that the sampling frequency range is between 1MHz and 5MHz and the actual data processing cycle does not exceed 100ms. The coupling term is calculated by dividing the total fault detection delay by the standard delay of 100ms to obtain the delay normalization value, and defining the detection sensitivity as the ratio of the minimum partial discharge pulse amplitude of successful detection to the standard pulse amplitude of 10mV. The product of the delay normalization value and the detection sensitivity is used as the value of the coupling term.

[0045] The topology-dynamics-based grid fluctuation robustness assessment algorithm abstracts each branch node within the charging module as a vertex of a weighted graph, using the admittance relationship between branches as the edge weights. It constructs a dynamic equation describing the propagation characteristics of grid fluctuations, characterizing the flow and attenuation of surge energy between nodes. Eigenvalue decomposition of the graph's Laplace matrix is ​​performed using spectral clustering analysis. The distribution of eigenvalues ​​reflects the system's modal characteristics under fluctuation excitation. Eigenvectors corresponding to smaller eigenvalues ​​reveal the spatial distribution of energy weak points. These weak points are prone to local voltage or current accumulation during grid surge impacts, leading to stress concentration at European standard through-wall terminals. The algorithm monitors the voltage fluctuation amplitude and phase difference of each branch node in real time. When a surge impact is detected, it dynamically adjusts the impedance matching strategy of the corresponding branch based on the location of the energy weak point identified by spectral clustering. Specifically, this involves connecting a dynamic reactor in series or a damping resistor in parallel in the branch with the energy weak point to increase the branch's impedance and disperse the surge energy flow to other paths. The impedance adjustment amount is determined based on the magnitude of the component of the characteristic vector corresponding to the node. The larger the component, the weaker the node is and the greater the impedance compensation required. Through the adaptive impedance matching, the system can reduce the local stress peak to below the safety threshold when facing complex fluctuations such as power grid surges, lightning strikes, or load abrupt changes. This prevents the insulation layer at the European standard through-wall terminal from aging faster due to repeated stress impacts. At the same time, it avoids single-point faults from spreading to the entire charging module through the admittance coupling path and causing a chain reaction of failures. The power grid fluctuation robustness assessment algorithm based on topology dynamics provides system-level robustness assurance for the entire fault detection scheme, enabling partial discharge monitoring and contact resistance parameter identification to operate in a more stable electrical environment. This reduces false alarms and missed detections caused by power grid fluctuations, significantly improving the reliability and accuracy of fault detection. Especially in scenarios with multiple guns operating in parallel and frequent power switching, the power grid fluctuation robustness assessment algorithm based on topology dynamics can effectively suppress electromagnetic coupling interference between branches, creating cleaner measurement conditions for the extraction of weak partial discharge signals.

[0046] The system stability is defined as the reciprocal of the voltage fluctuation variance of all branch nodes, normalized to the interval between 0 and 1. The voltage fluctuation variance is calculated in real time by monitoring the voltage fluctuation amplitude of each branch node using a grid fluctuation robustness assessment algorithm based on topology dynamics.

[0047] The fault confidence adjustment factor function is used to adjust the output layer activation threshold of the partial discharge characteristic decoupling model. This function is calculated based on three data points: insulation degradation index, contact resistance parameter, and system stability. The calculation method involves converting the insulation degradation index from a cosine similarity form to a degradation degree value using the formula 1 minus the cosine similarity value of the insulation degradation index. The normalized deviation is obtained by dividing the cumulative contact resistance deviation corresponding to the contact resistance parameter by the virtual impedance value. The system stability, degradation degree value, and normalized deviation are weighted and summed with weights of 0.5, 0.3, and 0.2 to obtain the fault confidence adjustment factor. When the fault confidence adjustment factor is within [0.75, 1.00], the output layer activation threshold of the partial discharge feature decoupling model is reduced by 30% to improve sensitivity. When the fault confidence adjustment factor is within [0.50, 0.75), the output layer activation threshold remains unchanged. When the fault confidence adjustment factor is <0.50, the output layer activation threshold is increased by 20% to reduce the false alarm rate. The adjustment of the output layer activation threshold directly affects the judgment criteria of the partial discharge feature decoupling model for partial discharge events. In high-risk conditions, the detection capability is enhanced by reducing the output layer activation threshold, while in low-risk conditions, oversensitivity is avoided by increasing the output layer activation threshold.

[0048] The cumulative contact resistance deviation is obtained by summing the deviations of the estimated contact resistance parameters relative to the virtual impedance from 100 consecutive sampling points during the online identification of contact resistance parameters using the recursive least squares method.

[0049] The fault feature database stores insulation degradation indices, contact resistance parameters, fault confidence adjustment factors, and corresponding timestamps and operating condition information for each detection cycle. The fault feature database uses a time-series data structure to organize records in chronological order for continuous training of the partial discharge feature decoupling model. Specifically, incremental training is triggered every 1000 sets of new data. The new data is merged with the original training dataset, and the weight parameters of the sparse autoencoder are retrained. At the same time, the health baseline feature space is updated, old samples with large differences from the distribution of the new data are removed, and the latest health status samples are added. Through this continuous learning mechanism, the partial discharge feature decoupling model adapts to the performance evolution of the European standard through-wall terminal over time, maintaining the long-term effectiveness of fault detection.

[0050] Optionally, the present invention also provides a method for implementing a European standard through-wall terminal fault detection system by means of a computer, wherein the computer is provided with a readable storage medium, the readable storage medium stores program instructions, and the program instructions execute the above-mentioned European standard through-wall terminal fault detection method when the computer is run.

