Electric vehicle charging module fault detection system and method based on deep learning

Abstract

This invention specifically relates to a deep learning-based fault detection system and method for electric vehicle charging modules, belonging to the field of electric vehicle charging infrastructure technology. It includes: a data acquisition and sensing module; a data preprocessing and feature engineering module; a deep learning model inference module; a fault diagnosis and graded alarm module; and a human-computer interaction and cloud management module. In this invention, a multi-dimensional synchronous acquisition system is established, taking into account both the high-frequency feature capture of transient faults in power devices and the gradual parameter acquisition of gradual faults such as component performance degradation. This ensures the integrity and effectiveness of fault features from the data source level. The signal preprocessing process can filter out grid fluctuations and environmental electromagnetic interference while fully preserving the weak features of early faults. Combined with a multi-model fusion inference architecture, it achieves accurate fault location at the topology and device levels of the charging module, and can effectively distinguish between normal operating condition fluctuations and equipment faults.

Description

Technical Field

[0001] This invention relates to the field of electric vehicle charging infrastructure technology, and in particular to a fault detection system and method for electric vehicle charging modules based on deep learning. Background Technology

[0002] With the large-scale development of the new energy vehicle industry and the continuous upgrading of users' energy replenishment needs, the construction scale, power level, and application scenarios of DC fast charging piles are constantly expanding, making them a core component of the new energy vehicle infrastructure system. As the core functional unit of DC charging piles for energy conversion, the charging module plays a crucial role in converting AC power from the grid into DC power suitable for electric vehicle batteries. Its operational reliability, safety, and full lifecycle health management capabilities directly determine the service continuity of the charging pile, equipment lifespan, and user energy replenishment experience. Because charging modules operate under conditions of high-frequency switching, high-current conversion, wide load range fluctuations, and complex grid environments, the performance degradation and failure risks of their internal power semiconductor devices, energy storage filter elements, drive control circuits, and other core components remain a core focus of technological research and engineering applications in the charging infrastructure industry.

[0003] Currently, fault detection and status monitoring technologies for charging modules have undergone multi-stage technological evolution and have formed a mature engineering application system. Traditional relay protection technologies based on fixed thresholds have achieved rapid response and reliable triggering for severe faults such as overvoltage, overcurrent, overheating, and short circuits, providing a solid technical guarantee for the basic safe operation of charging modules. Fault diagnosis methods based on classical signal processing such as Fourier transform and wavelet analysis have been widely applied in feature extraction scenarios of power electronic devices, effectively identifying signal distortion characteristics corresponding to typical faults. In recent years, with the deep application of deep learning technology in the field of industrial fault diagnosis, intelligent detection methods based on neural networks have also been gradually implemented in charging facility operation and maintenance scenarios, providing a new technical path for the automated identification and classification of charging module faults, and further enriching the industry's technical solutions.

[0004] Faced with the diverse operating conditions and multi-dimensional fault modes of charging modules, as well as the full lifecycle operation and maintenance management needs after the large-scale deployment of charging networks, the industry has raised higher requirements for the engineering capabilities of fault detection technology. In terms of detection coverage, it needs to simultaneously adapt to the feature capture requirements of microsecond-level transient hard faults and long-cycle gradual soft faults, ensuring rapid fault response while achieving early performance degradation identification and predictive maintenance. Regarding deployment adaptability, it needs to balance detection accuracy with the computing power limitations of the charging pile edge, achieving real-time inference while completing precise fault location at the device level, effectively distinguishing between normal operating condition fluctuations and equipment faults. In terms of large-scale application, it needs to build a complete system from data acquisition and fault diagnosis to operation and maintenance scheduling and model iteration, adapting to the centralized management needs of large-scale charging networks.

[0005] Based on this, developing an intelligent fault detection system that combines full-dimensional fault coverage, strong operating condition adaptability, edge-cloud collaboration capabilities, and management capabilities has become an important research and development direction for the intelligent upgrading of the charging infrastructure industry. Summary of the Invention

[0006] The purpose of this invention is to propose a fault detection system and method for electric vehicle charging modules based on deep learning in order to solve the above-mentioned problems.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A deep learning-based fault detection system for electric vehicle charging modules includes: The data acquisition and sensing module is used to collect multi-dimensional raw operating data during the operation of the electric vehicle charging module, and to complete the time synchronization of the data and the recording of abnormal trigger waveforms. The data preprocessing and feature engineering module is used to clean, denoise, perform time-frequency transformation, normalize, and augment the raw running data to generate standardized feature data that is adapted to the input of the deep learning model. The deep learning model inference module is used to deploy trained and lightweight optimized deep learning models to perform end-to-end fault feature extraction and fault state classification on standardized feature data, and output the predicted probability vectors of various faults. The fault diagnosis and graded alarm module is used to perform confidence verification and fault location on the probability vector output by the model, complete the fault severity grading, and trigger corresponding automatic handling and alarm commands. The human-computer interaction and cloud management module is used to realize local visualization of the charging module's operating status, remote centralized management of fault data, operation and maintenance scheduling, and remote iterative updates of deep learning models.

[0008] Preferably, the data acquisition and sensing module is configured with a multi-sampling-rate synchronous acquisition unit, and the acquired raw operating data includes three categories: The first category consists of electrical parameters with a 10kHz sampling rate, including the charging module input / output voltage, output current, real-time power, and PWM drive signal duty cycle. The second category consists of physical parameters for a 1Hz base sampling rate, including the junction temperature and case temperature of the IGBT / MOSFET inside the charging module, the temperature of the filter capacitor, the ambient temperature and humidity, and the insulation resistance. The third category is high-frequency waveform data with a sampling rate of 10MHz or higher, including PWM switching transient waveforms, IGBT collector-emitter voltage waveforms, output voltage ripple, and current transient waveforms; the multi-sampling-rate synchronous acquisition unit supports abnormal triggering waveform recording, which automatically triggers high-resolution waveform recording of 10 power frequency cycles before and after the fault when parameters are detected to be out of limit or waveform distortion is detected.

[0009] Preferably, the method further includes: The data cleaning, wavelet denoising, time-frequency domain transformation, normalization, data augmentation, and feature selection processes are performed sequentially. The wavelet denoising method uses a wavelet basis adapted to power electronic signals to complete multi-layer decomposition and reconstruction, which filters out white noise and electromagnetic interference on the power grid side while retaining transient fault characteristics. The time-frequency domain transformation uses Fast Fourier Transform and Short-Time Fourier Transform to convert a one-dimensional time-domain signal into a two-dimensional time-frequency spectrum, which serves as the input to a deep learning model. The data augmentation expands the number of fault samples by using sliding window slicing, amplitude scaling, and temporal stretching, thus solving the problems of fault sample scarcity and class imbalance.

[0010] Preferably, the deep learning model inference module adopts a multi-model fusion inference architecture including a 1D-CNN sub-model, an LSTM / GRU sub-model, and a temporal Transformer sub-model; The 1D-CNN sub-model is used to extract local distortion features of transient faults of power devices from the original time-series waveforms and identify hard faults such as IGBT open circuit and short circuit. The LSTM / GRU sub-model is used to capture long-term time-series dependencies of parameters and identify gradual faults such as capacitor aging and heat dissipation performance degradation. The time-series Transformer sub-model captures the global correlation features of multivariate time series through a multi-head self-attention mechanism and distinguishes between normal operating condition fluctuations and charging module faults.

