A real-time monitoring method and system for operation state of a direct-current charging pile of an automobile

Abstract

This invention discloses a method and system for real-time monitoring of the operating status of a DC charging pile for automobiles. The method includes: receiving real-time operating parameter data from the DC charging pile for automobiles and integrating it into a raw multidimensional dataset; performing nonlinear feature mapping on the raw multidimensional dataset to construct a set of state feature vectors, and decomposing each state feature vector into electrical sub-vectors and mechanical sub-vectors; using a distributed edge computing framework to perform real-time fusion analysis on the electrical sub-vectors and mechanical sub-vectors to generate a preliminary anomaly probability set; performing multi-objective optimization on the preliminary anomaly probability set to generate an optimized state index set; performing spatiotemporal consistency verification and confidence screening on the optimized state index set to determine the final operating status level and outputting a real-time monitoring alarm. Using this invention, the accuracy and real-time performance of fault detection can be improved.

Description

Technical Field

[0001] This invention belongs to the field of charging pile technology, and in particular to a method and system for real-time monitoring of the operating status of a DC charging pile for automobiles. Background Technology

[0002] With the increasing popularity of electric vehicles, the operational safety and stability of DC charging stations have become critical issues. Traditional monitoring methods often rely on single-parameter threshold alarms or offline diagnostics, which are insufficient to meet the real-time fault warning requirements under complex operating conditions. In existing technologies, electrical parameters (such as current and voltage) and mechanical parameters (such as vibration and temperature) are typically analyzed independently, leading to the neglect of fault correlations and high data processing latency, affecting response efficiency. Furthermore, conventional algorithms are not adaptable enough to nonlinear characteristics and dynamic operating conditions, easily resulting in false alarms or missed alarms. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for real-time monitoring of the operating status of DC charging piles for automobiles, so as to overcome the shortcomings of the prior art and improve the accuracy and real-time performance of fault detection.

[0004] One embodiment of this application provides a method for real-time monitoring of the operating status of a DC charging pile for automobiles, the method comprising:

[0005] It receives real-time operating parameter data from DC charging piles for automobiles, collects current, voltage, temperature and vibration signals based on a preset high-frequency sampling protocol, and integrates them into a raw multidimensional dataset.

[0006] The original multidimensional dataset is subjected to nonlinear feature mapping to construct a set of state feature vectors. Each state feature vector is decomposed into electrical sub-vectors and mechanical sub-vectors to separate and capture electrical characteristics and mechanical vibration characteristics.

[0007] A distributed edge computing framework is used to perform real-time fusion analysis on the electrical sub-vectors and the mechanical sub-vectors to generate a preliminary set of anomaly probabilities.

[0008] The preliminary anomaly probability set is optimized using a multi-objective approach, and a reinforcement learning algorithm is applied to optimize real-time responsiveness and fault prediction accuracy, thereby generating an optimized set of state indicators.

[0009] The set of optimized status indicators is subjected to spatiotemporal consistency verification and confidence screening to determine the final operating status level and output real-time monitoring alarms.

[0010] Optionally, the step of receiving real-time operating parameter data from the vehicle DC charging pile involves collecting current, voltage, temperature, and vibration signals based on a preset high-frequency sampling protocol and integrating them into a raw multidimensional dataset, including:

[0011] High-frequency waveforms of current and voltage are simultaneously acquired using a multi-channel anti-aliasing filter, while three-dimensional vibration signals are captured using a MEMS accelerometer to generate a heterogeneous sensing flow with timestamps.

[0012] An adaptive Kalman filter is used to align the heterogeneous sensor streams in the time domain, eliminating timing deviations caused by differences in the physical locations of the sensors and outputting a synchronized sensing matrix.

[0013] Wavelet packet decomposition is performed on the synchronized sensing matrix to extract energy entropy features in a specific frequency band and construct an initial multidimensional feature tensor.

[0014] The initial multidimensional feature tensor is reconstructed, and redundant dimensions are compressed through Tucker decomposition to generate a compact original multidimensional dataset.

[0015] Optionally, the step of performing nonlinear feature mapping on the original multidimensional dataset to construct a set of state feature vectors, and decomposing each state feature vector into electrical sub-vectors and mechanical sub-vectors to separate and capture electrical characteristics and mechanical vibration characteristics, includes:

[0016] The original multidimensional dataset 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.