[0051] The specific implementation methods of the above steps are described in detail below.

[0052] The specific implementation of step S1 involves first determining the sensor placement location at the contact interface of the European standard through-wall terminal. The placement location is chosen around the contact point with the highest current density. Multi-channel high-frequency voltage sensors are installed at the positive and negative terminals of the contact interface. These multi-channel high-frequency voltage sensors employ a differential probe structure to suppress common-mode interference. The probe bandwidth needs to cover the DC to 5MHz frequency band to capture transient voltage fluctuation signals. Simultaneously, a multi-channel high-frequency current sensor is installed on the main circuit conductor. This multi-channel high-frequency current sensor uses the Rogowski coil principle to achieve non-contact current measurement, avoiding the influence of additional impedance on circuit characteristics. The sampling frequencies of the multi-channel high-frequency voltage sensor and the multi-channel high-frequency current sensor are... The sampling frequency is uniformly set to 2MHz. The selection of the sampling frequency is based on the Nyquist sampling theorem to ensure that signal components with frequencies up to 1MHz can be reproduced without distortion. The 1MHz frequency covers the main spectral range of partial discharge pulses generated by insulation aging of European standard through-wall terminals. At the same time, a mV-level voltage signal is acquired from the measurement port of the 400A shunt. The voltage signal is amplified by a differential amplifier and used as the reference signal for the total loop current. The gain of the amplifier is set to 100 to 200 times to ensure that the mV-level signal is boosted to the volt level for subsequent analog-to-digital conversion processing. The purpose of these steps is to establish a high-precision, high-bandwidth, multi-parameter synchronous acquisition system to provide the original data foundation for subsequent signal processing and fault feature extraction.

[0053] The specific implementation of step S2 involves sending the high-frequency voltage and current signals acquired in step S1 into a compressed sensing dimensionality reduction processor. The compressed sensing dimensionality reduction processor first constructs a measurement matrix, which is generated using a Gaussian random matrix. The ratio of the number of rows to the number of columns in the matrix is ​​set to 15% to 25%, corresponding to the retention rate after dimensionality reduction. A low-dimensional observation vector is obtained by multiplying the high-dimensional original signal vector with the measurement matrix; this observation vector is the sparse feature coefficient. The compressed sensing dimensionality reduction processor embeds an orthogonal matching pursuit algorithm module. This algorithm iteratively selects the atom with the highest correlation to the residual signal from a preset sparse dictionary. Each iteration updates the residual vector and adds the selected atom to the support set. The termination condition is that the residual energy drops below 1% of the initial signal energy or the number of iterations reaches a preset upper limit. The preset upper limit is generally set to 2 to 3 times the dimension of the observation vector. The core of the orthogonal matching pursuit algorithm is to use the sparse prior knowledge of the signal to recover the key components of the original signal from the incomplete observation data. The key components correspond to the main frequency components and transient change events during the operation of the charging module. The sparse feature coefficients are transmitted to the charging main control module through the communication interface. During the transmission process, lossless compression coding is used to further reduce the amount of data. The purpose of this step is to solve the data explosion problem caused by high-frequency sampling at the megahertz level. By performing dimensionality reduction processing at the sampling end, the impact of massive original data on transmission bandwidth and processing performance is avoided.

[0054] The specific implementation of step S3 involves performing adaptive multi-scale wavelet packet transform decomposition on the sparse feature coefficients. First, wavelet basis functions are selected based on the energy distribution characteristics of the sparse feature coefficients. The Daubechies or Symlet wavelet families are preferred, as they possess good time-frequency localization and approximate symmetry, making them suitable for analyzing abrupt changes in non-stationary signals. The decomposition is set to 5 levels, recursively dividing the signal frequency band from DC to 1MHz into 32 sub-bands, each with a bandwidth of approximately 31.25kHz. The decomposition process employs a wavelet packet tree structure, further decomposing both the low-frequency and high-frequency sub-bands at each level. Unlike traditional wavelet transform strategies that only iteratively decompose the low-frequency portion, wavelet packet transform achieves uniform subdivision of the entire frequency domain. Subbands with a frequency range of 500kHz to 3MHz were extracted from 32 subbands as partial discharge characteristic components. The selection of these frequency bands was based on experimental statistical results. Partial discharge pulses generated by insulation aging exhibited peaks with a rise time of less than 100ns in the time domain, corresponding to frequency domain characteristics concentrated in the 500kHz to 3MHz range. This avoided the switching frequency of the 400kW power converter and the frequency range containing its 2nd to 5th harmonics. The adaptive characteristic was achieved by calculating the energy entropy of each subband. Subbands with energy entropy lower than a set threshold were selected as effective characteristic components. The threshold reference value was 0.3 to 0.5. Low energy entropy indicated a high energy concentration of the signal in that frequency band, corresponding to valuable fault characteristic information. The purpose of this step was to achieve accurate separation of weak partial discharge signals in a strong noise background.