[0011] Preferably, the method further includes: It adopts a collaborative deployment architecture of cloud-based training and edge-based inference; Supervised training, semi-supervised training, and incremental optimization of the model are completed in the cloud, and lightweight processing is achieved by model pruning, quantization, and knowledge distillation. The lightweight model is deployed at the edge to achieve millisecond-level real-time inference, and it also supports uploading locally collected fault data and misjudged data to the cloud to complete the iterative optimization of the model.

[0012] Preferably, the fault diagnosis and hierarchical alarm module is equipped with a dual reset reliability verification unit and a topology-level fault location unit; The dual-reset reliability verification unit first sets a threshold for the fault probability output by the model. Only when the fault prediction probability exceeds 95% and lasts for more than 3 inference cycles is the fault initially confirmed. Then, it combines the domain knowledge rules of the physical characteristics of the charging module to complete cross-validation and finally confirm the fault. The topology-level fault location unit maps the classification results output by the model to the specific topological locations and devices of the AC-DC rectifier stage and DC-DC converter stage of the charging module, thereby achieving accurate fault location at the device level.

[0013] Preferably, the fault diagnosis and graded alarm module is equipped with a three-level fault grading and differentiated automatic handling unit, including L1 warning level, L2 derating level and L3 shutdown level; The L1 warning level corresponds to early gradual faults, triggering predictive maintenance warnings and increasing the sampling frequency without interrupting normal charging; The L2 derating level corresponds to a non-fatal fault, triggering an automatic derating of the output power to 50%-80% of the rated power, while simultaneously pushing an operation and maintenance alarm. The L3 shutdown level corresponds to a fatal fault. It immediately blocks the PWM drive signal and disconnects the AC / DC side contactor, triggering an emergency shutdown and audible and visual alarms, while generating an emergency repair work order.

[0014] Preferably, the human-computer interaction and cloud management module includes an edge-end local human-computer interaction sub-module, which is equipped with a dual-access interface of user level and operation and maintenance level. The user-level interface displays the charging status, charging curve and fault prompt information in real time. The operation and maintenance level interface supports playback, amplification and numerical measurement of high-resolution waveforms before and after the fault, realizing on-site verification of the fault without disassembly, and also supports module parameter debugging, local self-test and model firmware upgrade.

[0015] Preferably, the method further includes: The cloud management platform enables centralized monitoring of the status of all charging modules in the network, full-process scheduling of fault alarms and maintenance work orders, and full lifecycle data storage and traceability. It is also equipped with a model OTA upgrade unit, which supports remotely distributing the cloud-optimized model weights to edge charging piles to complete the seamless hot update of the model.

[0016] A deep learning-based fault detection method for electric vehicle charging modules includes: The data acquisition and sensing module synchronously collects multi-dimensional raw operating data of the charging module, triggering high-resolution waveform recording under abnormal operating conditions. The data preprocessing and feature engineering modules clean, denoise, perform time-frequency transformation, normalize, and augment the raw running data to generate standardized feature data. The deep learning model inference module extracts and classifies fault features from standardized feature data, and outputs the predicted probability vectors of various faults. The fault diagnosis and graded alarm module completes the confidence verification of the probability vector, fault location and severity classification, and triggers corresponding automatic handling and alarm commands. Through human-computer interaction and cloud management modules, the system enables visualization of operational status, operation and maintenance scheduling, and model iteration updates, achieving full-process management of fault detection.

[0017] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention establishes a multi-dimensional synchronous acquisition system that takes into account both the high-frequency feature capture of transient faults in power devices and the gradual parameter acquisition of gradual faults such as component performance degradation, ensuring the integrity and effectiveness of fault features from the data source level. The signal preprocessing process can filter out power grid fluctuations and environmental electromagnetic interference while fully preserving the weak features of early faults. Combined with a multi-model fusion inference architecture, it can achieve accurate fault location at the topology and device levels of the charging module, and effectively distinguish between normal operating condition fluctuations and equipment faults.

[0018] 2. This invention constructs a complete technical system that integrates real-time edge processing with centralized cloud management, enabling closed-loop management and large-scale engineering implementation of charging module fault detection throughout the entire process. Through lightweight model optimization design, low-latency real-time inference and rapid fault response can be achieved at the charging pile edge. Combined with hierarchical alarms and differentiated automatic handling strategies, the continuity of charging services is maintained while ensuring equipment and personal safety. Attached Figure Description

[0019] Further details, features, and advantages of this application are disclosed in the following description of exemplary embodiments in conjunction with the accompanying drawings, in which: Figure 1 This is a system structure diagram of the present invention. Detailed Implementation

[0020] Several embodiments of this application will now be described in more detail with reference to the accompanying drawings to enable those skilled in the art to implement this application. This application may be embodied in many different forms and for various purposes and should not be limited to the embodiments set forth herein. These embodiments are provided to make this application thorough and complete, and to fully convey the scope of this application to those skilled in the art. The embodiments described do not limit this application.

[0021] Unless otherwise defined, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It will be further understood that terms such as those defined in commonly used dictionaries shall be interpreted as having a meaning consistent with their meaning in the relevant field and / or the context of this specification, and shall not be interpreted in an idealized or overly formal sense unless expressly defined herein.

[0022] Example 1 Its specific implementation method is combined with the appendix Figure 1 Please provide a detailed explanation.

[0023] In this embodiment, it includes: Data acquisition and sensing module This module is the "sensory system" of the entire system and is the data source foundation for all subsequent analysis and decision-making. The accuracy, synchronization, completeness and sampling rationality of the data it collects directly determine the final accuracy of fault detection. The core objective is to capture the operating characteristics of the charging module under all working conditions without distortion and in all dimensions, while distinguishing between fault signals and environmental interference signals.

[0024] The module's hardware carrier covers the entire data acquisition unit at the edge of the charging pile, including a high-precision intelligent metering chip, a wide-range sensor array, a high-speed ADC recording unit, the charging pile's main control MCU, and the edge acquisition terminal. All acquisition units achieve microsecond-level time synchronization, ensuring the timing consistency of multi-source data. In terms of functionality, the module not only completes the analog-to-digital conversion and acquisition of the raw signals but also synchronously records the real-time operating conditions of the charging pile, including different stages such as constant current, constant voltage, float charging, start-stop, and plugging / unplugging the charging gun. This provides a basis for subsequent data preprocessing to eliminate non-steady-state interference and avoids misjudging operating condition fluctuations as fault characteristics.

[0025] The core data collected by the module is divided into three categories, each corresponding to specific fault characteristics and detection targets, and possessing a clear engineering design logic. The first category is electrical parameters, covering three-phase / single-phase input voltage, input current, and power factor on the AC input side; bus voltage, output current, and real-time output power on the DC output side; and the real-time duty cycle and frequency of the PWM drive signal. These parameters are the core steady-state characteristics of the charging module's power conversion process. The standard sampling rate is set to the 10kHz level, which can completely capture the parameter changes during the module's steady-state operation. Among them, the input-side parameters are used to distinguish between grid-side fluctuations and module faults, while distortions in the output-side parameters directly correspond to abnormalities in the power circuit. Abnormal jumps in the duty cycle can locate faults in the drive circuit, such as a lack of duty cycle caused by the loss of the IGBT drive signal.

