Method and system for monitoring running state of automobile direct current charging pile in real time

Abstract

The invention discloses an automobile DC charging pile operation state real-time monitoring method and system, and the method comprises the steps: receiving real-time operation parameter data from an automobile DC charging pile, and integrating the real-time operation parameter data into an original multi-dimensional data set; performing nonlinear feature mapping on the original multi-dimensional data set, constructing a state feature vector set, and decomposing each state feature vector into an electrical sub-vector and a mechanical sub-vector; performing real-time fusion analysis on the electrical sub-vectors and the mechanical sub-vectors by using a distributed edge computing framework to generate a preliminary abnormal probability set; performing multi-objective optimization on the initial abnormal probability set to generate an optimized state index set; and performing space-time consistency verification and confidence coefficient screening on the optimized state index set, determining a final operation state level, and outputting a real-time monitoring alarm. According to the embodiment of the invention, the accuracy and real-time performance of fault detection can be improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of charging piles, and in particular to a method and system for real-time monitoring of the operating status of a DC charging pile for an automobile. Background Art

[0002] With the increasing popularity of electric vehicles, the operational safety and stability of DC charging stations have become critical issues. Traditional monitoring methods, which often rely on single-parameter threshold alarms or offline diagnosis, are unable to meet the demand for real-time fault warnings under complex operating conditions. Existing technologies often analyze electrical parameters (such as current and voltage) and mechanical parameters (such as vibration and temperature) independently, leading to overlooked fault correlations and high data processing latency, which affects response efficiency. Furthermore, conventional algorithms lack adaptability to nonlinear characteristics and dynamic operating conditions, making them prone to false positives and missed warnings. Summary of the Invention

[0003] The purpose of the present invention is to provide a method and system for real-time monitoring of the operating status of a DC charging pile for an automobile, so as to address the deficiencies in the prior art and improve the accuracy and real-time performance of fault detection.

[0004] An embodiment of the present application provides a method for real-time monitoring of the operating status of a DC charging pile for an automobile, the method comprising: Receive real-time operating parameter data from the vehicle's DC charging station, collect current, voltage, temperature, and vibration signals based on a preset high-frequency sampling protocol, and integrate them into a raw multidimensional data set; Performing nonlinear feature mapping on the original multidimensional data set to construct a state feature vector set, and decomposing each state feature vector into an electrical sub-vector and a mechanical sub-vector to separately capture electrical characteristics and mechanical vibration characteristics; Using a distributed edge computing framework, the electrical sub-vector and the mechanical sub-vector are fused and analyzed in real time to generate a preliminary abnormality probability set; Performing multi-objective optimization on the preliminary abnormal probability set, applying a reinforcement learning algorithm to optimize real-time responsiveness and fault prediction accuracy, and generating an optimized state indicator set; The optimization status indicator set is subjected to spatiotemporal consistency verification and confidence screening to determine the final operating status level and output a real-time monitoring alarm.

[0005] Optionally, the real-time operating parameter data from the DC charging pile of the vehicle is received, current, voltage, temperature and vibration signals are collected based on a preset high-frequency sampling protocol, and integrated into an original multidimensional data set, including: The high-frequency waveforms of current and voltage are collected synchronously through a multi-channel anti-aliasing filter, while the three-dimensional vibration signals are captured using a MEMS accelerometer to generate a heterogeneous sensing stream with a time stamp. Adaptive Kalman filtering is used to align heterogeneous sensor streams in the time domain, eliminating the timing deviation caused by the physical location differences of the sensors and outputting a synchronized sensor matrix. Perform wavelet packet decomposition on the synchronized sensing matrix, extract energy entropy features in a specific frequency band, and construct an initial multi-dimensional feature tensor; The initial multidimensional feature tensor is reconstructed and the redundant dimensions are compressed by Tucker decomposition to generate a compact original multidimensional dataset.

[0006] Optionally, performing nonlinear feature mapping on the original multidimensional dataset to construct a set of state feature vectors, and decomposing each state feature vector into an electrical sub-vector and a mechanical sub-vector to separately capture electrical characteristics and mechanical vibration characteristics, includes: The original multidimensional data set is input into the kernel function space, and nonlinear mapping is performed through the radial basis function kernel to generate a high-dimensional state feature space; Perform orthogonal feature decoupling in the high-dimensional state feature space and use the FastICA algorithm to separate electrical and mechanical features; Apply information entropy weighting to the decoupled electrical features to generate electrical sub-vectors with physical interpretability; Perform wavelet coherence analysis on mechanical characteristics, extract the phase coupling coefficient between the vibration signal and the cooling fan speed, and construct the mechanical sub-vector; The electrical sub-vector and the mechanical sub-vector are combined into a state feature vector set through the feature cascade layer, and a feature validity flag is added.

[0007] Optionally, the use of a distributed edge computing framework to perform real-time fusion analysis on the electrical sub-vector and the mechanical sub-vector to generate a preliminary abnormality probability set includes: Deploy a lightweight graph neural network on the edge computing node, use the electrical subvector as the node attribute and the charging pile topology connection relationship as the edge weight to build the electrical state graph structure; Perform spatiotemporal convolution on the mechanical sub-vectors to aggregate the vibration characteristics of adjacent charging piles in the spatial dimension, and apply causal convolution in the temporal dimension to generate a mechanical state evolution tensor; Design a dual-channel attention fusion mechanism, in which the electrical channel focuses on current mutation events and the mechanical channel focuses on abnormal vibration frequency bands, and outputs a cross-modal attention weight matrix; Based on the attention weight matrix, a gated recurrent unit is used to fuse the electrical state graph and the mechanical evolution tensor to generate a multi-dimensional state fusion vector; The multi-dimensional state fusion vector is input into the anomaly detection forest, and a preliminary anomaly probability set of each dimension is output.

[0008] Optionally, performing multi-objective optimization on the preliminary abnormal probability set, applying a reinforcement learning algorithm to optimize real-time responsiveness and fault prediction accuracy, and generating an optimized state indicator set includes: Construct a dual-objective reward function where real-time responsiveness is rewarded with associated processing latency, and prediction accuracy is rewarded with associated fault detection F1 score; Design a deep Q-network architecture that uses anomaly probability sets as state inputs and feature sampling frequencies as action space, and estimates the value function through a two-stream network. Adopting a non-dominated sorting strategy, the action strategy is selected on the Pareto optimal frontier, and the current sampling frequency and vibration analysis granularity are dynamically adjusted; The network parameters are updated through the proximal strategy optimization algorithm, and the optimized state indicator set is output.

[0009] Optionally, performing spatiotemporal consistency verification and confidence screening on the optimization status indicator set, determining a final operating status level, and outputting a real-time monitoring alarm includes: Construct a spatiotemporal consistency check window, use exponentially weighted moving average to filter out instantaneous interference in the time dimension, and compare the Mahalanobis distance of similar indicators of adjacent charging piles in the spatial dimension; Design a confidence screening funnel, where the first layer removes highly volatile indicators, and the second layer retains stable features to generate a subset of high-confidence indicators; Apply the fuzzy inference engine, input a subset of high-confidence indicators, and output a state level matrix through a preset rule base; Perform optimized alarm distribution, add timestamp hash chains to high-risk alarms, push them to the operation and maintenance terminal through priority queues, and trigger the device power reduction protection protocol at the same time.

[0010] Another embodiment of the present application provides a real-time monitoring system for the operating status of a DC charging station for an automobile, the system comprising: The receiving module is used to receive real-time operating parameter data from the vehicle DC charging station, collect current, voltage, temperature and vibration signals based on a preset high-frequency sampling protocol, and integrate them into a raw multidimensional data set; a mapping module, configured to perform nonlinear feature mapping on the original multidimensional data set, construct a set of state feature vectors, and decompose each state feature vector into an electrical sub-vector and a mechanical sub-vector to separately capture electrical characteristics and mechanical vibration characteristics; an analysis module, configured to perform real-time fusion analysis of the electrical sub-vector and the mechanical sub-vector using a distributed edge computing framework to generate a preliminary abnormality probability set; an optimization module for performing multi-objective optimization on the preliminary abnormal probability set, applying a reinforcement learning algorithm to optimize real-time responsiveness and fault prediction accuracy, and generating an optimized state indicator set; The screening module is used to perform spatiotemporal consistency verification and confidence screening on the optimization status indicator set, determine the final operating status level, and output a real-time monitoring alarm.

[0011] Yet another embodiment of the present application provides a storage medium, wherein the storage medium stores a computer program, wherein the computer program is configured to execute any of the above methods when run.

[0012] Yet another embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute any of the above methods.

[0013] Compared with the prior art, the present invention provides a real-time monitoring method for the operating status of a DC charging pile for a vehicle. The method receives real-time operating parameter data from the DC charging pile for a vehicle and integrates it into an original multidimensional data set; performs nonlinear feature mapping on the original multidimensional data set to construct a state feature vector set, and decomposes each state feature vector into an electrical sub-vector and a mechanical sub-vector; utilizes a distributed edge computing framework to perform real-time fusion analysis of the electrical sub-vector and the mechanical sub-vector to generate a preliminary abnormality probability set; performs multi-objective optimization on the preliminary abnormality probability set to generate an optimized state indicator set; performs spatiotemporal consistency verification and confidence screening on the optimized state indicator set to determine the final operating state level, and outputs a real-time monitoring alarm, thereby improving the accuracy and real-time performance of fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 A hardware structure block diagram of a computer terminal for a method for real-time monitoring of the operating status of a DC charging pile for a vehicle provided by an embodiment of the present invention; Figure 2 A flow chart of a method for real-time monitoring of the operating status of a DC charging station for a vehicle provided by an embodiment of the present invention; Figure 3 This is a structural diagram of a real-time monitoring system for the operating status of a DC charging pile for an automobile provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0015] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and are not to be construed as limiting the present invention.

[0016] The embodiment of the present invention first provides a method for real-time monitoring of the operating status of a vehicle DC charging pile. The method can be applied to electronic devices such as computer terminals, specifically ordinary computers.

[0017] The following describes it in detail by taking running on a computer terminal as an example. Figure 1The hardware structure block diagram of a computer terminal for a method for real-time monitoring of the operating status of a DC charging pile for a vehicle provided by an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, a memory, and a network interface connected via a system bus, wherein the memory may include a non-volatile storage medium and an internal memory.

[0018] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions that, when executed, enable a processor to execute any of the methods for real-time monitoring of the operating status of a DC charging station for a vehicle.