[0055] The specific implementation of step S4 involves organizing the partial discharge feature components into a 32-dimensional input vector and feeding it into the partial discharge feature decoupling model. Each dimension of the input vector corresponds to the energy coefficient of a sub-band. The partial discharge feature decoupling model follows a forward propagation process, sequentially passing through the input layer, the first sparse coding layer, the second sparse coding layer, the third sparse coding layer, the first decoding layer, the second decoding layer, and the output layer. The 64 neurons in the first sparse coding layer perform a nonlinear transformation on the input vector. This nonlinear transformation is achieved using the ReLU activation function, while applying L1 regularization constraints so that the output of most neurons approaches zero, with only about 5% to 10% of neurons remaining active. This sparse activation mode forces the network to learn the most significant feature patterns in the input data. The second and third sparse coding layers further compress the feature dimensions. The third sparse coding layer forms a 16-dimensional bottleneck feature vector. The neck structure forces the network to extract the most essential nonlinear feature representations. The first and second decoding layers use symmetrical neuron configurations to gradually reconstruct back to 32 dimensions. The output layer generates an 8-dimensional partial discharge feature vector. All reference vectors are extracted from the health baseline feature space. The cosine similarity between the partial discharge feature vector to be detected and each reference vector is calculated. The cosine similarity is obtained by dividing the vector dot product by the vector magnitude product. The maximum value among all cosine similarity values ​​is taken as the insulation degradation index. The value range of the insulation degradation index is between -1 and 1. When the insulation degradation index is greater than 0.85, it is judged as a healthy state. The insulation degradation index between 0.65 and 0.85 is judged as early insulation degradation. The insulation degradation index less than 0.65 is judged as severe insulation degradation. The above steps utilize the nonlinear mapping capability of deep neural networks to decouple the essential patterns related to the insulation state from the complex multidimensional feature space.

[0056] The specific implementation of step S5 involves constructing a contact state observer to estimate contact resistance parameters in real time. First, the voltage signals across the two ends of the European standard through-wall terminal are read from a multi-channel high-frequency voltage sensor. The difference between the two voltage signals is calculated to obtain the voltage drop. The current value corresponding to the voltage drop is extracted from the total loop current reference signal. The voltage drop is divided by the current value to obtain the actual contact impedance. The theoretical impedance value without contact degradation is retrieved according to the design specifications of the European standard through-wall terminal as the virtual impedance. The reference value of the virtual impedance is generally 0.1mΩ to 0.5mΩ. The difference between the actual contact impedance and the virtual impedance is calculated as the impedance deviation. The recursive least squares module is then activated for online parameter identification. The recursive least squares method treats the contact resistance parameter as a time-varying parameter to be identified, establishing a linear regression model of voltage drop and current. The model parameters are the contact resistance parameters. At each sampling moment, the measured voltage drop and current value are used to calculate... The model predicts the output, compares the predicted output with the actual measured voltage drop to obtain the prediction error, and updates the gain matrix and covariance matrix based on the prediction error. The gain matrix controls the magnitude of parameter correction, and the covariance matrix quantifies the confidence level of the parameter estimate. A forgetting factor of 0.98 is set to increase the weight of recent data. The forgetting factor causes the weight of data before 100 sampling points to decay to about 13% of the current data, enabling rapid tracking of dynamic changes in contact resistance parameters. The deviation of the estimated contact resistance parameter value relative to the virtual impedance of 100 consecutive sampling points is statistically analyzed, and the deviation is accumulated point by point to obtain the cumulative contact resistance deviation. When the cumulative contact resistance deviation exceeds 20% of the virtual impedance value, the contact degradation trend is determined to be significant. The purpose of this step is to solve the problem of highly nonlinear dynamic drift of contact resistance under the influence of current, stress, and vibration, and to avoid the false alarm risk of a single threshold criterion through online identification.

[0057] The specific implementation of step S6 involves inputting the insulation degradation index and contact resistance parameter into a two-layer game optimization model for collaborative optimization. The upper-layer game model's inputs include the insulation degradation index, contact resistance parameter, and the current detection threshold setting; its output is the optimized insulation degradation index threshold and contact resistance deviation threshold. The lower-layer game model's inputs include the current sampling frequency and data processing cycle; its output is the optimized sampling frequency and data processing cycle. The upper and lower-layer game models interact through a coupling term, which integrates two key performance indicators: total fault detection delay and detection sensitivity. The upper-layer game model increases sensitivity by lowering the detection threshold but increases the false alarm rate. The game model shortens the response time by increasing the sampling frequency but increases the computational burden. Driven by their respective objective functions, the two models search for the Nash equilibrium point, which corresponds to the parameter configuration that optimizes the overall system performance. The solution process adopts an alternating iterative algorithm. First, the parameters of the lower-level game model are fixed to solve for the optimal solution of the upper-level game model. Then, the parameters of the upper-level game model are fixed to solve for the optimal solution of the lower-level game model. The iterative process is repeated until the parameter changes of both models are less than a set threshold. The reference value of the set threshold is 0.001. After the iteration converges, the optimized detection threshold and the optimized sampling strategy are output. The above steps realize the dynamic balance optimization between fault detection sensitivity and system response speed.

[0058] It should be noted that the key technical ideas of this invention include partial discharge feature decoupling technology based on sparse autoencoders and grid fluctuation robustness assessment technology based on topological dynamics. The partial discharge feature decoupling technology based on sparse autoencoders introduces L1 regularization constraints into a deep neural network, forcing the network to learn the sparse representation pattern of partial discharge signals in the time-frequency domain. Compared to traditional Fourier transform methods that can only provide linear frequency domain decomposition, sparse autoencoders can capture high-order coupling relationships between nonlinear features. This effectively decouples the essential characteristics of weak partial discharge pulses even against the backdrop of strong electromagnetic noise generated by high-power switching in the charging module, avoiding the detection failure problem caused by noise masking in conventional methods. Combined with adaptive multi-scale wavelet packet transform preprocessing, this technology achieves highly sensitive identification of partial discharge events in the early stages of insulation aging. The grid fluctuation robustness assessment technology based on topology dynamics abstracts the complex branch network inside the charging module into a weighted graph topology. It utilizes spectral clustering analysis to identify energy weaknesses in the system under grid surge impacts. Compared to traditional protection strategies based on single-point monitoring, this technology assesses the fluctuation propagation path and energy distribution characteristics from a global network perspective. By dynamically adjusting the impedance matching of weak branches, it actively guides and disperses surge energy, preventing local stress concentration from evolving into a global fault. This creates a more stable electrical environment for partial discharge monitoring and contact resistance identification. The synergistic effect of these two key technologies lies in the fact that the topology dynamics assessment technology reduces noise levels at the signal acquisition end by suppressing grid fluctuation interference, providing higher-quality input data for the sparse autoencoder and significantly improving the accuracy of feature decoupling. Simultaneously, the fault features extracted by the sparse autoencoder are fed back to the topology dynamics model to optimize the impedance matching strategy, forming a closed-loop collaborative mechanism for fault detection and system protection. Compared to the existing architecture where detection and protection are independent, this collaborative mechanism exhibits stronger robustness and a lower false alarm rate under complex electromagnetic environments and dynamic load conditions.