[0026] The second category is physical parameters, focusing on the operating status and environmental factors affecting the core vulnerable components of the charging module. Sampling is based on a 1Hz fundamental frequency, with the sampling density automatically increased for scenarios with abnormal temperature rise. The most crucial data is the temperature of the power semiconductor devices. This is not a single-point temperature measurement, but rather individual junction and case temperatures are collected for each IGBT / MOSFET in the AC-DC rectifier stage and DC-DC converter stage, simultaneously collecting the temperature of the heat sink substrate and the inlet and outlet air temperatures of the heat sink. This distinguishes between abnormal heating of the device itself and a failure in the cooling system, such as fan stall or aging thermal grease. In addition, parameters such as the core temperature of the output filter electrolytic capacitor, ambient temperature and humidity, dust concentration in the equipment compartment, and insulation resistance are also collected. These parameters are the core data source for gradual fault warning; for example, an early sign of electrolytic capacitor aging is a slow rise in case temperature, while ambient humidity and salt spray data can help distinguish whether the degradation of insulation performance is due to environmental factors or a fault in the module's own insulation.

[0027] The third category is high-frequency waveform data, which is the core of capturing microsecond-level transient hard faults and a key design feature that distinguishes it from traditional detection systems. Traditional threshold detection can only capture steady-state anomalies after a fault occurs, while the core characteristics of fatal faults such as IGBT short circuits and open circuits only appear in switching transients, requiring high-frequency waveform recording for complete capture. The module uses a high-speed ADC with a sampling rate of over 10MHz, satisfying the Nyquist sampling theorem while fully reproducing the waveform details of PWM switching transients. The acquired waveforms include IGBT gate drive waveforms, collector-emitter voltage waveforms, diode reverse recovery waveforms, DC bus voltage ripple, and output current transient waveforms. Among these, the drive waveform is used to distinguish between drive circuit faults and power device faults. For example, if the drive waveform is normal but the output waveform is distorted, it can be determined that the device itself is damaged. The voltage and current transient waveforms can capture core fault characteristics such as voltage drops and current spikes during IGBT short circuits, and the missing half-cycle waveform during open circuits. The module supports both continuous acquisition and abnormal trigger acquisition modes. Under normal operating conditions, it adopts low-frequency inspection-type acquisition. When a slight parameter over-limit or waveform abnormality is detected, it immediately triggers high-resolution waveform recording of 10 power frequency cycles before and after the fault, which saves storage space at the edge end and can completely retain all the transient data of the fault.

[0028] Data Preprocessing and Feature Engineering Module This module serves as the core bridge connecting raw acquired data with deep learning models. Its core objective is to address the industry pain points of "dirty, messy, poor, and scarce" raw data in industrial scenarios. It performs data cleaning, transformation, enhancement, and feature selection to adapt the data to the input requirements of deep learning models. Simultaneously, it maximizes the preservation of fault characteristics and eliminates interference information, directly determining the model's convergence speed, generalization ability, and detection accuracy. The module's processing flow follows a temporal logic, sequentially completing data cleaning, noise reduction filtering, data alignment and completion, time-frequency domain transformation, normalization, data enhancement, and feature selection. The entire process aligns with the characteristics of power electronic signals and the fault diagnosis needs of charging modules.

[0029] The first step is data cleaning, which systematically processes invalid and interfering information in the raw data to ensure data quality from the source. First, missing values ​​are addressed. For 1-2 missing sampling points caused by short-term sensor disconnection or acquisition interference, linear interpolation and spline interpolation are used to fill in the missing data, ensuring the temporal continuity of the electrical signal. For long-term missing data exceeding 10 sampling points, these are directly marked as invalid and removed to prevent erroneous data from being introduced into the model training and inference process. Second, outliers are handled. A combination of the 3σ criterion and box plot method is used to identify isolated jumps in data during acquisition, such as a single-point spike in current exceeding 10 times the rated value. These data are often outliers caused by electromagnetic interference, rather than true fault characteristics, and are smoothed or removed. Simultaneously, temporal context is used to distinguish outliers from true fault spikes. True fault characteristics exhibit continuous temporal changes, rather than isolated single-point jumps. Finally, the operating conditions were screened, and non-steady-state data during the charging pile start-up and shutdown, no-load, and plug-in / plug-out stages were removed. The electrical signals in these scenarios fluctuate violently and are highly similar to fault characteristics. Only data from steady-state operating stages such as constant voltage and constant current were retained. At the same time, different operating conditions were labeled so that the model could learn the normal feature boundaries under different operating conditions.

[0030] The second step is noise reduction filtering, the core of which is to filter out various interference noises in the original signal while preserving the weak fault characteristics. The operating environment of charging piles is subject to a large amount of electromagnetic interference, including power frequency harmonics from the grid side, white noise, and EMI interference from switching operations. These noises can completely drown out the weak characteristics of early faults such as capacitor aging and IGBT performance degradation. Traditional FIR / IIR linear filtering can only filter out noise in fixed frequency bands and will also erase transient fault characteristics. Therefore, the module uses wavelet transform as the core noise reduction method. Wavelet transform has the characteristics of multi-resolution analysis and can perform localized analysis in both the time and frequency domains simultaneously, perfectly adapting to the processing requirements of non-stationary transient signals of power electronics. In specific implementation, wavelet bases such as db4 and sym6 adapted to power signals are used to perform multi-level wavelet decomposition on the original time domain signal, splitting the signal into approximate coefficients corresponding to the low-frequency main components and detail coefficients corresponding to high-frequency noise and transient characteristics. Then, the detail coefficients corresponding to noise are thresholded and shrunk, and the signal is reconstructed through inverse wavelet transform, ultimately achieving the goal of filtering out interference noise while completely preserving the distortion characteristics of the fault transient.

[0031] The third step is data alignment and time-frequency domain transformation, the core of which is to convert multi-source heterogeneous data into an input format that can be adapted to the deep learning model. First, the time synchronization and alignment of the multi-source data is completed. Since the sampling rates of different parameters vary greatly, for example, temperature parameters are sampled at 1Hz, conventional electrical parameters at 10kHz, and high-frequency waveforms at 10MHz, resampling and interpolation must be performed to align all parameters to the same time axis. This ensures that the multi-dimensional data corresponding to each timestamp is completely synchronized, allowing the model to learn the temporal correlation between different parameters, such as the correspondence between temperature rise and current distortion.

[0032] Subsequently, time-frequency domain transformation is performed, primarily employing Fast Fourier Transform (FFT) and Short-Time Fourier Transform (STFT). The original one-dimensional time-domain signal is difficult to intuitively represent the frequency component changes caused by faults, while most faults in the charging module are directly reflected in changes to the signal spectrum. Under normal operating conditions, the signal spectrum of the charging module is concentrated at the PWM switching frequency and its harmonics, exhibiting a discrete spectral distribution. When faults such as IGBT open circuits or capacitor aging occur, new subharmonic and interharmonic components appear, causing significant changes in the spectral distribution. These characteristics are difficult to capture in the time domain but are extremely evident in the frequency domain.

[0033] FFT can convert a one-dimensional time-domain signal into a frequency-domain spectrum, while STFT, through a sliding time window, converts the signal into a two-dimensional time-spectrum, a three-dimensional spectrum of "time-frequency-amplitude," which can not only show the changes in frequency components but also locate the time points where the changes occur. This spectrum can be directly used as input to a two-dimensional CNN model, converting fault features into image texture features, allowing the model to complete end-to-end feature learning. For multivariate time series, the module stitches the time-spectrum of each parameter into a multi-channel spectrum, similar to the RGB channels of a color image, allowing the model to simultaneously learn the associated fault features of multiple parameters.