[0017] Orthogonal feature decoupling is performed in a high-dimensional state feature space, and the FastICA algorithm is used to separate electrical and mechanical features.

[0018] Information entropy weighting is applied to the decoupled electrical features to generate physically interpretable electrical sub-vectors;

[0019] Wavelet coherence analysis is performed on the mechanical characteristics to extract the phase coupling coefficient between the vibration signal and the cooling fan speed, and a mechanical sub-vector is constructed.

[0020] The electrical and mechanical sub-vectors are combined into a set of state feature vectors through a feature cascade layer, and a feature validity flag is added.

[0021] Optionally, the step of 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 anomaly probability set includes:

[0022] A lightweight graph neural network is deployed on the edge computing node, with electrical sub-vectors as node attributes and charging pile topology connections as edge weights to construct an electrical state graph structure.

[0023] Spatiotemporal convolution is performed on the mechanical subvectors to aggregate the vibration features of adjacent charging piles in the spatial dimension, and causal convolution is applied in the temporal dimension to generate a mechanical state evolution tensor.

[0024] Design a dual-channel attention fusion mechanism, in which the electrical channel focuses on current surge events and the mechanical channel focuses on abnormal vibration frequency bands, and outputs a cross-modal attention weight matrix;

[0025] Based on the attention weight matrix, a gated cyclic unit is used to fuse the electrical state diagram and the mechanical evolution tensor to generate a multi-dimensional state fusion vector;

[0026] Input the multidimensional state fusion vector into the anomaly detection forest, and output a preliminary set of anomaly probabilities for each dimension.

[0027] Optionally, the step of performing multi-objective optimization on the preliminary anomaly probability set, applying reinforcement learning algorithms to optimize real-time responsiveness and fault prediction accuracy, and generating an optimized state index set includes:

[0028] Construct a dual-objective reward function, in which real-time responsiveness is rewarded for processing latency, and prediction accuracy is rewarded for fault detection F1 score.

[0029] Design a deep Q-network architecture, using the set of anomaly probabilities as the state input and the feature sampling frequency as the action space, and estimate the value function through a two-stream network;

[0030] A non-dominated sorting strategy is adopted to select the action strategy at the Pareto optimal front, and the current sampling frequency and vibration analysis granularity are dynamically adjusted.

[0031] The network parameters are updated using a near-end policy optimization algorithm, and the optimized set of state indicators is output.

[0032] Optionally, the step of performing spatiotemporal consistency verification and confidence screening on the optimized state indicator set to determine the final operating state level and outputting real-time monitoring alarms includes:

[0033] A spatiotemporal consistency verification window is constructed. In the time dimension, an exponentially weighted moving average is used to filter out instantaneous interference, and in the spatial dimension, the Mahalanobis distance of the same indicators of adjacent charging piles is compared.

[0034] Design a confidence screening funnel, in which the first layer removes high volatility indicators, the second layer retains stable features, and generates a subset of high confidence indicators;

[0035] The system uses a fuzzy inference engine, takes a subset of high-confidence indicators as input, and outputs a state level matrix through a preset rule base.

[0036] The system optimizes alarm distribution by adding a timestamp hash chain to high-risk alarms and pushing them to the operation and maintenance terminal via a priority queue, while simultaneously triggering the device power reduction protection protocol.

[0037] Another embodiment of this application provides a real-time monitoring system for the operating status of a DC charging pile for automobiles, the system comprising:

[0038] The receiving module is used to receive real-time operating parameter data from the DC charging pile for automobiles, collect current, voltage, temperature and vibration signals based on a preset high-frequency sampling protocol, and integrate them into a raw multidimensional dataset.

[0039] The mapping module is used to perform nonlinear feature mapping on the original multidimensional dataset, construct a set of state feature vectors, and decompose each state feature vector into electrical sub-vectors and mechanical sub-vectors to separate and capture electrical characteristics and mechanical vibration characteristics.

[0040] The analysis module is used to perform real-time fusion analysis on the electrical sub-vector and the mechanical sub-vector using a distributed edge computing framework to generate a preliminary set of anomaly probabilities.

[0041] The optimization module is used to perform multi-objective optimization on the preliminary anomaly probability set, apply reinforcement learning algorithms to optimize real-time responsiveness and fault prediction accuracy, and generate an optimized state index set.

[0042] The filtering module is used to perform spatiotemporal consistency verification and confidence filtering on the set of optimized status indicators, determine the final operating status level, and output real-time monitoring alarms.