[0059] It should be noted that this invention also solves the following technical problem: During the operation of a high-power charging system, the strong electromagnetic noise generated by the main frequency and harmonics of the power converter in the 500kHz to 3MHz frequency band completely covers the energy distribution range of the partial discharge signal, making it difficult for traditional frequency domain filtering methods to effectively separate weak partial discharge pulses. This invention uses a compressed sensing dimensionality reduction processor based on the principle of signal sparsity to project the high-dimensional original signal into a low-dimensional space, retaining only key feature components. Combined with adaptive multi-scale wavelet packet transform decomposition, it dynamically selects the optimal wavelet basis according to the local characteristics of the signal, achieving adaptive suppression of harmonic interference from the power converter. The L1 regularization constraint of the sparse autoencoder forces the neural network to learn the pulse-like intermittent characteristics of the partial discharge event, so that the average activation rate of the first sparse coding layer is controlled at 5% to 10%, ensuring that only a few neurons respond to the partial discharge feature components, thereby achieving effective decoupling of weak partial discharge signals in the context of strong electromagnetic noise. Furthermore, this invention also solves the technical problem of decreased accuracy in contact resistance parameter identification caused by power grid surge impacts. The power grid fluctuation robustness assessment algorithm based on topological dynamics identifies energy weak points by performing eigenvalue decomposition on the Laplace matrix through spectral clustering analysis. When a surge impact is detected, the impedance matching strategy of the corresponding branch is dynamically adjusted according to the magnitude of the eigenvector components, reducing the local stress peak to below the safety threshold. This creates a stable electrical environment for online identification of contact resistance parameters using the recursive least squares method, enabling the exponential weighting mechanism with a forgetting factor of 0.98 to accurately track the slow drift of contact resistance caused by temperature rise and oxidation, thus avoiding false alarms and missed detections caused by parameter estimation errors introduced by power grid fluctuations.

[0060] Specifically, the principle of this invention is as follows: The invention solves the aforementioned technical problem by deeply integrating insulation degradation monitoring and contact deterioration identification at the signal processing level. Its core logic lies in the compressed sensing dimensionality reduction processor, which utilizes the sparsity of partial discharge signals in the time-frequency domain to compress high-dimensional data at a 2MHz sampling rate to 15% to 25% of the original data volume. This preserves the key characteristic components of the partial discharge pulse while reducing real-time processing pressure. This allows the subsequent adaptive multi-scale wavelet packet transform decomposition to accurately extract partial discharge signals in the 500kHz to 3MHz frequency band using a 5-layer decomposition structure with limited computing resources. Furthermore, the sparse autoencoder forces the neural network to learn the pulse-like intermittent characteristics of partial discharge events on the time axis through L1 regularization constraints, avoiding overfitting to noise. The system eliminates misjudgments caused by component factors. Meanwhile, the virtual impedance reference standard constructed by the contact state observer based on Ohm's law and the online identification by the recursive least squares method form a closed-loop feedback, which can track the slow drift of contact resistance caused by temperature rise and oxidation in real time. The two-level game optimization model transforms the contradiction between detection sensitivity and response speed into a collaborative optimization problem between upper and lower level games through coupling terms. The grid fluctuation robustness assessment algorithm based on topological dynamics reveals the propagation law of surge energy between branch nodes through the eigenvalue decomposition of the Laplace matrix. It dynamically adjusts the impedance matching strategy at the energy weak point to suppress stress concentration, thereby eliminating the interference of grid fluctuations on partial discharge monitoring and contact resistance identification at the system level. This allows the insulation degradation index and contact resistance parameters to truly reflect the health status of the European standard through-wall terminals.

[0061] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0062] The specific implementation method of step S1 is the same as described above, and will not be repeated in detail here.

[0063] The specific implementation of step S2 is as follows: the compressed sensing dimensionality reduction processor performs projection transformation on the high-frequency voltage signal and the high-frequency current signal through a measurement matrix. The measurement matrix is ​​constructed based on a Gaussian random matrix. Measurement matrix The construction description is as follows:

[0064] ;

[0065] In the formula, for 3D measurement matrix, The number of measurements is taken as the dimension of the original data. 15% to 25%, For the original signal dimension, For the first A column of measurement vectors, whose elements follow a mean of 0 and a standard deviation of . Gaussian distribution, This represents the matrix transpose operation. The extraction of sparse feature coefficients employs the orthogonal matching pursuit algorithm, the iterative process of which is described below:

[0066] ;

[0067] ;

[0068] ;

[0069] ;

[0070] ;

[0071] In the formula, The initial residual vector, For the observed signal vector, For the initial support set, For the number of iterations, For the first The atomic index selected in the next iteration. For measurement matrix The List, This represents the inner product operation. The L2 norm of a vector. For the first The support set for the next iteration For the first The sparse coefficient vector reconstructed in the next iteration For the index in the measurement matrix belonging to The submatrix formed by the columns, For the first The residual vector of the next iteration. The iteration process continues until the L2 norm of the residual vector is less than a preset threshold or the number of iterations reaches the maximum value. The preset threshold is empirically 1% of the original signal energy, and the maximum number of iterations is 3 times the sparsity by default.