[0034] The fourth step is normalization and standardization. The core of this step is to eliminate the dimensional differences between different parameters and the parameter range differences between modules of different power levels. This prevents the model training process from being dominated by parameters with large dimensions, neglecting key fault characteristics in parameters with small dimensions. The different parameters of charging modules have extremely different dimensions. For example, voltage is on the order of hundreds of volts, current is on the order of tens of amperes, temperature is on the order of tens of degrees Celsius, and duty cycles are dimensionless values ​​of 0-1. Furthermore, the rated parameter ranges of a 20kW home charging module and a 60kW fast charging module are completely different. These differences must be eliminated through normalization. The module employs two adaptation methods: for electrical and temperature parameters that conform to a normal distribution, Z-Score normalization is used to map the data to a distribution with a mean of 0 and a variance of 1, eliminating the influence of dimensions; for the amplitude data of the time-spectrum graph, Min-Max normalization is used to linearly map the data to the [0,1] interval, adapting to the input requirements of the CNN model activation function.

[0035] To avoid data leakage, all normalized statistics, including mean, standard deviation, maximum and minimum values, are calculated only from the normal operating condition data of the training set. The test set and inference data are processed using pre-stored normal operating condition statistics to ensure the consistency of model input and generalization ability.

[0036] The fifth step is data augmentation, which addresses the industry pain points of scarce fault samples and class imbalance in industrial scenarios. Charging pile failures are low-probability events, and the number of fault samples can differ by thousands of times compared to the massive amount of data from normal operation. Samples of early and rare faults are particularly scarce, and direct training can lead to a model biased towards predicting normal states, resulting in serious missed detections. This module employs targeted data augmentation methods based on the signal characteristics of the charging module. For time-series signals, the core method is sliding window slicing. A segment of fault waveform data is sliced ​​using a fixed-length sliding window with a fixed step size, breaking down a long fault sequence into dozens of shorter sequences containing complete fault features, achieving a several-fold expansion of a single sample. Simultaneously, amplitude scaling, slight time-axis stretching and compression, and low-amplitude Gaussian noise superposition are used to simulate fault characteristics under different power levels and electromagnetic environments, improving the model's anti-interference ability and generalization. For the time-spectrum image, image augmentation methods such as random cropping, slight flipping, and brightness adjustment are used to expand the number of samples without changing the core fault spectrum characteristics, avoiding model overfitting. All data augmentation operations are performed only on the training set during the model training phase. No augmentation is applied to the test set or edge inference data to ensure the accuracy of model evaluation and actual inference.

[0037] The final step is feature selection. The module extracts massive features from the time domain, frequency domain, and time-frequency domain through preprocessing, including mean, variance, peak value, kurtosis, and margin in the time domain; harmonic distortion rate, spectral centroid, and spectral entropy in the frequency domain; and multi-parameter correlation features. Many redundant features increase the model's computational load and may even introduce noise leading to overfitting. The module employs Pearson correlation coefficient, mutual information, and maximum correlation minimum redundancy algorithms, combined with domain knowledge of power electronics fault diagnosis, to select core features with high correlation to the fault state and low redundancy, while eliminating invalid features irrelevant to the fault. This reduces the model's input dimensionality, improves edge inference speed, and further enhances the model's generalization ability.

[0038] Deep learning model inference module This module is the "intelligent brain" of the entire fault detection system. Its core function is to deploy a trained and lightweight optimized deep learning model to automatically extract fault features and classify and identify operating status in real time. Unlike traditional fixed threshold detection, this module can automatically learn the deep features of different faults through data-driven methods, realize early fault warning, unknown fault identification and multi-condition adaptive detection, while taking into account the low latency and low computing power requirements of edge deployment.

[0039] The module adopts a collaborative architecture of "cloud training + edge inference". The training, optimization, and validation of the model are completed on the cloud server, utilizing the massive computing power and full dataset of the cloud to complete the model's weight optimization and performance iteration. After being trained, the lightweight model is deployed in the edge control unit or T-BOX of the charging pile to process the pre-processed local data in real time and output fault detection results. Edge inference must meet millisecond-level latency requirements, especially for fatal faults such as IGBT short circuits, which must complete inference and response within one switching cycle. To address the different fault characteristics of the charging module, the module adopts a multi-model fusion architecture to adapt to the characteristics and patterns of different faults. Simultaneously, it achieves full coverage of all types of faults through model combination. The core model architecture includes three types.

[0040] The first category is one-dimensional convolutional neural networks (1D-CNNs), which are the core models for processing one-dimensional time-series electrical signals. They are particularly well-suited for detecting transient hard faults in power semiconductor devices, including IGBT open-circuit and short-circuit faults, diode open-circuit and short-circuit faults, and other fault types. The core characteristic of these faults is the local transient distortion of the voltage and current waveforms. Traditional methods require manual design of feature extraction rules, while 1D-CNNs can directly perform end-to-end feature extraction from the original time-domain waveforms without manual intervention. They also have advantages such as low computational cost, fast inference speed, and strong anti-interference capabilities, making them perfectly suited for real-time deployment requirements at the edge.

[0041] The core of 1D-CNN is a one-dimensional convolutional layer. The convolutional kernel slides along the time axis to perform convolution operations on the input time-series signal and extract local features at different scales. Small-sized convolutional kernels can capture high-frequency transient spikes, waveform edge distortions, and other features, corresponding to the fault details of IGBT switching instants. Large-sized convolutional kernels can extract low-frequency full-cycle waveform distortion features, corresponding to waveform anomalies throughout the entire switching cycle.

[0042] By cascading multiple convolutional and pooling layers, the model can extract features layer by layer from low-level waveform features to high-level fault semantic features. Then, a fully connected layer maps these features into probabilistic outputs for different fault categories. In practical applications, the model's input is preprocessed single-cycle or multi-cycle time-series waveforms, such as IGBT collector-emitter voltage waveforms and output current waveforms. Inference can be completed in microseconds, accurately identifying faults in different locations and types of power devices. Furthermore, it exhibits strong adaptability to modules with different operating conditions and power levels.

[0043] The second category is Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs). These are recurrent neural networks specifically optimized for long-term time series. Their core advantage is their ability to capture long-term dependencies in signals over time, solving the gradient vanishing problem of traditional RNNs. They are perfectly suited for detecting gradual and slowly changing faults in charging modules, which are also the core pain points of traditional detection systems. The most typical gradual faults in charging modules are aging of the output filter electrolytic capacitors, performance degradation of the heat dissipation system, and drift of IGBT junction temperature characteristics. The characteristic of these faults is not transient waveform distortion, but rather a slow drift of parameters over time, temperature cycles, and usage cycles. The early changes are extremely weak and may last for months or even years. Traditional threshold methods can only detect faults when they have developed to the end and the parameters exceed the threshold, making early warning impossible.

[0044] LSTM uses gating mechanisms such as input gates, forget gates, and output gates to remember timing information of hundreds or even thousands of time steps, learn the variation patterns of parameters under normal operating conditions, such as normal temperature drift with load changes and normal ripple fluctuations with ambient temperature, and then identify abnormal drifts that do not conform to normal patterns, such as a capacitor case temperature that continues to rise slowly with no load change, or a ripple amplitude that gradually increases while the output current is stable, thereby providing early warning in the early stages of fault occurrence.