[0019] The processor is used to provide computing and control capabilities and support the operation of the entire computer equipment.

[0020] The internal memory provides an environment for the operation of the computer program in the non-volatile storage medium. When the computer program is executed by the processor, the processor can execute any one of the methods for real-time monitoring of the operating status of a DC charging pile for a vehicle.

[0021] The network interface is used for network communication, such as sending assigned tasks, etc. Those skilled in the art will understand that Figure 1 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0022] It should be understood that the processor may be a central processing unit (CPU), other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor, etc.

[0023] See also Figure 2 The embodiment of the present invention provides a method for real-time monitoring of the operating status of a DC charging pile for a vehicle, which may include the following steps: S201, receiving real-time operating parameter data from a DC charging station for a vehicle, collecting current, voltage, temperature, and vibration signals based on a preset high-frequency sampling protocol, and integrating them into an original multidimensional data set; Specifically, high-frequency waveforms of current and voltage can be synchronously acquired through a multi-channel anti-aliasing filter, while three-dimensional vibration signals can be captured using a MEMS accelerometer to generate a heterogeneous sensing stream with a time stamp; Multi-channel anti-aliasing filtering and high-frequency waveform acquisition A multi-channel anti-aliasing filter (MAAF) is deployed at the power input of the DC charging station. Its core function is to prevent spectral aliasing distortion when sampling high-frequency signals. Current signals are acquired via Rogowski coils or Hall-effect sensors, while voltage signals are acquired via a high-precision voltage-divider resistor network. The filter utilizes a Butterworth low-pass topology, with a cutoff frequency set at 0.4 times the target sampling frequency (based on the Nyquist sampling theorem). For example, if the current sampling frequency is 100 kHz, the cutoff frequency is set at 40 kHz to suppress noise above this frequency (such as high-frequency harmonics from switching transistors). Each channel is equipped with an independent 24-bit analog-to-digital converter (ADC) that simultaneously captures the instantaneous waveform values ​​(IWV) of current and voltage at a sampling rate of 100 kHz. The sampling process is driven by a high-stability crystal oscillator, and the time synchronization error is controlled within ±50 nanoseconds (ns).

[0024] 3D Vibration Capture with MEMS Accelerometers Mechanical vibration signals are collected by MEMS accelerometers (Micro-Electro-Mechanical Systems Accelerometers) mounted on key locations of the charging pile, such as the power module heat sink and transformer housing. Based on a capacitive micromechanical structure, these sensors simultaneously measure acceleration (in g, representing gravitational acceleration) in all three axes (X, Y, and Z). Their sampling frequency is set at 5 kilohertz (kHz), covering the typical frequency range of mechanical vibration in charging piles (0 Hz to 2.5 kHz). The sensors include a built-in self-calibration circuit to eliminate zero-point errors caused by temperature drift (ranging from -40°C to 85°C). The vibration signals are output as digital signals and transmitted to the main control unit via a high-speed serial peripheral interface (SPI). Each vibration data point is accurately timestamped, provided by the same crystal oscillator used for electrical sampling, ensuring a consistent time base across all modal data.

[0025] Generation and integration of heterogeneous sensor streams Current and voltage waveform data, along with three-dimensional vibration signals, are collected by the Data Preprocessing Unit (DPU). This unit assigns a separate Data Stream Channel to each type of data and inserts a unified timestamp (with an accuracy of ±100 nanoseconds). Current and voltage waveforms are stored as arrays (e.g., 100,000 samples per second), while vibration data is packaged into structures based on three-axis acceleration components. This ultimately generates a heterogeneous sensing stream (HSS), whose data structure consists of four parts: Current Stream: Contains a sequence of timestamps and corresponding current amplitudes (unit: ampere (A)); Voltage Stream: a sequence of timestamps and voltage amplitudes (in volts (V)). Vibration Stream: timestamp sequence and X / Y / Z axial acceleration values ​​(unit: meters per second squared (m / s) 2 )).

[0026] The data stream is transmitted to the edge computing node in real time via Gigabit Ethernet.

[0027] Adaptive Kalman filtering is used to align heterogeneous sensor streams in the time domain, eliminating the timing deviation caused by the physical location differences of the sensors and outputting a synchronized sensor matrix. Timing deviation modeling and filter initialization Due to the different physical installation locations of the current sensor, voltage probe, and MEMS accelerometer within the charging pile, the signal transmission path lengths vary (typically 0.1 to 1 meter), resulting in fixed transmission delays (FTDs) in each data stream. For example, the current signal transmission delay is approximately 300 nanoseconds (ns), and the vibration signal delay is approximately 500 nanoseconds (ns). To eliminate this deviation, an adaptive Kalman filter (AKF) first establishes a state-space model: State Variable: The real time offset of each signal flow (unit: nanosecond (ns)); Observation Variable: The actual arrival time difference detected based on the reference edge of the charging pile control pulse (such as the IGBT drive signal).

[0028] When initializing the filter, the process noise covariance (Q) is set to 1e-6 and the observation noise covariance (R) is set to 1e-5, which represents the estimation of the system dynamics and measurement errors.

[0029] Dynamic adjustment process of time domain alignment During the runtime, AKF executes a Predict-update Cycle with a period of 10 milliseconds (ms): Prediction Phase: Based on the estimated time offset at the previous moment and the system dynamics equation (assuming the offset changes slowly), the current offset is predicted. Update Phase: When a reference edge is detected (e.g., the rising edge of a charging pile power module), the actual arrival time of the event in each signal stream is compared with the theoretical time (determined by the crystal oscillator global clock) to generate an observation residual. The Kalman gain (K) is used to weight the residual and correct the predicted value.

[0030] For example, if the detected current signal lags 400 nanoseconds (ns) relative to the reference edge, and the predicted value is 350 nanoseconds (ns), the offset is updated to 390 nanoseconds (ns) using a gain of K = 0.8. This process tracks delay drift caused by factors such as temperature changes and cable aging in real time.

[0031] Synchronize the output of the sensor matrix The aligned data are reconstructed into a synchronized sensing matrix (SSM). The rows of the matrix correspond to a unified time axis (with a time resolution of 1 microsecond (µs)), and the columns contain four types of synchronization signals: Column 1: current value after alignment (A); Column 2: voltage value after alignment (V); Columns 3-5: X / Y / Z axial vibration acceleration values ​​(m / s 2 ).

[0032] Missing values ​​in the matrix (due to sensor response delays) are filled using cubic spline interpolation. The SSM is finally output to subsequent processing modules to ensure that all sensor data are accurately correlated under the same time base.

[0033] Perform wavelet packet decomposition on the synchronized sensing matrix, extract energy entropy features in a specific frequency band, and construct an initial multi-dimensional feature tensor; Frequency Band Division of Wavelet Packet Decomposition Wavelet Packet Decomposition (WPD) uses the Daubechies4 Wavelet to perform a 4-level decomposition (Decomposition Level = 4) on each column of the SSM signal, generating 16 sub-bands. The frequency band division is based on binary recursion: the first level is (0 Hz to sampling frequency / 2), the second level is subdivided into 0 Hz to sampling frequency / 4 and sampling frequency / 4 to sampling frequency / 2, and so on. For example, for a 100 kHz current signal, the sub-band width of the fourth level is 100 kHz / 16 = 6.25 kHz. The key frequency bands are preset based on the characteristics of the charging pile fault: Electrical characteristic frequency band: 0 Hz to 1 kHz (DC component and low-frequency ripple), 1 kHz to 10 kHz (switching frequency harmonics); Mechanical characteristic frequency band: 100 Hz to 500 Hz (fan fundamental frequency), 1 kHz to 2 kHz (bearing resonance).

[0034] Energy entropy feature extraction Calculate the energy entropy (EE) in each sub-band. Its physical meaning reflects the randomness of the signal energy distribution. Specific steps: Step 1: Calculate the sum of the squares of the sub-band coefficients to obtain the band energy (BE). Step 2: Divide BE by the total signal energy and normalize it to a probability distribution (PD). Step 3: Calculate the entropy value using the Shannon Entropy Formula: EE = -Σ(PD ×log2(PD)).

[0035] For example, when the cooling fan bearings wear, the energy entropy in the 1 kHz to 1.5 kHz frequency band increases from 0.8 bits to 1.2 bits, indicating that the energy distribution is more dispersed (a sign of abnormality).

[0036] Initial multi-dimensional feature tensor construction All sub-band energy entropy values ​​of the five signals, namely current, voltage, and triaxial vibration, are organized into an initial multidimensional feature tensor (IMFT) according to the time window (WindowLength = 1 second). The tensor dimensions are: Dimension 1: the number of samples in the time window (e.g., 1000, corresponding to 1 second); Dimension 2: Number of signal types (current, voltage, X / Y / Z vibration, a total of 5 types); Dimension 3: Number of sub-bands (16 sub-bands in total with 4-layer decomposition).

[0037] For example, an IMFT slice can be represented as: time sample index × signal type × frequency band index, storing the energy entropy value of the signal in that frequency band at that moment. This tensor fully preserves the time-frequency domain characteristics of the original signal.

[0038] The initial multidimensional feature tensor is reconstructed and the redundant dimensions are compressed by Tucker decomposition to generate a compact original multidimensional dataset.

[0039] Necessity and strategy of tensor reconstruction IMFT has significant dimensional redundancy: Adjacent time samples are highly correlated (e.g., electrical parameters change slightly within 1 millisecond (ms)). The energy entropy of some frequency bands is close to zero (for example, the current signal has no energy above 10 kHz).

[0040] Tensor Reshaping reorganizes the original three-dimensional structure of IMFT (time × signal type × frequency band) into an optimized new dimension: New Mode 1: Combines time and signal type to form a joint dimension of “time-signal”; New Mode 2: retains the frequency band dimension but only selects eight key frequency bands whose total energy exceeds a threshold (e.g., 1% of the total energy); New Mode 3: Adds the dimension of “feature type” to leave room for expansion.

[0041] After reconstruction, the tensor size is compressed from 1000×5×16 to 500×8×1 (compression rate 83%), reducing the computational load.