[0072] The specific implementation of step S3 is as follows: the sparse feature coefficients are decomposed into multiple layers using adaptive multi-scale wavelet packet transform decomposition. The decomposition process is implemented using the discrete implementation of the Mallat algorithm. The calculation of the layer decomposition coefficients is expressed as follows:

[0073] ;

[0074] ;

[0075] In the formula, For the first Layer The first of the sub-bands Decomposition coefficients These are the coefficients of the low-pass filter. These are the coefficients of the high-pass filter. For the first Layer The first of the sub-bands Decomposition coefficients For the first Layer The first of the sub-bands Decomposition coefficients. Low-pass filter coefficients. and high-pass filter coefficients The coefficients are derived using the Daubechies wavelet basis db4, with a filter length of 8. The coefficients are obtained by solving the scaling equation of the wavelet function. A 5-level decomposition is used to generate... Each sub-band corresponds to a frequency band range of [number]. to In the formula The sampling frequency is 2MHz. The sub-band index is set to 0 to 31. The sub-bands with a frequency range of 500kHz to 3MHz are extracted and their corresponding indices are 16 to 96, which are used as the partial discharge characteristic components.

[0076] The specific implementation of step S4 is as follows: the 8-dimensional partial discharge feature vector output by the partial discharge feature decoupling model is compared with the health baseline feature space using cosine similarity calculation. The cosine similarity calculation formula is expressed as follows:

[0077] ;

[0078] In the formula, The cosine similarity value is... Let be the first part of the feature vector of the partial discharge to be detected. One portion, Let the first vector in the health baseline feature space be the first vector. One portion, For the first The standard deviation of each component in the training dataset The component index ranges from 1 to 8. Standard deviation The calculation is expressed as follows:

[0079] ;

[0080] In the formula, The total number of samples in the training dataset. For the first The first training sample localized feature vector One portion, For all training samples, the first The mean of each component. The insulation degradation index is defined as the maximum cosine similarity between the partial discharge feature vector to be detected and all vectors in the health baseline feature space, expressed as follows:

[0081] ;

[0082] In the formula, As an indicator of insulation degradation, Let be the total number of vectors in the health baseline feature space. The feature vector of the partial discharge to be detected and the first Cosine similarity of health baseline vectors.

[0083] The specific implementation of step S5 is as follows: the contact state observer calculates the actual contact impedance using the virtual impedance principle. The calculation of the actual contact impedance is expressed as follows:

[0084] ;

[0085] In the formula, for The actual contact resistance unit at any moment , for The voltage drop across a through-wall terminal in the European standard is measured in volts (V). for The current unit (A) corresponding to the current reference signal of the total loop at any given time. Virtual impedance value. The theoretical contact impedance of a healthy European standard through-wall terminal under rated operating conditions was obtained through experimental calibration. The experimental steps included: Step 1, selecting 10 brand-new European standard through-wall terminal samples and running 100 insertion and removal cycles at a current of 400A; Step 2, measuring the contact resistance value after each insertion and removal cycle using the four-wire method; Step 3, taking the median of all measured values ​​as the virtual impedance value. The empirical value is 0.5m Up to 1.2m The parameter update formula for recursive least squares is expressed as follows:

[0086] ;

[0087] ;

[0088] ;

[0089] In the formula, for Gain vector at time step, for The covariance matrix at time t, for The observation vector at time t takes the value of 1. The forgetting factor is set to 0.98. for Units of estimated contact resistance parameters at any given time , for Units of estimated contact resistance parameters at any given time The calculation of the cumulative contact resistance deviation is expressed as follows:

[0090] ;

[0091] In the formula, This is the cumulative amount of contact resistance deviation. Units of virtual impedance , This is the index for the sampling points.

[0092] The specific implementation of step S6 is as follows: the objective function of the upper-level game model of the two-level game optimization model is expressed as follows:

[0093] ;

[0094] In the formula, The objective function value of the upper-level game model. The weight for the normalized value of the false negative rate is set to 0.6. This represents the actual number of samples that were missed. This represents the total number of fault samples. The weight for the deviation from the normalized value of the threshold is set to 0.3. The threshold for insulation degradation index, The deviation threshold is set to a weight of 0.1 based on the deviation from the normalized value. This is the contact resistance deviation threshold. This is a coupling term. The objective function of the lower-level game model is expressed as follows:

[0095] ;

[0096] In the formula, The objective function value of the lower-level game model. The actual data processing cycle is in milliseconds (ms). The sampling frequency is in MHz. The calculation of the coupling term is expressed as follows:

[0097] ;

[0098] In the formula, The total fault detection delay is in milliseconds (ms). The minimum partial discharge pulse amplitude unit for successful detection is mV.