[0045] GRU is a lightweight variant of LSTM that combines the input gate and forget gate into an update gate, significantly reducing the number of model parameters and computational cost, resulting in faster inference speed and better suitability for edge deployment requirements, while fully retaining the ability to capture long-term temporal dependencies.

[0046] In practical engineering applications, 1D-CNN is usually combined with LSTM / GRU to build a CNN-LSTM hybrid architecture. First, 1D-CNN is used to extract local waveform features at each time step, and then the feature sequence is input into LSTM / GRU to capture long-term temporal changes, so as to achieve synchronous detection of transient hard faults and gradual soft faults.

[0047] The third category is Transformer models optimized for time series analysis, with the core being a variant of the Time-SeriesTransformer. This is a cutting-edge technology in industrial fault detection in recent years. Its core advantage lies in capturing long-range correlations and global dependencies between multivariate time series through a self-attention mechanism. This addresses the core pain point of traditional models, which cannot distinguish between "normal operating condition fluctuations and grid interference" and "module-specific faults," significantly reducing the system's false detection rate. The actual operating conditions of charging piles are extremely complex. Voltage fluctuations and harmonic interference on the grid side, as well as changes in the charging demand of electric vehicle batteries on the load side, all lead to normal fluctuations in voltage and current parameters. These fluctuations are highly similar to parameter changes caused by faults. Traditional models only focus on the numerical changes of a single parameter, making them prone to misjudgment. The Transformer's multi-head self-attention mechanism can simultaneously calculate the correlation weights between all input variables, learning the coordinated changes of different parameters under normal operating conditions. For example, under normal operating conditions, when the output current increases, the IGBT temperature rises accordingly, and the PWM duty cycle adjusts accordingly; all parameter changes are coordinated and conform to physical laws. However, when the module itself malfunctions, the correlation between parameters is broken. For example, the output current remains unchanged but the IGBT temperature rises sharply, or the duty cycle adjusts but the output voltage does not respond. The Transformer can accurately capture such anomalies in correlation, thus accurately distinguishing between normal operating fluctuations and real faults. Simultaneously, the multi-head attention mechanism can focus on features at different time scales, including millisecond-level switching transient features, second-level parameter change features, and minute-level temperature drift features, achieving synchronous extraction of multi-scale features and covering all fault types. To address the computational limitations at the edge, lightweight Transformer variants, including TST, Informer, and Autoformer, are used in engineering. By optimizing the calculation method of the attention mechanism, the number of model parameters and computational load are significantly reduced, enabling real-time inference at the edge.

[0048] In terms of model training and optimization, the module adopts a training strategy that prioritizes supervised learning and supplements it with semi-supervised learning to address the issues of scarce fault samples and the inability to detect unknown faults. Supervised learning uses labeled normal and fault samples, with the cross-entropy loss function as the core, to optimize the model's fault classification accuracy. Semi-supervised learning, on the other hand, utilizes massive amounts of unlabeled normal operating data to train autoencoders and generative adversarial networks (GANs) to learn the feature distribution of normal operating conditions. When the reconstruction error of the input data exceeds a set threshold, it can be judged as an anomaly. This approach can detect novel and unknown faults that have never appeared before, solving the problem of incomplete fault sample coverage. To address the deployment requirements at the edge, the trained model undergoes systematic lightweight optimization, including model pruning to remove redundant weights and channels; model quantization to convert 32-bit floating-point weights to 8-bit integers, reducing computation and storage space; and knowledge distillation to guide the training of smaller models using a larger model, allowing the smaller models to inherit the feature extraction capabilities of the larger model. Ultimately, this reduces the number of model parameters and computation by over 90% while maintaining minimal loss in detection accuracy, making it suitable for the computing power of low-end MCUs in charging piles and achieving millisecond-level real-time inference. Simultaneously, the module supports incremental model updates. The edge device uploads detected abnormal, fault, and misjudged data to the cloud. The cloud then uses the new data to perform incremental training and optimization of the model before deploying it back to the edge device via OTA, enabling continuous iteration of model performance.

[0049] Fault diagnosis and hierarchical alarm module This module is the core hub connecting model inference results with actual operation and maintenance. Its core objective is to convert the probability vectors output by the model into actionable fault location information, hierarchical handling strategies, and operation and maintenance instructions, thereby realizing the engineering implementation from "fault detection" to "fault handling". It solves the pain point of traditional detection systems that "only alarm, do not locate, and do not handle". It can not only identify the occurrence of faults, but also accurately locate the fault location, assess the fault risk, and trigger automatic handling actions to prevent the fault from escalating and causing equipment damage and safety accidents.

[0050] The core processing logic of the module is divided into four stages: post-processing of model output and confidence verification, accurate fault location, fault severity classification, automatic handling and alarm triggering. The entire process combines the inference results of deep learning with prior knowledge in the field of power electronics to achieve high accuracy and high reliability in fault diagnosis.

[0051] The first step is the post-processing and confidence verification of the model output. The raw output of the deep learning model is the probability value of each fault category. The result with the highest probability cannot be directly used as the final diagnostic conclusion, otherwise false alarms will be caused by single-point misjudgment, affecting the normal operation of the charging pile and the user experience. The module first sets up a dual-reset confidence verification mechanism. The first layer is probability threshold verification. Only when the predicted probability of a certain fault category exceeds 95% and continues for more than 3 consecutive inference cycles is the occurrence of the fault initially confirmed, because real faults are continuous, while model misjudgments are mostly single-point and random. The second layer is domain knowledge rule verification. Combined with the physical characteristics of the charging module, the fault results output by the model are cross-validated. For example, if the model outputs an IGBT short circuit fault, it must simultaneously match the physical characteristics of current spike exceeding the limit, instantaneous drop in DC bus voltage, and sharp rise in the corresponding IGBT temperature. Only when multiple conditions are met simultaneously is the fault finally confirmed, avoiding model misjudgment. For the multi-model fusion architecture, the module also performs multi-model result fusion verification. Only when the three models 1D-CNN, LSTM, and Transformer output the same fault result at the same time is the fault confirmed to have occurred, further reducing the false detection rate.

[0052] The second aspect is precise fault location, which is one of the core values ​​of this module. Traditional fault detection can only determine whether a module has malfunctioned, but cannot pinpoint the specific component. Maintenance personnel need to disassemble the device and check each component individually, resulting in extremely low maintenance efficiency. This system, however, can achieve precise location at the topology and component levels. The core power topology of the charging module is divided into an AC-DC power factor correction rectifier stage and a DC-DC isolation converter stage. The AC-DC rectifier stage is mostly a three-phase full-bridge topology, consisting of six IGBTs forming three bridge arms, with each bridge arm containing two power transistors. The DC-DC converter stage is mostly an LLC resonant topology, containing core components such as switching transistors, resonant cavities, high-frequency transformers, output rectifier diodes, and filter electrolytic capacitors. During the model training phase, the module refines fault labels to specific topological locations and device types. Fault classification not only includes IGBT open circuits, IGBT short circuits, diode open circuits, and capacitor aging, but also corresponds to specific bridge arms and device numbers. Therefore, the model's classification results can be directly mapped to specific devices in the charging module topology, achieving precise topology-level localization of faults such as "AC-DC rectifier stage A-phase upper IGBT short circuit," "DC-DC side output rectifier diode D2 open circuit," and "output filter electrolytic capacitor C5 aging with 30% capacity decay." Simultaneously, the module combines multivariate data correlation analysis to perform secondary verification of the localization results. For example, if a short circuit is located in the A-phase upper IGBT, corresponding features such as A-phase input current distortion, abnormal IGBT temperature rise, and abnormal drive waveform must also match to ensure the accuracy of the localization results and provide precise guidance for subsequent operation and maintenance.