[0042] Second paragraph: Tucker decomposition of the core operations Tucker Decomposition decomposes the reconstructed tensor into the product of the core tensor and the factor matrix: Step 1: Set the target rank to (50, 3, 1), which means retaining 50%, 37.5%, and 100% of the information of the original tensor in the three dimensions; Step 2: Use the Higher-Orthogonal Iteration (HOOI) algorithm to iteratively solve the problem: Fix two dimensions and perform singular value decomposition (SVD) on the remaining dimensions; Truncate the singular vectors (keep the first 50, 3, and 1 principal components) and update the factor matrix; Step 3: After convergence, the core tensor size is 50×3×1, and the factor matrix sizes are 500×50 (time-signal), 8×3 (frequency band), and 1×1 (feature type).

[0043] Generation of compact datasets The core tensors and factor matrices together form a compact raw multidimensional dataset (CRMDD). This dataset is stored in binary format and contains: Core tensor (50×3×1): stores coupling relationships across time, frequency bands, and feature types; Time-signal factor matrix (500 × 50): Each column represents a time-signal joint mode (e.g., current slow-changing mode); Band factor matrix (8×3): Each column represents the band clustering result (such as "low-frequency electrical band group"); Feature type factor matrix (1×1): scalar value is 1, preserving dimensionality consistency.

[0044] For example, a core tensor element of (25,2,1) = 0.8 indicates that the correlation strength between the 25th time-signal pattern and the second frequency band group is 0.8. The dataset is only 12% of the original IMFT, laying the foundation for subsequent real-time analysis.

[0045] First, multiple sensors are used to synchronously collect key physical quantities during charging pile operation, including electrical parameters (current, voltage) and mechanical parameters (temperature, vibration). High-frequency sampling is employed to ensure the capture of transient abnormal signals. A pre-set protocol aligns heterogeneous sensor data in the time domain to form a structured multidimensional dataset, providing a unified data foundation for subsequent analysis and enabling comprehensive awareness of the charging pile's operating status. High-frequency sampling can capture transient fault characteristics that are often missed by traditional methods. Multidimensional data integration addresses the heterogeneity of sensor data and creates conditions for refined status analysis.

[0046] S202, performing nonlinear feature mapping on the original multidimensional data set to construct a state feature vector set, and decomposing each state feature vector into an electrical sub-vector and a mechanical sub-vector to separately capture electrical characteristics and mechanical vibration characteristics; Specifically, the original multidimensional data set can be input into the kernel function space, and nonlinear mapping can be performed through the radial basis function kernel to generate a high-dimensional state feature space; Principle of constructing kernel function space The original multidimensional dataset contains quantized values ​​of current, voltage, temperature, and vibration signals, with dimensions up to dozens of dimensions (e.g., 12 dimensions for current harmonic components, 8 dimensions for voltage ripple characteristics, 5 dimensions for temperature gradients, and 18 dimensions for vibration spectrum energy). Complex nonlinear coupling relationships exist between these parameters (e.g., sudden current changes can lead to increased vibration). Direct linear analysis loses key correlation information. The essence of the kernel function space is to project the original data into a higher-dimensional implicit space through mathematical transformations, making the originally nonlinearly separable data linearly separable. In specific implementation, the system creates an independent kernel mapping unit (KMU) for each charging pile. This unit receives all parameter values ​​for a time slice (e.g., a 500-millisecond window) in the original dataset and treats it as a multidimensional coordinate point.

[0047] How the Radial Basis Function Kernel Works The Radial Basis Function (RBF) kernel is a core mapping tool. Its mathematical essence is to calculate the "similarity distance" between data points. This kernel function has two key parameters: the kernel width coefficient (gamma, γ) and the truncation threshold (epsilon, ε). γ controls the extent of the mapped feature space (typical value 0.01-0.1), and ε determines whether to ignore small correlations (typical value 1e-5). For example, when processing the correlation between current harmonic distortion (THD) and chassis vibration acceleration, the RBF kernel calculates the "attraction strength" between the two in high-dimensional space: if a sudden increase in THD from 5% to 15% is accompanied by a 30dB increase in vibration energy in the 200-400Hz frequency band, the two events will be pulled closer to the same cluster in the mapped feature space. The mapping process is implemented through matrix operations: a similarity matrix (SM) is constructed for the original data point set, and the matrix elements SM_ij represent the kernel function output values ​​of the i-th and j-th time slice data, ultimately generating a high-dimensional state feature space (HDSFS) with a dimension much higher than the original data (for example, from 43 dimensions to 1000 dimensions).

[0048] Physical meaning guarantee of feature space To prevent high-dimensional space from becoming a mathematical game, the system ensures interpretability by injecting physical constraints. For example: The current-voltage characteristic subspace is forced to satisfy the deformation constraints of Kirchhoff's law; The vibration characteristic subspace must conform to the mechanical resonance frequency formula (e.g., an approximate expression of f_res = 1 / (2π√(k / m))); The mapping of temperature gradient and electrical loss must satisfy the first law of thermodynamics.

[0049] In the resulting HDSFS, each dimension corresponds to a nonlinear interaction pattern (such as a "high-harmonic structural resonance" pattern), laying the foundation for subsequent feature decoupling. The number of feature space dimensions is dynamically controlled by the Automatic Relevance Determination (ARD) module and is typically maintained between 20 and 30 times the original dimensionality.

[0050] Perform orthogonal feature decoupling in the high-dimensional state feature space and use the FastICA algorithm to separate electrical and mechanical features; Necessity and Challenges of Orthogonal Decoupling During charging pile operation, electrical signals (such as DC bus pulsation) and mechanical vibrations (such as fan bearing wear) often interfere with each other. For example, 2000Hz current oscillations caused by IGBT switching can excite chassis panel resonance, which in turn affects the heat sink temperature field. In the HDSFS generated in step 1, these coupled features remain mixed within the same hyperplane. The goal of orthogonal feature decoupling is to project the two feature types into mutually perpendicular subspaces, ensuring that the dot product of the electrical and mechanical eigenvectors is zero (orthogonality condition). To achieve this goal, the system uses the FastICA algorithm (Fast Independent Component Analysis), which excels at separating statistically independent source components from mixed signals.

[0051] Parameterized implementation of the FastICA algorithm The core of FastICA is to maximize non-Gaussianity, using Negentropy (J) as the optimization objective. The algorithm contains four key parameters: Iteration tolerance error (Tolerance, Tol=1e-6): controls the convergence accuracy; Maximum number of iterations (MaxIter=200): prevents infinite loops; Nonlinear function type (G(u)=tanh(u)): used to measure non-Gaussianity; Whitening preprocessing flag (Whitening=True): eliminates the correlation between features.

[0052] Take the example of separating cooling fan vibration and PFC circuit noise: Input: 1000-dimensional mixed features in HDSFS; Process: The algorithm continuously adjusts the unmixing matrix W through fixed-point iteration; Output: Two independent component vectors - electrical component (including switching ripple and common mode interference) and mechanical component (including bearing rotation fundamental frequency and blade passing frequency).

[0053] Physical verification of the decoupling effect To ensure the credibility of the separation results, the system sets up a triple verification mechanism: Mutual Information Test (MI<0.05): Verify the independence of electrical / mechanical vectors; Energy retrospective comparison: The total energy of the electrical vector must have an error of less than 3% compared with the meter reading; Fault injection test: When a fan stuck fault is artificially created, the bearing characteristic energy in the mechanical vector should increase by 20dB while the electrical vector remains unchanged.

[0054] The verified decoupling features will be marked as Electrical Independent Component Set (EICS) and Mechanical Independent Component Set (MICS) for use in subsequent sub-vector construction.

[0055] Apply information entropy weighting to the decoupled electrical features to generate electrical sub-vectors with physical interpretability; The Engineering Significance of Information Entropy Weighting The decoupled electrical signature (EICS) may contain dozens of independent components (such as switching frequency components, ground return noise, battery backflow current, etc.), but not all components are equally important for fault prediction. The goal of information entropy weighting is to quantify the "diagnostic value" of each feature. The formula for calculating entropy (H) is based on Shannon's theory and reflects the degree of chaos in the feature. For example: The current harmonic entropy value is low during stable operation (H≈0.2): strong regularity and low diagnostic value; The entropy value of the current mutation when the contactor contacts are burned is high (H≈0.9): the abnormal information is rich and the weight should be increased.

[0056] Calculation process of weighting factors The system implements a three-step weighted strategy: Characteristic entropy calculation: Sample entropy (SampEn) of 500 time windows is calculated for each electrical component. Parameter settings: embedding dimension m = 2, similarity tolerance r = 0.2 times the standard deviation; Weight distribution: Using the reciprocal normalization method, the weight of the i-th feature W_i = (1 - H_i) / Σ(1 - H_j); Physical meaning injection: Superimpose expert rule correction factors (RCFs). For example, the RCF of the DC output voltage peak-to-peak ripple (Vpp) is constant at 1.2 (because it is sensitive to capacitor failure), while the RCF of the common-mode current is only 0.8.

[0057] Structured output of electrical subvectors The weighted feature set is organized into an Electrical Sub-Vector (ESV), whose data structure consists of three layers: Core parameter layer: 8-10 high-weight features (e.g., DC link voltage fluctuation (weight 0.15), IGBT junction temperature estimation slope (weight 0.12)); Auxiliary indicator layer: 15-20 low- to medium-weight features (e.g., PFC efficiency offset (weight 0.06)); Metadata header: Contains sampling timestamp (Timestamp, 64-bit integer), charging pile ID (CP_ID, 16-byte string), and weighted version number (Weight_Ver=2.3).

[0058] The final ESV length is controlled within 30 dimensions, ensuring real-time performance by sacrificing a small amount of secondary information. For example, in a 150kW charging station, the ESV update period is 200 milliseconds.

[0059] Perform wavelet coherence analysis on mechanical characteristics, extract the phase coupling coefficient between the vibration signal and the cooling fan speed, and construct the mechanical sub-vector; Technical Motivation for Wavelet Coherence Analysis The primary source of mechanical vibration in charging piles is the cooling fan (contributing >60%), but vibration from the gearbox and contactor can also interfere. Wavelet Coherence Analysis (WCA) is used to quantify the correlation between the vibration signal and the fan speed signal in the time-frequency domain. Compared to traditional FFTs, this wavelet transform uses the Morlet wavelet as the basis function, with parameters set as follows: center frequency ω0 = 6 (optimal time-frequency resolution balance) and scales = 64 (covering the 10-2000 Hz frequency band).