[0099] The specific implementation of step S7 is to construct a Laplace matrix based on the topological dynamics-based power grid fluctuation robustness assessment algorithm. The definition of the Laplace matrix is ​​as follows:

[0100] ;

[0101] In the formula, For Laplace matrix, For degree matrix, This is the adjacency matrix. The degree matrix is... For a diagonal matrix, the first... diagonal elements Adjacency matrix elements Represents a node With nodes The admittance value between. The calculation is expressed as follows:

[0102] ;

[0103] In the formula, For nodes With nodes The impedance value between This is obtained by real-time monitoring of the branch voltage-to-current ratio. The eigenvalue decomposition of the Laplace matrix is ​​expressed as follows:

[0104] ;

[0105] In the formula, The eigenvector matrix, It is an eigenvalue diagonal matrix. for The transpose of the matrix. The calculation of system stability is expressed as follows:

[0106] ;

[0107] In the formula, For system stability, The total number of branch nodes, For the first The voltage fluctuation variance at each node. The calculation is expressed as follows:

[0108] ;

[0109] In the formula, The default length of the sliding window is 50 sampling points. For nodes exist The voltage value at a given time is measured in V. Nodes within the sliding window The average voltage value is measured in volts (V).

[0110] The specific implementation of step S8 is as follows: the calculation of the fault confidence adjustment factor is expressed as follows:

[0111] ;

[0112] In the formula, This is the fault confidence adjustment factor. When... When the 400A contactor is triggered to disconnect, When the corresponding branch switch is disconnected, the command is triggered. Only abnormal data is recorded at that time.

[0113] The specific implementation method of step S9 is the same as described above, and will not be repeated in detail here.

[0114] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To verify the application effect of this invention in a real charging scenario, technicians built a test environment and deployed the fault detection system of this invention in the input module of a high-power dual-gun charging pile to monitor the insulation degradation and contact abnormalities of the European standard through-wall terminals under high load conditions in real time. The test environment is based on... Figure 2 The 400kW dual-gun charging pile input module structure shown is connected to a three-phase AC 400V power grid via a European standard through-wall terminal. After passing through a surge protector, a 400A contactor, a 500A fuse, and a 400A shunt, the power is distributed to the dual gun circuits. The main charging control module is responsible for the control logic of the main circuit and branch circuit switches. Technicians deployed multi-channel high-frequency voltage sensors and multi-channel high-frequency current sensors at the three-phase contact interfaces of the European standard through-wall terminal, with a uniform sampling frequency of 2MHz. Simultaneously, the main circuit current reference signal was obtained from the mV-level voltage signal output by the 400A shunt. During the test, it ran continuously for 48 hours, simulating a dual-gun parallel charging scenario. Gun 1's output power remained at 250kW, while Gun 2's output power dynamically switched between 100kW and 150kW, with the total load fluctuating between 350kW and 400kW.

[0115] During the first 12 hours of testing, the system operated smoothly. Signals collected by the multi-channel high-frequency voltage sensors, after being processed by the compressed sensing dimensionality reduction processor, maintained a stable sparse feature coefficient retention rate of 18% of the original data volume, effectively alleviating the real-time processing pressure on the charging main control module. Technicians then input the dimensionality-reduced sparse feature coefficients into an adaptive multi-scale wavelet packet transform decomposition module, setting the decomposition level to 5 layers to extract partial discharge feature components in the frequency range of 500kHz to 3MHz. For example... Figure 3As shown, under healthy conditions, the amplitude distribution of partial discharge characteristic components is concentrated in the low-energy range, with the peak energy of all 32 sub-band components not exceeding 5mV. Technicians input the extracted partial discharge characteristic components into a partial discharge feature decoupling model for nonlinear feature extraction. This model is based on a deep neural network structure of a sparse autoencoder, comprising an input layer receiving a 32-dimensional wavelet packet decomposition coefficient vector, a first sparse coding layer containing 64 neurons with L1 regularization constraints, a second sparse coding layer containing 32 neurons, and a third sparse coding layer containing 16 neurons forming a bottleneck structure. The output layer outputs an 8-dimensional partial discharge feature vector. The partial discharge feature vector output by the partial discharge feature decoupling model is compared with the healthy baseline feature space using cosine similarity calculation. The resulting insulation degradation index fluctuates between 0.92 and 0.96, significantly higher than the healthy judgment threshold of 0.85, indicating good insulation condition at this stage.

[0116] During the 15th hour of testing, technicians simulated a grid surge by adjusting the load switching frequency. The output power of gun 2 surged from 100kW to 150kW within 10 seconds, then dropped back to 100kW within 30 seconds, repeating this cycle five times. A grid surge robustness assessment algorithm based on topology dynamics monitored the voltage fluctuation amplitude and phase difference of each branch node in real time. Through spectral clustering analysis and eigenvalue decomposition of the Laplace matrix of the graph, the algorithm identified the node corresponding to gun 2 branch as an energy weak point. During a surge, the voltage fluctuation amplitude at this node reached 23V, significantly higher than the 8V of gun 1 branch. The algorithm dynamically adjusted the impedance matching strategy of gun 2 branch by connecting a dynamic reactor in series in this branch. The impedance adjustment amount was determined based on the magnitude of the eigenvector components. The system successfully reduced the local stress peak to below the safety threshold. The system stability was obtained by calculating the reciprocal of the variance of voltage fluctuations at all branch nodes and normalizing it to the 0-1 range. Before the surge, the system stability was 0.94, which dropped to 0.78 during the surge, and then recovered to 0.86 after impedance matching adjustment.