[0053] The third step is fault severity classification. The module classifies faults into three levels based on their impact on personal safety, equipment safety, and charging function, combined with the stage of fault development. Different levels correspond to different handling strategies to avoid "one-size-fits-all" shutdown operations. Under the premise of ensuring safety, the availability of charging piles is maximized, while realizing full life cycle management of faults.

[0054] Level L1 is the early warning level, also known as early fault warning. Faults at this level do not affect the current charging function and do not pose an immediate safety risk. However, if not addressed promptly, they will gradually develop into more serious faults. The core objective is to achieve predictive maintenance and eliminate faults in their early stages. The triggering conditions for this level are: the model detects early, weak fault characteristics with a fault occurrence probability between 60% and 95%, or the parameter change trend matches the early characteristics of a gradual fault, but has not yet exceeded the normal threshold range. Typical fault types include: dust accumulation on the heat sink leading to decreased heat dissipation performance, resulting in a temperature rise 5-10°C higher than normal under the same operating conditions; premature aging of the output filter capacitor, with ESR increasing by 10%-20% and capacitance decaying by 5%-10%; abnormal IGBT junction temperature fluctuations, slight distortion of drive signal edges, and a slow decrease in insulation resistance, etc. The corresponding automatic handling strategy is as follows: the current charging process is not interrupted and the module maintains normal power output; warning information is displayed on the local human-machine interface, and warning data and fault characteristics are uploaded to the cloud management platform and incorporated into the predictive maintenance plan; the sampling frequency and model inference frequency of the module are automatically increased, the changing trend of fault characteristics is continuously tracked, and the fault level is immediately upgraded once the characteristics deteriorate; warning information is pushed to maintenance personnel to remind them to complete the inspection and handling during the next inspection.

[0055] Level L2 is the derating operation level. Faults at this level have already affected some of the module's performance. Continuing to operate at full power would lead to rapid deterioration of the fault and even safety risks. However, the module still possesses some safe charging capabilities. The core objective is to retain basic charging functionality while ensuring safety, avoiding direct shutdown that would impact user experience. The trigger conditions for this level are: the model confirms the occurrence of a fault, the fault probability exceeds 95%, the fault type is non-fatal, and the parameters exceed normal thresholds but have not yet reached safety protection limits. Typical fault types include: moderate aging of the output filter capacitor, with ESR increasing by 30%-50% and capacitance decaying by 10%-20%; partial faults in the cooling system, such as a single cooling fan stalling, resulting in reduced cooling capacity; slight abnormalities in the drive signal of a single bridge arm IGBT; and decreased grid adaptability. The corresponding automatic handling strategy is as follows: immediately implement automatic derating of the module, limiting the maximum output power of the module to 50%-80% of the rated power according to the severity of the fault, reducing the working stress of the module and preventing further deterioration of the fault; display a derating operation prompt on the local human-machine interface to inform the user of the current maximum output power, and simultaneously record the derating event and fault information on the cloud platform; continuously monitor changes in fault characteristics, and if the characteristics continue to deteriorate, immediately upgrade to L3 level to execute shutdown protection; if the characteristics return to normal, gradually restore the rated power output; push alarm information to maintenance personnel, prompting them to complete on-site maintenance and replace the faulty components within the specified time.

[0056] Level L3 is the emergency shutdown level. Faults at this level are fatal, having already caused or about to cause equipment damage, fire, electric shock, or other safety accidents. Immediate shutdown protection must be implemented. The core objective is to ensure personal and equipment safety and prevent the fault from escalating into serious losses. The triggering conditions for this level are: the model confirms a serious fault occurrence, the fault probability exceeds 95%, parameters exceed safety protection limits, or fault characteristics directly threaten safety. Typical fault types include: IGBT / MOSFET shoot-through short circuit, rectifier bridge arm short circuit, output side short circuit, severely excessive input overvoltage / undervoltage, module insulation breakdown, excessive leakage current, and IGBT temperature exceeding the limit junction temperature.

[0057] The corresponding automatic handling strategy is as follows: immediately execute a hardware-level emergency shutdown, first blocking all PWM drive signals and stopping the switching action of power devices, then immediately disconnecting the AC input contactor and DC output relay to completely cut off the power input and output of the module to prevent the fault from escalating; display a red emergency shutdown alarm on the local HMI to inform the user that the charging pile is faulty and cannot be used, and simultaneously trigger on-site audible and visual alarms; immediately upload the fault information, high-frequency waveform data before and after the fault, and full operating data to the cloud platform, and push an emergency alarm to the maintenance personnel to generate a repair work order; lock the operating status of the module and prevent it from restarting until the maintenance personnel complete the fault repair and safety test and manually unlock it before it can resume operation.

[0058] In addition to the core graded handling, the module also has a root cause analysis function. By combining the operating data, working condition data and environmental data before the fault occurred, it can locate the root cause of the fault through correlation analysis. For example, whether the capacitor aging is caused by long-term over-temperature operation or excessive power grid harmonics, or whether the IGBT short circuit is caused by a drive circuit failure or overheating breakdown due to poor heat dissipation, it can guide maintenance personnel not only to replace the faulty components, but also to solve the root cause of the fault and prevent the recurrence of similar faults.

[0059] Human-computer interaction and cloud management module This module serves as the interface between the system and end users, maintenance personnel, and operation and management personnel. It also undertakes the full lifecycle management function of the entire charging network. It is divided into two main parts: the edge-end local human-machine interaction sub-module and the cloud management platform. It achieves full-process coverage from local equipment monitoring and fault handling to remote centralized management and model iteration. It is the core carrier of the entire system from single-device intelligence to large-scale operation of the charging network.

[0060] The edge-end local human-machine interaction submodule is deployed on the touchscreen terminal of the charging pile. It is divided into an interface for ordinary charging users and a professional operation and maintenance interface for maintenance personnel. The two types of interfaces are strictly isolated in terms of permissions to ensure the security of equipment operation. The interface for ordinary users is designed with simplicity and intuitiveness as the core principle. The main interface is a real-time charging status dashboard, which clearly displays basic information such as current charging voltage, charging current, output power, amount charged, charging time, and estimated full charge time. It also overlays the normal charging curve and the residual abnormal curve calculated by the model. If there is an anomaly, it is marked with a bright color so that users can intuitively understand the charging status. Different prompts are set for different levels of faults: L1 level warning is only displayed as a slight prompt in the corner of the interface, which does not affect the user's charging operation; L2 level derating operation will clearly inform the user of the current maximum output power of the charging pile and the reason for the derating; L3 level emergency shutdown will display the fault information with a full-screen red mark, informing the user that the charging pile is unusable, and providing the operator's customer service contact information. The interface also has an emergency stop button, which users can manually trigger in case of emergencies to cut off the power to the charging pile and ensure personal safety.

[0061] The professional maintenance interface, designed for maintenance personnel, requires a dedicated account and password for login, granting them higher operational privileges. Its core services include on-site maintenance, troubleshooting, and equipment debugging. The interface's core is a panoramic health status dashboard, using green, yellow, and red to mark core components within a module that are functioning normally, under warning, or experiencing a fault, respectively. Clicking on the corresponding component allows users to view its historical operating curves, temperature data, waveform information, fault records, and other comprehensive data, enabling maintenance personnel to quickly grasp the module's health status without disassembling the device.