[0043] Another embodiment of this application provides a storage medium storing a computer program, wherein the computer program is configured to execute the method described in any of the preceding claims when running.

[0044] Another embodiment of this application provides an electronic device including 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 method described in any of the preceding claims.

[0045] Compared with existing technologies, the present invention provides a real-time monitoring method for the operating status of a DC charging pile for automobiles. This method receives real-time operating parameter data from the DC charging pile and integrates it into a raw multidimensional dataset. It then performs nonlinear feature mapping on the raw multidimensional dataset to construct a set of state feature vectors, decomposing each state feature vector into electrical and mechanical sub-vectors. A distributed edge computing framework is used to perform real-time fusion analysis on the electrical and mechanical sub-vectors to generate a preliminary anomaly probability set. This preliminary anomaly probability set is then optimized using multi-objective methods to generate an optimized state index set. Finally, the optimized state index set undergoes spatiotemporal consistency verification and confidence screening to determine the final operating status level and output a real-time monitoring alarm, thereby improving the accuracy and real-time performance of fault detection. Attached Figure Description

[0046] Figure 1A hardware structure block diagram of a computer terminal for a real-time monitoring method for the operating status of a DC charging pile for automobiles, provided in an embodiment of the present invention;

[0047] Figure 2 A flowchart illustrating a method for real-time monitoring of the operating status of a DC charging pile for automobiles, provided in an embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of a real-time monitoring system for the operating status of a DC charging pile for automobiles, provided in an embodiment of the present invention. Detailed Implementation

[0049] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0050] This invention first provides a method for real-time monitoring of the operating status of a DC charging pile for automobiles. This method can be applied to electronic devices, such as computer terminals, specifically ordinary computers.

[0051] The following detailed explanation uses a computer terminal as an example. Figure 1 This is a hardware structure block diagram of a computer terminal for a real-time monitoring method of the operating status of a DC charging pile for automobiles, provided in an embodiment of the present invention. Figure 1 As shown, the computer device includes a processor, memory, and network interface connected via a system bus, wherein the memory may include non-volatile storage media and internal memory.

[0052] The non-volatile storage medium can store an operating system and a computer program. This computer program includes program instructions that, when executed, cause the processor to perform any method for real-time monitoring of the operating status of a DC charging station for automobiles.

[0053] The processor provides computing and control capabilities, supporting the operation of the entire computer device.

[0054] The internal memory provides an environment for the execution of computer programs in non-volatile storage media. When the computer program is executed by the processor, it enables the processor to execute any method for real-time monitoring of the operating status of a DC charging pile for automobiles.

[0055] This network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 1 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0056] It should be understood that the processor can be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. Among these, a general-purpose processor can be a microprocessor or any conventional processor.

[0057] See Figure 2 The present invention provides a method for real-time monitoring of the operating status of a DC charging pile for automobiles, which may include the following steps:

[0058] S201 receives real-time operating parameter data from the DC charging pile for automobiles, collects current, voltage, temperature and vibration signals based on a preset high-frequency sampling protocol, and integrates them into a raw multidimensional dataset.

[0059] 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 flow with timestamps.

[0060] Multi-channel anti-aliasing filtering and high-frequency waveform acquisition

[0061] A multi-channel anti-aliasing filter (MAAF) is deployed at the power input of the DC charging pile. Its core function is to prevent spectral aliasing distortion during high-frequency signal sampling. Current signals are acquired through a Rogowski coil or Hall effect sensor, while voltage signals are obtained through a high-precision voltage divider resistor network. The filter employs a Butterworth low-pass topology, with the cutoff frequency set to 0.4 times the target sampling frequency (according to the Nyquist sampling theorem). For example, if the current sampling frequency is 100 kHz, the cutoff frequency is set to 40 kHz to ensure that noise above this frequency (such as high-frequency harmonics from switching transistors) is suppressed. Each channel is equipped with an independent 24-bit analog-to-digital converter (ADC) to synchronously capture 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, with time synchronization error controlled within ±50 nanoseconds (ns).