[0117] During the 24th hour of testing, technicians observed slight oxidation at the contact interface of phase A of the European standard through-wall terminal. The contact status observer, based on the principle of virtual impedance, monitored the dynamic relationship between the voltage drop of this phase and the current corresponding to the total circuit current reference signal in real time, and used the recursive least squares method to identify the contact resistance parameters online. The forgetting factor in the recursive formula was set to 0.98, making the algorithm focus more on recent data to track the dynamic changes in the contact resistance parameters. As shown in Table 1, the estimated contact resistance parameters of phase A under healthy conditions are... Ω, rising to [value] in the early stages of contact degradation Ω, significant degradation period reached The cumulative deviation of Ω relative to the virtual impedance increased from 9.5% in the initial stage to 23.8% in the significant stage, exceeding the 20% judgment threshold and triggering an early warning of a significant contact degradation trend.

[0118] Table 1. Monitoring data of contact resistance parameters at different operating stages

[0119]

[0120] Synchronized with the abnormal contact resistance parameters, the insulation degradation indicators also showed a downward trend. Technicians observed that the amplitude of the partial discharge characteristic components of phase A began to increase at the 26th hour, such as... Figure 4 As shown, the peak energy of the characteristic components in the 500kHz to 3MHz frequency band increases from 3.8mV in the healthy state to 8.2mV in the early stage of degradation, and further increases to 15.7mV in the significant degradation stage. The cosine similarity calculation results between the partial discharge characteristic vector output by the partial discharge characteristic decoupling model and the healthy baseline characteristic space show that the insulation degradation index decreases from 0.94 in the healthy state to 0.72 in the early stage of degradation, entering the early stage of insulation degradation judgment range, and further decreases to 0.58 at 30 hours, reaching below the threshold for severe insulation degradation judgment.

[0121] A two-layer game-theoretic optimization model dynamically optimizes thresholds and sampling strategies based on insulation degradation indices and contact resistance parameters. The upper-layer model aims to minimize the false negative rate. By adjusting the insulation degradation index threshold and the contact resistance deviation threshold, the threshold is adjusted from 0.65 to 0.72 at the 28th hour, and the contact resistance deviation threshold is adjusted from 0.20 to 0.15, reducing the risk of false negatives. The lower-layer model aims to maximize response speed. By optimizing the sampling frequency and data processing cycle, the sampling frequency is increased from 2MHz to 2.8MHz, and the data processing cycle is shortened from the standard 50ms to 38ms, significantly improving the real-time performance of fault detection. The upper and lower-layer game-theoretic models are linked by a coupling term, which is the product of the total fault detection delay and the detection sensitivity. During the significant degradation period, the coupling term value is 1.24, higher than the 0.87 in the healthy state, reflecting the system's priority in ensuring detection capabilities under high-risk conditions.

[0122] At the 35th hour of testing, technicians calculated the fault confidence adjustment factor based on insulation degradation indicators, contact resistance parameters, and system stability. The cosine similarity value of the insulation degradation indicator was 0.58, which converted to a degradation degree value of 0.42. The cumulative contact resistance deviation corresponding to the contact resistance parameter was 23.8%, divided by the virtual impedance value. The normalized deviation of Ω was 0.24, and the system stability was 0.86. A fault confidence adjustment factor of 0.68 was obtained by weighting the values ​​with weights of 0.5, 0.3, and 0.2, falling within the interval [0.50, 0.75]. The system maintained the output layer activation threshold unchanged. At the 38th hour, the insulation degradation of phase A further worsened, with the insulation degradation index decreasing to 0.48 and the fault confidence adjustment factor increasing to 0.78, entering the interval [0.75, 1.00]. The system reduced the output layer activation threshold of the partial discharge characteristic decoupling model by 30% to improve sensitivity and simultaneously triggered a 400A contactor disconnect command to cut off the main circuit power input, preventing a safety accident caused by insulation breakdown.

[0123] like Figure 5 As shown, technicians recorded the dynamic evolution trajectory of the fault confidence adjustment factor throughout the entire testing process. During the healthy operation phase, the fault confidence adjustment factor remained in the range of 0.15 to 0.25. In the early stage of contact degradation, it rose to the range of 0.45 to 0.55. During the significant degradation period, it exceeded the 0.75 threshold and finally reached 0.78, triggering the overall circuit power outage protection. The system stores the insulation degradation index, contact resistance parameters, and fault confidence adjustment factor at each moment into the fault feature database of the charging main control module. After accumulating 1000 sets of new data, an incremental training is triggered. The new data is merged with the original training dataset, and the weight parameters of the sparse autoencoder are retrained. At the same time, the health baseline feature space is updated, old samples with large differences from the distribution of the new data are removed, and the latest health status samples are added. Through a continuous learning mechanism, the partial discharge feature decoupling model adapts to the performance evolution of the European standard through-wall terminal over time, maintaining the long-term effectiveness of fault detection.