[0060] Phase coupling coefficient extraction process The analysis process specifically includes: Signal preprocessing: synthesize the vector magnitude (VectorMagnitude, VM = √(x 2 +y 2 +z 2 )); Fan speed synchronization: Get the pulse frequency of the fan PWM control signal through the CAN bus (e.g. 2500 RPM corresponds to 41.67Hz); Coherence calculation: Based on the continuous wavelet transform (CWT), the time-frequency domain coherence value is calculated: Coh(f,t) = |S(fan_cwt * vib_cwt*)| 2 / (S(|fan_cwt| 2 ) * S(|vib_cwt| 2)), where S is the smoothing operator and * represents the complex conjugate.

[0061] Key coefficient extraction: Calculate the phase coupling coefficient (PCC) at the fan fundamental frequency (f0) and the blade pass frequency (BPF = f0 × number of blades): PCC_f0 = max( |Coh(f0,t)| × cos(Δφ(t)) ) t∈[t-0.5s,t]. Δφ(t) is the real-time phase difference between the vibration signal and the speed signal.

[0062] Engineering packaging of mechanical sub-vectors The Mechanical Sub-Vector (MSV) built based on PCC includes three types of indicators: Main coupling indicators: fundamental frequency coupling coefficient PCC_f0 (normal range 0.7-0.9), blade passing frequency coupling coefficient PCC_BPF; Fault characteristic index: coherent energy ratio at bearing fault characteristic frequency (BPFI / BPFO); Auxiliary verification values: Kurtosis (K>3.5 warning) and envelope spectrum entropy of the three-dimensional vibration signal.

[0063] Example of a typical MSV structure (unit: dimensionless): [PCC_f0=0.83, PCC_BPF=0.76, BPFI_Energy=0.12, Kurtosis_X=3.2, EnvEntropy=0.45].

[0064] This vector is updated every 500 milliseconds, and when PCC_f0<0.6, a "fan out of step" warning is triggered.

[0065] The electrical sub-vector and the mechanical sub-vector are combined into a state feature vector set through the feature cascade layer, and a feature validity flag is added.

[0066] Structural design of the cascade layer The Feature Concatenation Layer (FCL) does not simply concatenate data, but performs intelligent alignment and compression: Timeline alignment: Since the update period of ESV (200ms) is different from that of MSV (500ms), cubic spline interpolation is used to resample ESV to 500ms timestamps; Dimensionality reduction: Principal component analysis (PCA) was used to reduce ESV from 30 dimensions to 15 dimensions and MSV from 8 dimensions to 5 dimensions, retaining 95% of the original information. Data normalization: Z-score normalization (mean = 0, variance = 1) was used for ESV, and Min-Max normalization (range [0, 1]) was used for MSV.

[0067] Generation of state feature vector sets The structure of the State Feature Vector (SFV) after cascading is as follows: Electrical segment (15 dimensions): includes weighted DC ripple, IGBT temperature trend, insulation resistance estimation, etc. Mechanical segment (5D): including PCC_f0, PCC_BPF, bearing fault indicators, etc. Cross-features (2D): such as the interaction term of “current harmonic distortion rate × fundamental frequency vibration amplitude”.

[0068] The total dimension of each SFV is 22, and the example segments are: [0.32, -1.2, 0.05, ..., 0.83, 0.76, 1.2, 0.45, 0.12].

[0069] Intelligent setting of feature validity flags To cope with sensor failure or communication packet loss, the system adds a Feature Validity Flag (FVF) to each SFV. The flag is an 8-bit binary number: Bit0: Current sampling validity (1 valid); Bit1: Voltage sampling validity; Bit2: Temperature sensor status; Bit3: Vibration signal signal-to-noise ratio (SNR>40dB=1); Bit4: Fan speed signal synchronization status; Bit5: FastICA convergence flag; Bit6: Entropy weighting factor credibility (H calculation sample number>300=1); Bit7: Wavelet coherence spectrum resolution flag.

[0070] For example, when the vibration sensor falls off, Bit 3 is set to 0 and the "mechanical feature failure" alarm is triggered, and the system automatically switches to pure electrical analysis mode.

[0071] The raw data is mapped into a high-dimensional feature space using a kernel function, and independent component analysis (ICA) is used to decouple electrical and mechanical features. The electrical sub-vector includes parameters such as current harmonic distortion, while the mechanical sub-vector focuses on vibration spectrum features, enabling targeted extraction of different physical properties. Feature decoupling prevents mutual interference between electrical and mechanical signals, improving the purity of feature representation. The separated sub-vectors facilitate subsequent, targeted analysis of specific fault types, such as electrical overload or mechanical looseness.

[0072] S203, using a distributed edge computing framework to perform real-time fusion analysis on the electrical sub-vector and the mechanical sub-vector to generate a preliminary abnormality probability set; Specifically, a lightweight graph neural network can be deployed on edge computing nodes, with electrical subvectors as node attributes and charging pile topology connections as edge weights to construct an electrical state graph structure. The Edge Computing Nodes (ECN) deployed at charging stations utilize embedded hardware with neural network acceleration capabilities (such as the NVIDIA Jetson TX2 module). Their memory capacity typically includes 8GB of LPDDR4 memory and a computing power of 1.3 TFLOPS. Each node manages a group of physically adjacent charging piles (e.g., 6-8), forming a local computing cluster. The Lightweight Graph Neural Network (L-GNN) utilizes a simplified GraphSAGE architecture, consisting of two aggregation layers and one fully connected layer (FCL). The network inputs the electrical sub-vectors (ESVs) of each charging pile within its jurisdiction as attributes of the graph nodes. The ESVs are typically 12-dimensional (e.g., including quantitative features such as current harmonic distortion (THD), voltage sag depth (SagDepth), and insulation resistance (IR). The connection relationship (Edge) between nodes is determined by the topological connection relationship (TCR) of the charging piles. If two charging piles share the same power distribution cabinet (PDC) or are in the same cable loop, a bidirectional edge is established. The edge weight (EW) is calculated based on the strength of the electrical coupling: when the rated power difference between the two piles is less than 20kW, the weight is set to 0.9 (strong coupling); when the power difference is 20-50kW, the weight is 0.6 (medium coupling); and when the power difference is greater than 50kW, the weight is 0.3 (weak coupling).

[0073] The graph structure is constructed dynamically. When new charging stations are added or distribution cabinets are modified, edge nodes automatically identify topology changes and update the adjacency matrix (AM) by scanning Modbus-TCP broadcast packets on the local area network. The first-layer aggregation operation of the L-GNN performs a "weighted neighbor averaging" operation: for a target node (e.g., charging station #05), the ESV data of its first-order neighbors (e.g., stations #03, #04, and #07) are collected and weighted by edge weight (EW) (e.g., a weight of 0.9 for station #03 and 0.6 for station #04). The weighted result is concatenated with the target node's own ESV and input into a 128-dimensional fully connected layer (FCL) for a nonlinear transformation. The second-layer aggregation extends to second-order neighbors, ultimately outputting a 64-dimensional electrical state embedding vector (ESEV). This process is performed every 200 milliseconds (ms) to ensure responsiveness to rapid events such as grid fluctuations.

[0074] The key optimization lies in reducing the computational load. Parameter quantization compresses the L-GNN weights from 32-bit floating point to 8-bit integers (INT8), reducing the model size by 75%. Sparse matrix multiplication is used in the aggregation phase to process only non-zero edge weights (for example, if a pile is connected to only three neighbors, the zero-weight calculation for the remaining nodes is skipped). Experiments show that processing a single graph inference on an edge node takes no more than 15 milliseconds (ms), meeting real-time requirements. The resulting Electrical State Graph Structure (ESGS) contains three core data items: node attributes (ESEV), edge connections (AM), and edge weights (EW), providing a structured representation of the electrical state for subsequent fusion analysis.

[0075] Perform spatiotemporal convolution on the mechanical sub-vectors to aggregate the vibration characteristics of adjacent charging piles in the spatial dimension, and apply causal convolution in the temporal dimension to generate a mechanical state evolution tensor; The mechanical sub-vector (MSV) contains vibration signal features (such as the root mean square (RMS) acceleration of the X / Y / Z axes, the bearing fault characteristic frequency amplitude (BPFO_amp), and the cooling fan fundamental frequency component (Fan_Fundamental). Spatio-temporal convolution (SPC) consists of a cascade of spatial and temporal convolutions. Spatial convolution first aggregates the features of adjacent charging piles in the spatial dimension: With the target charging pile (e.g., #05) as the center, physically adjacent charging piles (e.g., #04 and #06) within a 5-meter radius are selected. Their MSV data are weighted by the inverse of their distance (e.g., if #04 is 3 meters away, the weight is 1 / 3, which is approximately 0.33; if #06 is 4 meters away, the weight is 1 / 4, which is 0.25). These weighted aggregated features are then concatenated with the target charging pile's own MSV to form the Spatial Enhanced Mechanical Vector (SEMV).

[0076] Causal convolution is used in the temporal dimension to process continuous time series data. The time window length is set to 64 samples (corresponding to 6.4 seconds of data at a sampling rate of 10Hz). The convolution kernel size is 3×1 (width×height) with a stride of 1 to ensure that the output length is the same as the input. The core constraint of causal convolution is to prevent future data leakage: the output at time t depends only on the input at time t-2, t-1, and t (due to the kernel size of 3). Two layers of causal convolution are stacked: the first layer uses 16 kernels to extract short-term features (such as impact events), and the second layer uses 32 kernels to capture long-term trends (such as fan wear and degradation). Each layer is followed by a ReLU activation function and max pooling with a pooling kernel size of 2×1, compressing the temporal dimension by 50%.

[0077] The resulting Mechanical State Evolution Tensor (MSET) has dimensions of [time steps × feature channels × spatial nodes]. For example, when processing five adjacent piles, the output tensor has a shape of 32 × 32 × 5 (32 time steps, 32-dimensional feature channels, and 5 spatial nodes). This tensor improves fault sensitivity through three characteristics: Spatial correlation: Compare the vibration of a single pile with that of neighboring piles. If the vibration amplitude of the fan at pile #05 suddenly increases to more than twice the average value of its neighbors, it is marked as abnormal. Time dependence: Causal convolution identifies a sustained upward trend in vibrational energy entropy (EE) (e.g., EE increases >5% over 10 consecutive windows). Multi-scale features: Shallow kernels capture instantaneous impacts (such as 2kHz transient pulses caused by loose bolts), while deep kernels identify slow degradation (such as 80Hz sideband energy growth caused by bearing wear).