[0062] The core function is high-resolution waveform playback. The system automatically saves high-frequency waveform data for 10 power frequency cycles before and after a fault, with a sampling rate of over 10MHz, completely restoring all waveform details at the moment of the fault, including PWM drive waveforms, IGBT voltage and current waveforms, and bus voltage ripple. Maintenance personnel can zoom in, zoom out, pan, and perform numerical measurements on the waveform data on the interface, comparing the differences between normal and fault waveforms to verify the accuracy of fault location. This completely replaces traditional professional oscilloscopes, allowing on-site fault analysis without the need for specialized equipment. In addition, the interface provides device debugging functions. Maintenance personnel can adjust module operating parameters, sampling frequency, and model inference thresholds on-site, manually trigger waveform recording and module self-tests, complete local model firmware upgrades, and fill in fault handling records, including fault cause, handling method, replaced components, and personnel involved, which are simultaneously uploaded to the cloud platform for full-process traceability of fault handling.

[0063] The cloud management platform is the central hub for the operation and management of the entire charging network. It communicates in real time with the edge terminals of all charging piles via 4G / 5G / Ethernet, enabling centralized, remote, and intelligent management of hundreds or thousands of charging piles and charging modules. It is designed for operation and management personnel, maintenance and scheduling personnel, and algorithm developers of charging pile operators. Its core functions are divided into five categories.

[0064] The first category is centralized equipment status monitoring. The platform collects real-time operating data, health status data, and fault data from all charging modules and displays them centrally on a large visual screen. Users can view equipment status hierarchically by region, site, charging pile, and module, including the total number of online devices, the number of devices operating normally, the number of warnings, the number of devices shut down due to faults, and key indicators such as charging pile utilization and failure rate at each site, achieving a comprehensive overview of the entire network's equipment status. Administrators can click on any charging module to view its real-time operating data, historical operating curves, health status scores, fault records, and other detailed information, remotely monitoring the operation of the entire network's equipment without needing to be physically present.

[0065] The second category is fault alarm and operation and maintenance scheduling functions, which is the core operation and maintenance capability of the platform. When a fault is detected at the edge, the fault information will be immediately uploaded to the platform. The platform will trigger the corresponding alarm and work order process according to the fault level: Level 3 emergency faults will immediately push emergency alarms to the corresponding operation and maintenance personnel via SMS, telephone, and APP push, including the location of the fault site, charging pile number, fault type, and location information. At the same time, an emergency repair work order will be automatically generated, and the nearest operation and maintenance personnel will be dispatched to the site; Level 2 derating faults will generate ordinary operation and maintenance work orders, be included in the operation and maintenance plan, and require operation and maintenance personnel to complete the handling within the specified time; Level 1 early warning faults will be summarized to generate predictive maintenance plans, and operation and maintenance personnel will be arranged to handle them uniformly during the next inspection, realizing the transformation from "passive emergency repair" to "proactive predictive maintenance", significantly reducing operation and maintenance costs and improving equipment availability. The platform also features full-process work order management, from work order generation, dispatch, acceptance, on-site handling, to completion and acceptance, the entire process is traceable. It can automatically calculate core indicators such as average fault repair time, average fault-free time, and work efficiency of maintenance personnel, providing data support for operation and maintenance management.

[0066] The third category is data statistics and analysis functions. The platform conducts big data analysis on operational data, fault data, and maintenance data collected from the entire network to uncover the value behind the data. This includes statistics on the failure rate, common fault types, and average lifespan of charging modules of different brands and models, providing data support for operators to select equipment suppliers; statistics on the frequency of occurrence, root cause distribution, and regional and seasonal distribution patterns of different fault types, such as the significant increase in the incidence of heat dissipation system failures and IGBT overheating failures in high-temperature environments during summer, helping operators to formulate preventive measures in advance; analysis of common root causes of faults, such as excessive grid harmonics causing a large number of IGBT failures, which can be addressed by adding harmonic mitigation devices at the front end to reduce the failure rate at the source; and statistics on the operating efficiency, charging pile utilization rate, and maintenance costs of the entire charging network, providing data basis for operators' operational decisions.

[0067] The fourth category is the OTA (Over-The-Air) upgrade and iteration function of the model. This is the core advantage that distinguishes this system from traditional detection systems. Traditional detection systems, once deployed, have fixed detection logic that cannot be upgraded. However, this system can remotely update model weights via OTA, enabling continuous iteration of model performance. The entire iteration process is as follows: the edge device uploads detected fault data, abnormal data, and misjudged data to the cloud dataset management platform in real time; data labelers annotate the new data, supplementing it to the training dataset and enriching the coverage of fault samples; algorithm developers use the new dataset to incrementally train the model, optimizing model weights, improving the detection capability for new and rare faults, and reducing false positive and false negative rates; after the new model passes testing and verification, it is remotely deployed to the corresponding charging pile's edge device via the OTA platform; the edge device completes a hot update of the model, replacing the old model, without affecting the normal operation of the charging pile. The entire process forms a complete positive flywheel: the more charging piles connected to the system, the richer the collected fault data, and the better the model's detection performance. The platform also has comprehensive model version management functions, recording the version number, training time, performance indicators, and distribution range of each model. It supports model version rollback, and if a new model has an anomaly, it can be immediately rolled back to an older stable version remotely to ensure the stable operation of the system.

[0068] The fifth category is historical data storage and full lifecycle traceability. The platform stores all operational data for the entire lifecycle of all charging modules, including real-time electrical parameters, temperature data, high-frequency waveform data, fault records, maintenance records, model upgrade records, etc., with a storage period of 3-5 years, meeting the requirements of industry supervision and accident traceability. When a safety accident or equipment damage occurs, all operational data and waveform data prior to the accident can be retrieved to completely reconstruct the entire process of the accident, analyze the root cause of the accident, clarify responsibilities, and provide data support for subsequent equipment optimization and testing strategy upgrades.

[0069] In addition, the module has a comprehensive security protection mechanism with strict hierarchical permission management. Different roles have different operation permissions, and all operations are logged and traceable throughout the process. The communication between the edge and the cloud adopts end-to-end encrypted transmission to prevent data theft or tampering. The model OTA upgrade adopts an official signature verification mechanism. Only model firmware that has been signed and certified can be received and updated by the edge, preventing malicious firmware injection and ensuring the network security and data security of the entire system. Example

[0070] A deep learning-based fault detection method for electric vehicle charging modules includes the following parts: The data acquisition and sensing module synchronously collects multi-dimensional raw operating data of the charging module, triggering high-resolution waveform recording under abnormal operating conditions. The data preprocessing and feature engineering modules clean, denoise, perform time-frequency transformation, normalize, and augment the raw running data to generate standardized feature data. The deep learning model inference module extracts and classifies fault features from standardized feature data, and outputs the predicted probability vectors of various faults. The fault diagnosis and graded alarm module completes the confidence verification of the probability vector, fault location and severity classification, and triggers corresponding automatic handling and alarm commands. Through human-computer interaction and cloud management modules, the system enables visualization of operational status, operation and maintenance scheduling, and model iteration updates, achieving full-process management of fault detection.