[0124] The advancements of this invention compared to traditional methods are reflected in technological breakthroughs at multiple levels. Traditional fault detection methods typically rely on a single sensor type and fixed threshold judgment, lacking the ability to coordinate the monitoring of high-frequency partial discharge signals and the dynamic evolution of contact resistance, leading to false alarms or missed detections in complex electromagnetic environments. This invention, through the synchronous acquisition of multi-channel high-frequency voltage and current sensors, combined with sparse feature extraction by a compressed sensing dimensionality reduction processor, achieves efficient capture of weak partial discharge signals at a 2MHz sampling rate, while reducing the data volume to 15% to 25% of the original, breaking through the bottleneck of the real-time processing bandwidth of the charging main control module. Adaptive multi-scale wavelet packet transform decomposition, through a 5-layer decomposition structure, accurately locks the partial discharge characteristic components in the 500kHz to 3MHz frequency band, effectively avoiding interference from the main frequency and harmonics of the 400kW power converter, and solving the problem of feature submersion in strong noise backgrounds caused by traditional fixed-frequency filtering methods. The partial discharge (PD) feature decoupling model is based on a deep learning architecture with a sparse autoencoder. L1 regularization forces the model to learn the sparse pattern of PD signals in the time-frequency domain, avoiding overfitting of noise components by traditional shallow feature extraction methods and improving the accuracy of distinguishing between insulation degradation and healthy states. The contact state observer, based on the virtual impedance principle combined with recursive least squares, achieves online adaptive identification of contact resistance parameters. Compared to traditional offline measurement methods, it can track the dynamic evolution of contact degradation in real time, providing a reliable basis for early warning. The two-layer game optimization model achieves a dynamic balance between minimizing the fault detection miss rate and maximizing the response speed through collaborative optimization of the upper and lower game models, overcoming the limitation of traditional fixed threshold strategies that cannot adapt to changes in operating conditions. The grid fluctuation robustness assessment algorithm based on topological dynamics identifies energy weaknesses through spectral clustering and dynamically adjusts the impedance matching strategy, effectively suppressing the interference of surge impacts on PD monitoring and creating a more stable electrical environment for the extraction of weak PD signals. This proactive adaptability is unattainable by traditional passive protection methods. The fault confidence adjustment factor mechanism, through the fusion decision of insulation degradation index, contact resistance parameter, and system stability, achieves adaptive adjustment of the activation threshold of the output layer of the partial discharge characteristic decoupling model. This enhances detection capability under high-risk conditions and reduces false alarm rate under low-risk conditions, demonstrating the significant advantage of intelligent decision-making over traditional static thresholds. The continuous learning mechanism of the fault feature database, through incremental training and dynamic updates of the health baseline feature space, enables the system to adapt to the performance evolution of European standard through-wall terminals over time, maintaining long-term effectiveness—a self-evolutionary capability that traditional one-time training models cannot possess.

[0125] It should be noted that the variables involved in this invention are explained in detail in Tables 2 and 3.

[0126] Table 2. Variable Explanation Table (Part 1)

[0127]

[0128] Table 3. Variable Explanation Table (Part Two)

[0129]

[0130] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A fault detection method for European standard through-wall terminals, characterized in that, Multi-channel high-frequency voltage sensors and multi-channel high-frequency current sensors are deployed at the contact interface of the European standard through-wall terminals. Simultaneously, the total loop current reference signal output from the shunt is acquired. The acquired high-frequency voltage and current signals are processed in real-time using a compressed sensing dimensionality reduction processor to extract sparse feature coefficients. Adaptive multi-scale wavelet packet transform is applied to these sparse feature coefficients to extract partial discharge feature components. These components are then input into a partial discharge feature decoupling model for nonlinear feature extraction, outputting a partial discharge feature vector. The partial discharge feature vector is then compared with the health benchmark feature space using cosine similarity calculation to obtain an insulation degradation index. A virtual... An impedance-based contact state observer monitors the dynamic relationship between voltage drop and current in real time and uses recursive least squares to identify contact resistance parameters online. Insulation degradation index and contact resistance parameters are input into a two-layer game optimization model to output optimized detection thresholds and optimized sampling strategies. A topology dynamics-based grid fluctuation robustness assessment algorithm is then activated to identify energy weaknesses under grid surge impacts and output system stability. Based on insulation degradation index, contact resistance parameters, and system stability, a fault confidence adjustment factor is calculated to trigger a graded response command. Detection data at each time point is stored in a fault feature database for continuous training of the partial discharge feature decoupling model.

2. The method according to claim 1, characterized in that, The compressed sensing dimensionality reduction processor projects the high-dimensional original signal into a low-dimensional space through a measurement matrix based on the principle of signal sparsity, and recovers the sparse signal from the observations using an orthogonal matching pursuit algorithm.

3. The method according to claim 2, characterized in that, The sparse feature coefficient retention rate is controlled between 15% and 25% of the original data volume.

4. The method according to claim 3, characterized in that, The adaptive multi-scale wavelet packet transform decomposition is set to 5 layers to extract partial discharge characteristic components in the frequency band from 500kHz to 3MHz.

5. The method according to claim 4, characterized in that, The partial discharge feature decoupling model is a deep neural network based on a sparse autoencoder, including an input layer, a first sparse coding layer, a second sparse coding layer, a third sparse coding layer, a first decoding layer, a second decoding layer, and an output layer.

6. The method according to claim 5, characterized in that, The first sparse coding layer contains 64 neurons and is subject to L1 regularization constraints, the second sparse coding layer contains 32 neurons, and the third sparse coding layer contains 16 neurons.

7. The method according to claim 6, characterized in that, The sparsity in the sparse autoencoder is achieved through L1 regularization constraints, which force the activation values ​​of most neurons to approach zero, retaining only the activation responses of a few key neurons.

8. The method according to claim 7, characterized in that, The health benchmark feature space is composed of partial discharge feature vectors of a large number of healthy European standard through-wall terminal samples. The maximum value of the cosine similarity between the partial discharge feature vector corresponding to the signal to be detected and all vectors in the health benchmark feature space is taken as the insulation degradation index.

9. The method according to claim 8, characterized in that, The virtual impedance principle estimates the equivalent impedance value of the contact resistance in real time by constructing a contact state observer, and reflects the degree of deviation of the contact state by the difference between the actual impedance and the virtual impedance.

10. The method according to claim 9, characterized in that, The recursive least squares method estimates the contact resistance parameters in real time through recursive updates, and sets the forgetting factor to 0.98 to achieve exponential weighting of historical data.

Images