[0078] Design a dual-channel attention fusion mechanism, in which the electrical channel focuses on current mutation events and the mechanical channel focuses on abnormal vibration frequency bands, and outputs a cross-modal attention weight matrix; The Dual-Channel Attention Fusion Mechanism (DCAFM) consists of two independent branches: the electrical channel (EC) and the mechanical channel (MC). The electrical channel's Attention Generation Module (AGM) specifically detects current mutation events (CMEs): an event flag is triggered when the current change rate dI / dt exceeds a threshold (e.g., 100A / ms). For these events, three key metrics are calculated: Mutation Duration (MD); Mutation Amplitude (MA); Mutation Phase (MP, distinguishing between charging start / end / steady state phase).

[0079] These three metrics are mapped to electrical attention weights (EAWs) using a three-layer perceptron (MLP), ranging from 0 to 1. For example, if a sudden change of 300 A / ms (MA=300) occurs during the charging termination phase and lasts for 50 ms (MD=50), the EAW is 0.92 (high risk).

[0080] The AGM of the mechanical channel focuses on the abnormal vibration frequency band (AVFB). Four typical fault frequency bands are preset: Bearing fault frequency band (BPFI / BPFO, usually 500Hz-2kHz); Gear mesh frequency (GMF, usually 1kHz-4kHz); Electromagnetic noise frequency band (EM_Noise, >5kHz); Structural resonance frequency band (Resonance, characteristic frequency to be identified).

[0081] A short-time Fourier transform (STFT) is performed on the MSET to calculate the energy ratio (ER) of each frequency band. When the ER of a frequency band exceeds the historical baseline by 30%, it is considered abnormal activation. The mechanical attention weight (MAW) is weighted by the degree of frequency band abnormality: if both the BPFO band ER exceeding the standard by 40% and the GMF band ER exceeding the standard by 25% are detected, the MAW is 0.85.

[0082] The final output, the Cross-Modality Attention Weight Matrix (CMAWM), is the Cartesian product of EAW and MAW. The matrix dimensions are N×M (N electrical event types × M mechanical frequency band types). Each element W_ij represents the association weight between electrical event i and mechanical frequency band j, adjusted by a learnable parameter matrix. For example: The weights of the current mutation (EC event type 1) and the bearing fault frequency band (MC frequency band 1) are initialized to 0.7 (strong correlation); The weights of voltage flicker (EC type 2) and electromagnetic noise frequency band (MC frequency band 3) are initialized to 0.4 (weak correlation); The matrix real-time update cycle is 100 milliseconds (ms), providing dynamic priority guidance for subsequent fusion.

[0083] Based on the attention weight matrix, a gated recurrent unit is used to fuse the electrical state graph and the mechanical evolution tensor to generate a multi-dimensional state fusion vector; The Gated Recurrent Unit (GRU) uses a two-layer stacked structure with a hidden layer dimension of 128. Its input is a joint representation of the electrical state graph structure (ESGS) and the mechanical state evolution tensor (MSET): Electrical part: Flatten the electrical state embedding vector (ESEV) output by the graph neural network into a 64-dimensional vector; Mechanical part: Calculate the mean over time of MSET to generate a 32-dimensional feature vector; Concatenate the two to obtain a 96-dimensional raw fusion vector (RFV).

[0084] The fusion process introduces the attention weight matrix (CMAWM) for dynamic modulation. It is divided into three steps: Step 1 - Electrical feature weighting: According to the type of electrical event currently detected (such as CME type 3), the corresponding row vector W_i* (dimension 1×M) is extracted from CMAWM and the 64-dimensional features of ESEV are scaled.

[0085] Step 2 - Mechanical feature filtering: Based on the activation status of the mechanical abnormal frequency band (such as the activation of bands 2 and 4), the corresponding column vector W_*j (dimension N×1) is extracted from CMAWM and multiplied element-wise with the 32-dimensional feature of MSET (Element-wise Multiplication).

[0086] Step 3 - Gated Fusion: The weighted electrical features are concatenated with the filtered mechanical features and then fed into the GRU unit. The GRU's update gate determines how much of the historical state is retained, while the reset gate controls the proportion of new features written. For example, if mechanical vibration suddenly increases (MSET energy rises by 40%) and there is no sudden change in electrical vibration, the reset gate opens 70% of the channels to write new data.

[0087] The output of the GRU is compressed to 48 dimensions through the fully connected layer to generate a Multi-Dimensional State Fusion Vector (MDSFV). This vector has three major characteristics: Temporal Memory: The GRU hidden state retains the state evolution trajectory of the past 120 seconds; Cross-modal correlation: Current mutation events (e.g., dI / dt = 150A / ms) and high-frequency bearing vibrations (>1kHz energy) are mapped to similar regions in vector space. Enhanced fault sensitivity: Experiments show that the detection sensitivity of cooling fan stuck faults has increased from 73% for a single mode to 92% after fusion. The vector is updated every 200 milliseconds (ms) and serves as the input baseline for anomaly detection.

[0088] The multi-dimensional state fusion vector is input into the anomaly detection forest, and a preliminary anomaly probability set of each dimension is output.

[0089] The Anomaly Detection Forest (ADF) consists of 100 deep randomized isolation trees (iTrees). Each tree randomly selects four feature dimensions of the MDSFV (e.g., dimensions 12, 25, 33, and 41) for node splitting. The splitting threshold is randomly generated, using criteria such as whether the value of dimension 12 is greater than 0.7. ADF uses an unsupervised training model, requiring only normal state data (NSD) for training. During training, 5% Gaussian noise is injected into each iTree to improve model robustness.

[0090] The calculation of anomaly probability is carried out in three steps: Step 1 - Path length calculation: For the input MDSFV vector, count the number of edges it passes through from the root node to the leaf node in each iTree (i.e., path length, PL).

[0091] Step 2 - Anomaly Score Conversion: Based on the average PL of all iTrees in the forest, the anomaly score (AS) is calculated using the formula S = 2^(-E(PL) / c(n)), where c(n) is the path length normalization factor (for example, c = 12.8 when n = 1000).

[0092] Step 3 - Dimensional Probability Generation: Using Feature Contribution Decomposition (FCD), we perturb the target dimension d (e.g., the vibration energy dimension 25) by ±10% to generate two new vectors. The difference in AS before and after the perturbation, ΔAS, is then compared. A larger ΔAS indicates a higher anomaly contribution in that dimension. This is ultimately mapped to the Dimensional Anomaly Probability (DAP), using the formula P_d = sigmoid(5×ΔAS).

[0093] The final output of the Preliminary Anomaly Probability Set (PAPS) contains 48 probability values ​​(aligned with the MDSFV dimension). Each probability value P_d represents the anomaly confidence level of the corresponding feature dimension (range 0-1). The system sets two thresholds: Warning Threshold (WT=0.65): When the single dimension P_d>0.65, a yellow warning is marked; Alarm Threshold (AT=0.85): A red alarm is triggered when P_d of three or more dimensions is greater than 0.85 or P_d of a single dimension is greater than 0.95.

[0094] For example, when the "current harmonic THD dimension probability = 0.92" and the "bearing vibration sideband dimension probability = 0.88" are detected, an alarm event for charging module IGBT aging is generated.

[0095] A lightweight analysis model is deployed at the edge node, processing electrical features through graph neural networks and mechanical features through spatiotemporal convolution. This is then integrated with an attention mechanism for cross-modal fusion. This process leverages the low latency of edge computing to achieve millisecond-level anomaly detection. The distributed architecture meets real-time requirements and avoids cloud-based transmission delays. Multimodal fusion enhances comprehensive detection, enabling simultaneous identification of complex faults such as contactor arcing (electrical) and fan imbalance (mechanical).

[0096] S204, performing multi-objective optimization on the preliminary abnormal probability set, applying a reinforcement learning algorithm to optimize real-time responsiveness and fault prediction accuracy, and generating an optimized state indicator set; Specifically, a dual-objective reward function can be constructed, where real-time responsiveness rewards are associated with processing delay, and prediction accuracy rewards are associated with fault detection F1 score; The system first defines a dual-objective reward function as the evaluation benchmark for reinforcement learning. This function contains two core reward items: Real-time Responsiveness Reward (RRR): Its calculation is directly related to Processing Latency (PL). PL refers to the time difference (in milliseconds) between sensor data acquisition and the generation of an anomaly probability. RRR is designed as a negative correlation function. For example, when PL ≤ 5ms, the reward is +10 points. The reward decreases linearly by 2 points for every 1ms increase in PL, and returns to zero when PL > 20ms. This design forces the system to prioritize low latency.

[0097] Prediction Accuracy Reward (PAR): This is calculated based on the Fault Detection F1-Score (FDF1). FDF1 is the harmonic mean of precision (PREC) and recall (REC) (formula: 2 × (PREC × REC) / (PREC + REC)). PAR is designed as a positive correlation function. For example, if FDF1 ≥ 0.95, the reward is +15 points. For every 0.01 decrease in FDF1, the reward decreases by 3 points. If FDF1 < 0.8, the reward is zero.

[0098] The reward function needs to resolve multiple conflicting objectives: for example, a high sampling frequency can improve FDF1 (increase PAR) but increase the computational load (reduce RRR). The system introduces a dynamic weight factor (DWF) for arbitration: During peak charging periods (identified through historical data), the RRR weight is set to 0.7 and the PAR weight is set to 0.3 to ensure that response speed is prioritized.

[0099] In the equipment aging stage (determined by the cumulative operating time), the PAR weight is set to 0.8 and the RRR weight is set to 0.2, focusing on the accuracy of fault prediction.

[0100] The final reward (TR) is calculated as: TR = DWF_RRR × RRR + DWF_PAR × PAR. This formula transforms the multi-objective optimization problem into a single-objective maximization problem.

[0101] The reward function is linked to physical constraints. For example, if the internal temperature of a charging station exceeds a safety threshold (e.g., 65°C), the RRR weight is forcibly set to 1.0, triggering Downclocking Operation Mode (DOM). This temporarily sacrifices some accuracy to reduce CPU load and heat dissipation. All weighting strategies and thresholds are stored in a Configurable Rule Base (CRB), allowing operators to adjust them as needed.

[0102] Design a deep Q-network architecture that uses anomaly probability sets as state inputs and feature sampling frequencies as action space, and estimates the value function through a two-stream network. The Deep Q-Network (DQN) uses a dual-stream architecture (DSA) to process state input: The State Input (SI) is the Preliminary Anomaly Probability Set (PAPS), which includes dimensions such as Electrical Anomaly Probability (EAP), Mechanical Anomaly Probability (MAP), and Thermal Anomaly Probability (TAP). For example, a typical PAPS is [EAP=0.82, MAP=0.15, TAP=0.03].