[0071] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

[0072] It should be noted that, in this document, the use of relational terms such as "first" and "second" is merely for distinguishing one entity or operation from another, and does not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0073] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0074] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0075] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep learning-based fault detection system for electric vehicle charging modules, characterized in that, include: The data acquisition and sensing module is used to collect multi-dimensional raw operating data during the operation of the electric vehicle charging module, and to complete the time synchronization of the data and the recording of abnormal trigger waveforms. The data preprocessing and feature engineering module is used to clean, denoise, perform time-frequency transformation, normalize, and augment the raw running data to generate standardized feature data that is adapted to the input of the deep learning model. The deep learning model inference module is used to deploy trained and lightweight optimized deep learning models to perform end-to-end fault feature extraction and fault state classification on standardized feature data, and output the predicted probability vectors of various faults. The fault diagnosis and graded alarm module is used to perform confidence verification and fault location on the probability vector output by the model, complete the fault severity grading, and trigger corresponding automatic handling and alarm commands. The human-computer interaction and cloud management module is used to realize local visualization of the charging module's operating status, remote centralized management of fault data, operation and maintenance scheduling, and remote iterative updates of deep learning models.

2. The electric vehicle charging module fault detection system based on deep learning according to claim 1, characterized in that, The data acquisition and sensing module is equipped with a multi-sampling-rate synchronous acquisition unit, and the acquired raw operational data includes three categories: The first category consists of electrical parameters with a 10kHz sampling rate, including the charging module input / output voltage, output current, real-time power, and PWM drive signal duty cycle. The second category consists of physical parameters for a 1Hz base sampling rate, including the junction temperature and case temperature of the IGBT / MOSFET inside the charging module, the temperature of the filter capacitor, the ambient temperature and humidity, and the insulation resistance. The third category is high-frequency waveform data with a sampling rate of 10MHz or higher, including PWM switching transient waveforms, IGBT collector-emitter voltage waveforms, output voltage ripple, and current transient waveforms; the multi-sampling-rate synchronous acquisition unit supports abnormal triggering waveform recording, which automatically triggers high-resolution waveform recording of 10 power frequency cycles before and after the fault when parameters are detected to be out of limit or waveform distortion is detected.

3. The electric vehicle charging module fault detection system based on deep learning according to claim 2, characterized in that, Also includes: The data cleaning, wavelet denoising, time-frequency domain transformation, normalization, data augmentation, and feature selection processes are performed sequentially. The wavelet denoising method uses a wavelet basis adapted to power electronic signals to complete multi-layer decomposition and reconstruction, which filters out white noise and electromagnetic interference on the power grid side while retaining transient fault characteristics. The time-frequency domain transformation uses Fast Fourier Transform and Short-Time Fourier Transform to convert a one-dimensional time-domain signal into a two-dimensional time-frequency spectrum, which serves as the input to a deep learning model. The data augmentation expands the number of fault samples by using sliding window slicing, amplitude scaling, and temporal stretching, thus solving the problems of fault sample scarcity and class imbalance.

4. The electric vehicle charging module fault detection system based on deep learning according to claim 1, characterized in that, The deep learning model inference module adopts a multi-model fusion inference architecture, including a 1D-CNN sub-model, an LSTM / GRU sub-model, and a temporal Transformer sub-model. The 1D-CNN sub-model is used to extract local distortion features of transient faults of power devices from the original time-series waveforms and identify hard faults such as IGBT open circuit and short circuit. The LSTM / GRU sub-model is used to capture long-term time-series dependencies of parameters and identify gradual faults such as capacitor aging and heat dissipation performance degradation. The time-series Transformer sub-model captures the global correlation features of multivariate time series through a multi-head self-attention mechanism and distinguishes between normal operating condition fluctuations and charging module faults.

5. The electric vehicle charging module fault detection system based on deep learning according to claim 4, characterized in that, Also includes: A collaborative deployment architecture is adopted, with cloud-based training and edge-based inference. Supervised training, semi-supervised training, and incremental optimization of the model are completed in the cloud, and lightweight processing is achieved by model pruning, quantization, and knowledge distillation. The lightweight model is deployed at the edge to achieve millisecond-level real-time inference, while also supporting the uploading of locally collected fault data and misjudged data to the cloud to complete the iterative optimization of the model.

6. The electric vehicle charging module fault detection system based on deep learning according to claim 1, characterized in that, The fault diagnosis and graded alarm module is equipped with a dual reset reliability verification unit and a topology-level fault location unit; The dual-reset reliability verification unit first sets a threshold for the fault probability output by the model. Only when the fault prediction probability exceeds 95% and lasts for more than 3 inference cycles is the fault initially confirmed. Then, it combines the domain knowledge rules of the physical characteristics of the charging module to complete cross-validation and finally confirm the fault. The topology-level fault location unit maps the classification results output by the model to the specific topological locations and devices of the AC-DC rectifier stage and DC-DC converter stage of the charging module, thereby achieving accurate fault location at the device level.

7. The electric vehicle charging module fault detection system based on deep learning according to claim 6, characterized in that, The fault diagnosis and graded alarm module is equipped with three levels of fault grading and differentiated automatic handling units, including L1 warning level, L2 derating level and L3 shutdown level; The L1 warning level corresponds to early gradual faults, triggering predictive maintenance warnings and increasing the sampling frequency without interrupting normal charging; The L2 derating level corresponds to a non-fatal fault, triggering an automatic derating of the output power to 50%-80% of the rated power, while simultaneously pushing an operation and maintenance alarm. The L3 shutdown level corresponds to a fatal fault. It immediately blocks the PWM drive signal and disconnects the AC / DC side contactor, triggering an emergency shutdown and audible and visual alarms, while generating an emergency repair work order.

8. The electric vehicle charging module fault detection system based on deep learning according to claim 1, characterized in that, The human-computer interaction and cloud management module includes an edge-end local human-computer interaction sub-module, which is equipped with a dual-access interface of user level and operation and maintenance level; the user level interface displays the charging status, charging curve and fault prompt information in real time. The operation and maintenance interface supports playback, amplification, and numerical measurement of high-resolution waveforms before and after a fault, enabling on-site verification without disassembly. It also supports module parameter debugging, local self-testing, and model firmware upgrades.

9. The electric vehicle charging module fault detection system based on deep learning according to claim 8, characterized in that, Also includes: The cloud management platform enables centralized monitoring of the status of all charging modules in the network, full-process scheduling of fault alarms and maintenance work orders, and full lifecycle data storage and traceability. It is also equipped with a model OTA upgrade unit, which supports remotely distributing the cloud-optimized model weights to edge charging piles to complete the seamless hot update of the model.

10. A deep learning-based method for detecting faults in electric vehicle charging modules, as described in any one of claims 1-9, characterized in that, include: The data acquisition and sensing module synchronously collects multi-dimensional raw operating data of the charging module, triggering high-resolution waveform recording under abnormal operating conditions. The data preprocessing and feature engineering modules clean, denoise, perform time-frequency transformation, normalize, and augment the raw running data to generate standardized feature data. The deep learning model inference module extracts and classifies fault features from standardized feature data, and outputs the predicted probability vectors of various faults. The fault diagnosis and graded alarm module completes the confidence verification of the probability vector, fault location and severity classification, and triggers corresponding automatic handling and alarm commands. Through human-computer interaction and cloud management modules, the system enables visualization of operational status, operation and maintenance scheduling, and model iteration updates, achieving full-process management of fault detection.

Images