[0103] Action Space (AS) is defined as a combination of feature sampling frequencies: Current / voltage sampling frequency (IV-Sampling Frequency, IV-SF): Three levels are available: 10kHz, 20kHz, and 50kHz.

[0104] Vibration signal sampling frequency (Vib-Sampling Frequency, Vib-SF): 1kHz, 2kHz, 5kHz three levels are available.

[0105] Temperature sampling frequency (Temp-Sampling Frequency, Temp-SF): Fixed to 1Hz (due to slow temperature changes).

[0106] The specific structure of the dual-stream network is as follows: Real-time Response Stream (RRS): Input is PAPS, and features are extracted through two fully connected layers (number of neurons: 128 → 64). The output is a response value estimate (RVE), which represents the expected delayed reward for taking a certain action.

[0107] Accuracy Prediction Stream (APS): Inputs PAPS, captures temporal dependencies through a long short-term memory network (LSTM unit number = 32), and outputs an accuracy value estimate (AVE), which represents the expected accuracy reward for taking a certain action.

[0108] The dual-stream outputs are concatenated in the fusion layer (FL) and passed through a fully connected layer (number of neurons = 32) to generate the final Q-value matrix (QVM). The matrix dimension is 3 (IV-SF option) × 3 (Vib-SF option) = Q-values ​​for 9 actions.

[0109] Network training utilizes an Experience Replay (ER) mechanism. The system stores historical state transition data (e.g., state S_t="[0.75, 0.2, 0.1]", action A_t="IV-SF=20kHz, Vib-SF=2kHz", reward R_t=+8, and new state S_t+1="[0.68, 0.18, 0.05]") in a replay buffer (RB). 32 sets of data (batch size=32) are randomly sampled each time to update network parameters to avoid training oscillations caused by timing correlation. The target network synchronizes weights with the main network every 1000 steps to improve stability.

[0110] Adopting a non-dominated sorting strategy, the action strategy is selected on the Pareto optimal frontier, and the current sampling frequency and vibration analysis granularity are dynamically adjusted; The non-dominated sorting strategy (NDSS) is used to resolve action selection under multi-objective conflicts: For the 9 action options in the current Q value matrix, calculate their dual-objective scores: Target 1: Real-time responsiveness score (mapped from RVE to latency reduction percentage, e.g. RVE=15 → latency reduction of 12%).

[0111] Objective 2: Prediction accuracy score (mapped from AVE to F1 score improvement, e.g., AVE=20 → F1 score improvement of 0.05).

[0112] Perform Pareto Ranking (PR): The first level (Pareto optimal frontier) is defined as all actions that are not dominated by other actions (dominance). For example, action A (decreases latency by 15%, increases F1 by 0.03) dominates action B (decreases latency by 10%, increases F1 by 0.02), but action C (decreases latency by 5%, increases F1 by 0.06) does not dominate action A.

[0113] Second level: A collection of actions that are dominated by the first level actions but dominate other lower level actions, and so on.

[0114] Rules for selecting actions on the Pareto front: Crowding Distance Calculation (CDC): For each action in the frontier, calculate the sum of its Euclidean distances to its neighboring actions in the target space. A larger distance indicates a sparser (and more optimal) decision space for that action.

[0115] Dynamic Preference Injection (DPI): Adjust preference weights based on the runtime scenario: If the charging power is > 150kW (fast charging mode), select the high-responsiveness action with the largest congestion distance (such as IV-SF=50kHz).

[0116] If mechanical abnormality is detected three times in a row (MAP > 0.6), a high-precision action is selected (e.g., Vib-SF = 5kHz).

[0117] The final output is the optimized action policy (OAP), for example: "IV-SF=50kHz,Vib-SF=2kHz".

[0118] Action execution is linked to Analysis Granularity (AG): When a high vibration sampling frequency (such as 5kHz) is selected, Fine Wavelet Packet Decomposition (FWPD) is enabled simultaneously to decompose the vibration signal into 32 sub-bands (originally 16), improving the accuracy of abnormal frequency band identification.

[0119] When the current sampling frequency drops to 10kHz, Dimensionality Reduction Feature Extraction (DRFE) is triggered, using principal component analysis (PCA) to compress the electrical feature dimensions from 128 to 64, reducing computational complexity. All granularity adjustments are recorded in the Policy Execution Log (PEL) for auditing.

[0120] The network parameters are updated through the proximal strategy optimization algorithm, and the optimized state indicator set is output.

[0121] The Proximal Policy Optimization (PPO) algorithm is responsible for network updates. Its core is the Clipped Objective Function (COF): Calculate the probability ratio of the new and old strategies: Old Policy (OP): The probability of the network selecting the historical action A_t before the update (e.g., probability = 0.65).

[0122] New Policy (NP): The probability of the network selecting the same action A_t after the update (e.g., probability = 0.72).

[0123] Probability Ratio (PR) = NP / OP = 0.72 / 0.65 ≈ 1.108.

[0124] Construct the basic objective function: Advantage Function (AF) × PR. AF is calculated using Generalized Advantage Estimation (GAE), which integrates immediate rewards and future values ​​(with a discount factor of γ = 0.99).

[0125] Apply pruning mechanism to prevent policy mutation: Set the clipping threshold (CT) = ±0.2.

[0126] If PR > 1.2 (i.e., NP is more than 20% higher than OP), force PR in the objective function to be replaced with 1.2.

[0127] If PR < 0.8 (i.e., NP is more than 20% lower than OP), PR is forcibly replaced with 0.8.

[0128] For example, when PR = 1.108, the threshold is not exceeded, and the objective function value = AF × 1.108; if PR = 1.25, the objective function value is calculated as 1.2. This mechanism ensures that the step size of each parameter update is controlled to avoid training divergence.

[0129] Output Optimized State Indicator Set (OSIS): Metrics Generation: Apply the optimized DQN to the real-time data stream and output the state evaluation results corresponding to the action strategy, including: Electrical Health Index (EHI): ranges from 0 to 100, with >80 indicating normal.

[0130] Mechanical Vibration Severity (MVS): ranges from 0 to 10, with <2 being normal.

[0131] Composite Risk Score (CRS): ranges from 0 to 1, with a score <0.3 triggering an alert.

[0132] Dynamic calibration: OSIS is backtested monthly using historical fault data. If FDF1 continuously falls below 0.85, the Network Retraining Procedure (NRP) is automatically initiated to update the DQN weights. Finally, OSIS is pushed to the cloud diagnostic platform via a message queue (e.g., MQTT).

[0133] A dual-objective reward function is constructed, dynamically balancing detection speed and accuracy through deep reinforcement learning. The algorithm automatically adjusts sampling frequency and analysis granularity, selecting the optimal strategy from the Pareto optimal solution set to achieve self-optimization of detection performance. This addresses the conflict between response speed and accuracy encountered in traditional methods and adapts to diverse operating conditions. Its autonomous learning feature reduces reliance on manual parameter tuning and improves system adaptability.

[0134] S205 , performing spatiotemporal consistency verification and confidence screening on the optimization status indicator set, determining a final operating status level, and outputting a real-time monitoring alarm.

[0135] Specifically, a spatiotemporal consistency check window can be constructed, and an exponentially weighted moving average can be used to filter out instantaneous interference in the time dimension, while the Mahalanobis distances of similar indicators of adjacent charging piles can be compared in the spatial dimension. After receiving the optimized state indicator set, the system first constructs a spatio-temporal consistency check window (STCCW). This window is a dynamic data buffer that covers the most recent 20 consecutive seconds of indicator data in the temporal dimension (the window length is configurable) and spatially associates the current charging station with its three physically adjacent charging stations (based on a preset topological map). The temporal dimension is processed using the Exponential Weighted Moving Average (EWMA) algorithm, the core of which is to assign a weight decay factor (WDF) to the indicator value at each time point within the window. The WDF is set to 0.2 (indicating a weight of 0.8 for new data and an exponential decay for historical data). The EWMA value is calculated independently for each indicator (such as the DC bus voltage fluctuation rate). For example, if a current harmonic distortion indicator experiences a transient spike within a window (possibly caused by grid interference), EWMA reduces the weight of older data (for example, the weight of data from 10 seconds ago is only 10.7% of the initial value), limiting the impact of the spike to a local area. It then outputs a smoothed trend value (STV), effectively filtering out transient noise. The system also records the standard deviation (SD) of each indicator within the window for subsequent confidence assessment.

[0136] The core of spatial dimension verification is calculating the Mahalanobis distance (MD). For the same indicator (e.g., the "IGBT module temperature rise rate" of all adjacent charging stations), the system collects the STV data for each charging station within the current verification window to form a multidimensional dataset. First, the global covariance matrix (GCM) for this indicator is calculated. This matrix, obtained through training with historical normal data, represents the correlation between the indicators of each charging station (e.g., the strength of the correlation between the temperature rise of charging station A and the cooling fan speed of charging station B). For a specific indicator value of the current target charging station (e.g., the temperature rise rate STV = 1.5°C / min), the Mahalanobis distance between it and the mean vector (MV) of the indicator of the adjacent charging station group is calculated: MD = √[(target value - group mean) × inverse covariance matrix × (target value - group mean) transposed]

[0137] This distance is dimensionless. If the MD exceeds a preset threshold (e.g., 3.0), it indicates that the pile's performance deviates significantly from the behavior of the adjacent pile group (possibly due to a fault). For example, if a pile's cooling fan is malfunctioning, its temperature rise rate MD may reach 4.2, while a normal pile group's MD is typically less than 1.5. The spatial calibration output is the spatial deviation coefficient (SDC), defined as the ratio of the MD to the threshold.

[0138] The spatiotemporal verification results need to be dynamically integrated. The system generates a spatiotemporal consistency score (STCS) for each indicator. The calculation formula is: STCS = temporal smoothing factor × (1 / SD within the window) + spatial alignment factor × (1 / SDC).

[0139] The Time Smoothness Factor (TSF) is calculated as the inverse of the EWMA residual (the difference between the actual value and the STV). Smaller residuals indicate larger TSFs. The Spatial Alignment Factor (SAF) is calculated as the inverse of the normalized SDC. The STCS value ranges from 0 to 1.0, with values ​​closer to 1 indicating greater temporal and spatial consistency (for example, a voltage ripple coefficient has an STCS of 0.92). Indicators below a threshold (e.g., 0.6) are marked as "temporal and spatial anomaly candidates," triggering subsequent in-depth verification. All computations are performed at the edge nodes in less than 50 milliseconds, ensuring real-time performance.

[0140] Design a confidence screening funnel, where the first layer removes highly volatile indicators, and the second layer retains stable features to generate a subset of high-confidence indicators; The Confidence Screening Funnel (CSF) is a two-stage filtering architecture. The first layer, Volatility Filtering Layer (VFL), calculates the Dynamic Volatility Index (DVI) for all indicators that pass spatiotemporal validation. The DVI is composed of three elements: Short-Term Coefficient of Variation (STCV): The ratio of the standard deviation of the indicator value in the last 5 seconds to the mean; Historical Volatility Deviation (HVD): Compares the deviation of the current STCV with the baseline value under the same operating conditions in the past 24 hours; Sensor Health Index (SHI): Quantifies the reliability of voltage / temperature sensors based on their self-diagnostic signals (such as zero drift).

[0141] For example, if the STCV of a vibration acceleration indicator is 0.35 (normal should be <0.1), the HVD is 2.8 times (baseline is 0.12), and the SHI is 0.7 (out of 1.0), then the DVI = STCV × 0.5 + HVD × 0.3 + (1-SHI) × 0.2 = 0.52. If the DVI is > 0.4 (the threshold is adjustable), it is considered a high volatility indicator and is eliminated.

[0142] The second layer, the Stable Feature Preservation Layer (SFPL), performs Feature Stability Authentication (FSA) on the remaining indicators. This step introduces Sliding Window Autocorrelation Analysis (SWAA): The autocorrelation coefficient (AC) of the indicator series is calculated within a 10-second window with a step size of 1 second. If the AC is greater than 0.8 (strong autocorrelation), it indicates that the indicator changes regularly (for example, cooling fan speed increases linearly with increasing temperature), and the feature is considered stable. Simultaneously, the predictability score (PS) of the indicator is fitted using an ARIMA model (autoregressive integrated moving average model). The root mean square error (RMSE) of the prediction residuals is calculated. A lower RMSE indicates a higher PS. For example, the PS of DC output current can reach 0.95 (RMSE = 0.5A), while the PS of an indicator affected by ambient humidity may be only 0.6.

[0143] Finally, a High-Confidence Indicator Subset (HCIS) is generated. Each indicator in the HCIS is assigned an Integrated Confidence Tag (ICT) using the following formula: ICT = 0.4 × STCS + 0.3 × (1 - DVI) + 0.3 × PS.

[0144] Tag values ​​range from 0 to 1.0, and only metrics with an ICT value greater than 0.75 are retained. For example, the retained metrics after filtering might include: "DC bus voltage smoothing value (ICT = 0.88)", "IGBT junction temperature trend value (ICT = 0.82)", and "Fan speed autocorrelation coefficient (ICT = 0.79)". The HCIS data volume is reduced by over 60% compared to the original data set, while its reliability is significantly improved, providing clean input for status level determination.

[0145] Apply the fuzzy inference engine, input a subset of high-confidence indicators, and output a state level matrix through a preset rule base; The core of the Fuzzy Inference Engine (FIE) is the Predefined Rule Base (PRB), which contains three types of rules: Electrical safety rules: For example, "If the DC overvoltage probability is > 0.7 and the insulation resistance value is < 500kΩ, then the arc risk level = high"; Mechanical health rules: For example, "If the vibration energy in the 4kHz frequency band is > 0.15g 2 / Hz and bearing temperature > 85°C, then mechanical wear level = severe"; Thermal management rules: For example, "If the heatsink temperature difference is > 15°C and the fan speed deviates from the set value by > 20%, then the cooling failure level = medium."

[0146] Each rule is described using fuzzy logic. An input metric (e.g., "overvoltage probability") is mapped to a fuzzy set (e.g., "low / medium / high") using a membership function (MF). For example, the MF for overvoltage probability is defined as: low (0-0.3, triangular function), medium (0.2-0.7, trapezoidal function), and high (0.6-1.0, sigmoid function).

[0147] The engine performs Multi-Stage Inference (MSI): Step 1: Fuzzification: Convert the exact value in the HCIS (e.g., "bearing temperature = 88°C") into a fuzzy set. By querying the temperature membership function (e.g., "normal" < 80°C, "warning" 80-90°C, "dangerous" > 85°C), 88°C is assigned a "warning" membership of 0.6 and a "dangerous" membership of 0.4.

[0148] Step 2: Rule Activation: Traverse the PRB and match the fuzzy sets in the conditional part (IF part). For example, for the rule "If bearing temperature = Dangerous and vibration energy = High, then fault level = Urgent," since the membership of temperature "Dangerous" is 0.4 and the membership of vibration energy "High" is 0.8, the minimum value of 0.4 is used as the rule activation strength.

[0149] Step 3: Defuzzification: For each activated rule for an output variable (e.g., "fault level"), use the centroid method to calculate the exact value. For example, if the "Emergency" level is activated at 0.4 and the "High Risk" level is activated at 0.3, the output value is (0.4 × 90 + 0.3 × 70) / (0.4 + 0.3) = 81.4 (percentile).

[0150] The final output is the State Level Matrix (SLM). This matrix is ​​a two-dimensional structure with 6 rows and 3 columns: The rows represent six core dimensions: electrical safety, mechanical health, thermal management, communication status, efficiency index, and environmental adaptability; The columns represent three levels: normal (0-30 points), warning (31-70 points), and alarm (71-100 points).

[0151] Each cell is populated with the defuzzified score for that dimension (e.g., Thermal Management = 65 points, falling into the "Alert" column). The matrix is ​​supplemented with a Global Health Index (GHI), which is a weighted average of the scores for each dimension (electrical weighting 0.3, mechanical 0.25, thermal management 0.2, etc.). For example, a GHI of 58 indicates that the entire system is in an alert state.

[0152] Perform optimized alarm distribution, add timestamp hash chains to high-risk alarms, push them to the operation and maintenance terminal through priority queues, and trigger the device power reduction protection protocol at the same time.

[0153] Alert Optimization Distribution (AOD) first processes high-risk alerts (HRAs), defined as those with a score > 75 in any dimension of the SLM or a GHI > 70. A timestamped hash chain (THC) is added to the HRA: Generate the current alarm timestamp (e.g. 2025-04-11 14:23:05.876); Extract the hash value of the previous alert (the first alert uses the initial value 0x0000); Concatenate the current timestamp, alert content, and previous hash value, and input the SHA-256 algorithm to generate a new hash (e.g., 0x9f86d081...); The new hash and timestamp are embedded in the alarm message. THC ensures that the alarm sequence cannot be tampered with (for example, maintenance personnel cannot delete the "IGBT overtemperature" record) and that the event chain (for example, "fan failure → temperature surge → power reduction") can be traced.

[0154] Alerts are pushed through a Priority Queue (PQ). The queue defines four priorities: P0 (Emergency): Electrical safety / mechanical health alarm (such as arc risk, bearing fragmentation), delay requirement is < 1 second; P1 (high risk): thermal management alarm or dual-dimensional warning (such as cooling failure + efficiency drop), delay < 3 seconds; P2 (Warning): Single-dimensional warning or environmental abnormality (such as excessive humidity), delay < 10 seconds; P3 (prompt): performance degradation or maintenance reminder (such as the number of contactor operations is approaching the end of life), delay < 60 seconds.

[0155] The queues use the Weighted Fair Queuing (WFQ) algorithm, with P0 allocated 50% of the bandwidth, P1 30%, and the remaining 20%. Push targets include the operations and maintenance personnel's mobile app (via the MQTT protocol), the cloud monitoring platform (HTTPS API), and local industrial control screens (WebSocket).

[0156] The P0 / P1 level alarm synchronously triggers the device power reduction protection protocol (Derating Power Protection Protocol, DP 3 ): Gradual current reduction: If an over-temperature alarm occurs, the output current is reduced at a rate of 50A per second (e.g., from 200A to 150A) until the temperature drops back to within the threshold. Voltage clamping: If an overvoltage alarm occurs, the charging pile output voltage is forced to not exceed 105% of the vehicle's requested value; Safety lock: If an insulation fault (<100kΩ) is detected, the DC contactor will be immediately disconnected and locked for restart; Power reduction records: The power reduction magnitude (e.g., 30%) and duration are recorded in the blockchain for accident responsibility tracing. The SLM is continuously monitored during protocol execution. If the GHI drops below 60, power is gradually restored (in increments of 20A per second).

[0157] Spatiotemporal filtering eliminates transient interference, compares adjacent device states using Mahalanobis distance, and integrates a fuzzy inference engine to comprehensively assess risk levels. Confirmed high-risk faults are timestamped to prevent tampering and trigger a tiered response mechanism. Spatiotemporal verification effectively suppresses false alarms, while confidence screening ensures alarm reliability. This tiered response mechanism balances operational efficiency with device safety, preventing overprotection from impacting the user charging experience.

[0158] It can be seen that the real-time operating parameter data from the automobile DC charging pile is received and integrated into an original multidimensional data set; nonlinear feature mapping is performed on the original multidimensional data set to construct a state feature vector set, and each state feature vector is decomposed into an electrical sub-vector and a mechanical sub-vector; the electrical sub-vector and the mechanical sub-vector are fused and analyzed in real time using a distributed edge computing framework to generate a preliminary abnormality probability set; multi-objective optimization is performed on the preliminary abnormality probability set to generate an optimized state indicator set; spatiotemporal consistency verification and confidence screening are performed on the optimized state indicator set to determine the final operating state level, and a real-time monitoring alarm is output, thereby improving the accuracy and real-time performance of fault detection.

[0159] Another embodiment of the present invention provides a real-time monitoring system for the operation status of a DC charging station for a vehicle. Figure 3 , the system may include: The receiving module 301 is used to receive real-time operating parameter data from the DC charging station of the vehicle, collect current, voltage, temperature and vibration signals based on a preset high-frequency sampling protocol, and integrate them into an original multi-dimensional data set; A mapping module 302 is configured to perform nonlinear feature mapping on the original multidimensional dataset, construct a set of state feature vectors, and decompose each state feature vector into an electrical sub-vector and a mechanical sub-vector to separately capture electrical characteristics and mechanical vibration characteristics; An analysis module 303 is configured to perform real-time fusion analysis of the electrical sub-vector and the mechanical sub-vector using a distributed edge computing framework to generate a preliminary abnormality probability set; An optimization module 304 is configured to perform multi-objective optimization on the preliminary abnormal probability set, apply a reinforcement learning algorithm to optimize real-time responsiveness and fault prediction accuracy, and generate an optimized state indicator set; The screening module 305 is used to perform spatiotemporal consistency verification and confidence screening on the optimization status indicator set, determine the final operating status level, and output a real-time monitoring alarm.

[0160] An embodiment of the present invention further provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps of any one of the above method embodiments when running.

[0161] Specifically, in this embodiment, the above-mentioned storage medium may be configured to store a computer program for performing the following steps: S201, receiving real-time operating parameter data from a DC charging station for a vehicle, collecting current, voltage, temperature, and vibration signals based on a preset high-frequency sampling protocol, and integrating them into an original multidimensional data set; S202, performing nonlinear feature mapping on the original multidimensional data set to construct a state feature vector set, and decomposing each state feature vector into an electrical sub-vector and a mechanical sub-vector to separately capture electrical characteristics and mechanical vibration characteristics; S203, using a distributed edge computing framework to perform real-time fusion analysis on the electrical sub-vector and the mechanical sub-vector to generate a preliminary abnormality probability set; S204, performing multi-objective optimization on the preliminary abnormal probability set, applying a reinforcement learning algorithm to optimize real-time responsiveness and fault prediction accuracy, and generating an optimized state indicator set; S205 , performing spatiotemporal consistency verification and confidence screening on the optimization status indicator set, determining a final operating status level, and outputting a real-time monitoring alarm.

[0162] An embodiment of the present invention further provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any one of the above method embodiments.

[0163] Specifically, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0164] Specifically, in this embodiment, the processor may be configured to execute the following steps through a computer program: S201, receiving real-time operating parameter data from a DC charging station for a vehicle, collecting current, voltage, temperature, and vibration signals based on a preset high-frequency sampling protocol, and integrating them into an original multidimensional data set; S202, performing nonlinear feature mapping on the original multidimensional data set to construct a state feature vector set, and decomposing each state feature vector into an electrical sub-vector and a mechanical sub-vector to separately capture electrical characteristics and mechanical vibration characteristics; S203, using a distributed edge computing framework to perform real-time fusion analysis on the electrical sub-vector and the mechanical sub-vector to generate a preliminary abnormality probability set; S204, performing multi-objective optimization on the preliminary abnormal probability set, applying a reinforcement learning algorithm to optimize real-time responsiveness and fault prediction accuracy, and generating an optimized state indicator set; S205 , performing spatiotemporal consistency verification and confidence screening on the optimization status indicator set, determining a final operating status level, and outputting a real-time monitoring alarm.

[0165] The above describes in detail the structure, features and effects of the present invention based on the embodiments shown in the drawings. The above is only a preferred embodiment of the present invention, but the scope of implementation of the present invention is not limited to what is shown in the drawings. Any changes made in accordance with the concept of the present invention, or modifications to equivalent embodiments with equivalent changes, which do not exceed the spirit covered by the description and drawings, should be within the scope of protection of the present invention.

Claims

1. A method for real-time monitoring of the operating status of a DC charging pile for an automobile, characterized in that: The method comprises: Receive real-time operating parameter data from the vehicle's DC charging station, collect current, voltage, temperature, and vibration signals based on a preset high-frequency sampling protocol, and integrate them into a raw multidimensional data set; Performing nonlinear feature mapping on the original multidimensional data set to construct a state feature vector set, and decomposing each state feature vector into an electrical sub-vector and a mechanical sub-vector to separately capture electrical characteristics and mechanical vibration characteristics; Using a distributed edge computing framework, the electrical sub-vector and the mechanical sub-vector are fused and analyzed in real time to generate a preliminary abnormality probability set; Performing multi-objective optimization on the preliminary abnormal probability set, applying a reinforcement learning algorithm to optimize real-time responsiveness and fault prediction accuracy, and generating an optimized state indicator set; The optimization status indicator set is subjected to spatiotemporal consistency verification and confidence screening to determine the final operating status level and output a real-time monitoring alarm.

2. The method according to claim 1, characterized in that The system receives real-time operating parameter data from the DC charging pile of the vehicle, collects current, voltage, temperature and vibration signals based on a preset high-frequency sampling protocol, and integrates them into an original multi-dimensional data set, including: The high-frequency waveforms of current and voltage are collected synchronously through a multi-channel anti-aliasing filter, while the three-dimensional vibration signals are captured using a MEMS accelerometer to generate a heterogeneous sensing stream with a time stamp. Adaptive Kalman filtering is used to align heterogeneous sensor streams in the time domain, eliminating the timing deviation caused by the physical location differences of the sensors and outputting a synchronized sensor matrix. Perform wavelet packet decomposition on the synchronized sensing matrix, extract energy entropy features in a specific frequency band, and construct an initial multi-dimensional feature tensor; The initial multidimensional feature tensor is reconstructed and the redundant dimensions are compressed by Tucker decomposition to generate a compact original multidimensional dataset.

3. The method according to claim 2, characterized in that The performing nonlinear feature mapping on the original multidimensional data set, constructing a state feature vector set, and decomposing each state feature vector into an electrical sub-vector and a mechanical sub-vector to separately capture electrical characteristics and mechanical vibration characteristics, includes: The original multidimensional data set is input into the kernel function space, and nonlinear mapping is performed through the radial basis function kernel to generate a high-dimensional state feature space; Perform orthogonal feature decoupling in the high-dimensional state feature space and use the FastICA algorithm to separate electrical and mechanical features; Apply information entropy weighting to the decoupled electrical features to generate electrical sub-vectors with physical interpretability; Perform wavelet coherence analysis on mechanical characteristics, extract the phase coupling coefficient between the vibration signal and the cooling fan speed, and construct the mechanical sub-vector; The electrical sub-vector and the mechanical sub-vector are combined into a state feature vector set through the feature cascade layer, and a feature validity flag is added.

4. The method according to claim 3, characterized in that The distributed edge computing framework is used to perform real-time fusion analysis on the electrical sub-vector and the mechanical sub-vector to generate a preliminary abnormality probability set, including: Deploy a lightweight graph neural network on the edge computing node, use the electrical subvector as the node attribute and the charging pile topology connection relationship as the edge weight to build the electrical state graph structure; Perform spatiotemporal convolution on the mechanical sub-vectors to aggregate the vibration characteristics of adjacent charging piles in the spatial dimension, and apply causal convolution in the temporal dimension to generate a mechanical state evolution tensor; Design a dual-channel attention fusion mechanism, in which the electrical channel focuses on current mutation events and the mechanical channel focuses on abnormal vibration frequency bands, and outputs a cross-modal attention weight matrix; Based on the attention weight matrix, a gated recurrent unit is used to fuse the electrical state graph and the mechanical evolution tensor to generate a multi-dimensional state fusion vector; The multi-dimensional state fusion vector is input into the anomaly detection forest, and a preliminary anomaly probability set of each dimension is output.

5. The method according to claim 4, characterized in that The multi-objective optimization of the preliminary abnormal probability set is performed, and a reinforcement learning algorithm is applied to optimize the real-time responsiveness and fault prediction accuracy to generate an optimized state indicator set, including: Construct a dual-objective reward function where real-time responsiveness is rewarded with associated processing latency, and prediction accuracy is rewarded with associated fault detection F1 score; Design a deep Q-network architecture that uses anomaly probability sets as state inputs and feature sampling frequencies as action space, and estimates the value function through a two-stream network. Adopting a non-dominated sorting strategy, the action strategy is selected on the Pareto optimal frontier, and the current sampling frequency and vibration analysis granularity are dynamically adjusted; The network parameters are updated through the proximal strategy optimization algorithm, and the optimized state indicator set is output.

6. The method according to claim 5, characterized in that The performing of spatiotemporal consistency verification and confidence screening on the optimized state indicator set, determining the final operating state level, and outputting a real-time monitoring alarm includes: Construct a spatiotemporal consistency check window, use exponentially weighted moving average to filter out instantaneous interference in the time dimension, and compare the Mahalanobis distance of similar indicators of adjacent charging piles in the spatial dimension; Design a confidence screening funnel, where the first layer removes highly volatile indicators, and the second layer retains stable features to generate a subset of high-confidence indicators; Apply the fuzzy inference engine, input a subset of high-confidence indicators, and output a state level matrix through a preset rule base; Perform optimized alarm distribution, add timestamp hash chains to high-risk alarms, push them to the operation and maintenance terminal through priority queues, and trigger the device power reduction protection protocol at the same time.

7. A real-time monitoring system for the operating status of a DC charging pile for a car, characterized in that: The system comprises: The receiving module is used to receive real-time operating parameter data from the vehicle DC charging station, collect current, voltage, temperature and vibration signals based on a preset high-frequency sampling protocol, and integrate them into a raw multidimensional data set; a mapping module, configured to perform nonlinear feature mapping on the original multidimensional data set, construct a set of state feature vectors, and decompose each state feature vector into an electrical sub-vector and a mechanical sub-vector to separately capture electrical characteristics and mechanical vibration characteristics; an analysis module, configured to perform real-time fusion analysis of the electrical sub-vector and the mechanical sub-vector using a distributed edge computing framework to generate a preliminary abnormality probability set; an optimization module for performing multi-objective optimization on the preliminary abnormal probability set, applying a reinforcement learning algorithm to optimize real-time responsiveness and fault prediction accuracy, and generating an optimized state indicator set; The screening module is used to perform spatiotemporal consistency verification and confidence screening on the optimization status indicator set, determine the final operating status level, and output a real-time monitoring alarm.

8. The system according to claim 7, characterized in that The receiving module is specifically configured to: The high-frequency waveforms of current and voltage are collected synchronously through a multi-channel anti-aliasing filter, while the three-dimensional vibration signals are captured using a MEMS accelerometer to generate a heterogeneous sensing stream with a time stamp. Adaptive Kalman filtering is used to align heterogeneous sensor streams in the time domain, eliminating the timing deviation caused by the physical location differences of the sensors and outputting a synchronized sensor matrix. Perform wavelet packet decomposition on the synchronized sensing matrix, extract energy entropy features in a specific frequency band, and construct an initial multi-dimensional feature tensor; The initial multidimensional feature tensor is reconstructed and the redundant dimensions are compressed by Tucker decomposition to generate a compact original multidimensional dataset.

9. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 6 when executed.

10. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 6.

Images