[0062] Three-dimensional vibration capture using MEMS accelerometers

[0063] Mechanical vibration signals are acquired using MEMS (Micro-Electro-Mechanical Systems Accelerometer) sensors mounted on key components of the charging pile (such as the power module heatsink and transformer housing). This sensor, based on a capacitive micromechanical structure, can simultaneously measure acceleration values ​​(in g, gravitational acceleration) along the X, Y, and Z axes. Its sampling frequency is set to 5 kHz, covering the typical frequency band of mechanical vibration in charging piles (0 Hz to 2.5 kHz). The sensor incorporates a self-calibration circuit to eliminate zero-point errors caused by temperature drift (range: -40°C to 85°C). The vibration signal is output as a digital signal and transmitted to the main control unit via a high-speed serial peripheral interface (SPI). Each vibration data point is appended with a precise timestamp, provided by the same crystal oscillator source as the electrical sampling, ensuring a consistent time base across modal data.

[0064] Generation and integration of heterogeneous sensing flows

[0065] Current and voltage waveform data, along with three-dimensional vibration signals, are aggregated into a Data Preprocessing Unit (DPU). This unit allocates an independent data stream channel to each data type and inserts a uniform timestamp (accuracy ±100 nanoseconds). Current and voltage waveforms are stored in array format (e.g., 100,000 sampling points per second), while vibration data is packaged into a structure based on triaxial acceleration components. This ultimately generates a Heterogeneous Sensing Stream (HSS), whose data structure comprises four parts:

[0066] Current Stream: Includes a timestamp sequence and corresponding current amplitude (unit: amperes (A));

[0067] Voltage Stream: A sequence of timestamps and voltage amplitudes (unit: volts (V)).

[0068] Vibration Stream: Timestamp sequence and X / Y / Z axial acceleration values ​​(unit: meters per second squared (m / s)) 2 )).

[0069] The data stream is transmitted in real time to the edge computing node via Gigabit Ethernet.

[0070] An adaptive Kalman filter is used to align the heterogeneous sensor streams in the time domain, eliminating timing deviations caused by differences in the physical locations of the sensors and outputting a synchronized sensing matrix.

[0071] Timing Deviation Modeling and Filter Initialization

[0072] Due to the different physical installation locations of the current sensor, voltage probe, and MEMS accelerometer within the charging pile, the varying signal transmission path lengths (typically 0.1 meters to 1 meter) result in a fixed transmission delay (FTD) for 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 bias, the Adaptive Kalman Filter (AKF) first establishes a state-space model:

[0073] State Variable: The actual time offset of each signal stream (unit: nanoseconds (ns));

[0074] 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).

[0075] During filter initialization, the process noise covariance (Q) is set to 1e-6 and the observation noise covariance (R) is set to 1e-5, representing the estimation of system dynamics and measurement errors.

[0076] Dynamic adjustment process of time domain alignment

[0077] During the runtime phase, AKF executes a prediction-update cycle every 10 milliseconds (ms):

[0078] Prediction Phase: Based on the estimated time offset from the previous moment and the system dynamics equations (assuming the offset changes slowly), predict the current offset.

[0079] Update Phase: When a reference edge is detected (such as the rising edge of the power module switch in a charging pile), the actual arrival time of the event in each signal stream is compared with the theoretical time (determined by the global clock of the crystal oscillator), generating an observation residual. The predicted value is corrected by using the Kalman gain (K) weighted residual.

[0080] For example, if a current signal is detected to lag behind the reference edge by 400 nanoseconds (ns), while the predicted value is 350 nanoseconds (ns), the offset is updated to 390 nanoseconds (ns) using a gain K=0.8. This process tracks the delay drift caused by factors such as temperature changes and cable aging in real time.

[0081] Output of synchronized sensing matrix

[0082] The aligned data is reconstructed into a synchronized sensing matrix (SSM). The rows of the matrix correspond to a uniform time axis (time resolution 1 microsecond (µs)), and the columns contain four types of synchronization signals:

[0083] Column 1: Aligned current value (A);

[0084] Column 2: Aligned voltage values ​​(V);

[0085] Columns 3-5: X / Y / Z axial vibration acceleration values ​​(m / s²) 2 ).

[0086] Missing values ​​in the matrix (due to sensor response delays) are filled using cubic spline interpolation. The final output SSM is sent to subsequent processing modules to ensure that all sensor data are accurately correlated at the same time reference.

[0087] Wavelet packet decomposition is performed on the synchronized sensing matrix to extract energy entropy features in a specific frequency band and construct an initial multidimensional feature tensor.

[0088] Frequency band division of wavelet packet decomposition

[0089] Wavelet Packet Decomposition (WPD) uses the Daubechies4 Wavelet base to decompose each signal column of the SSM into 4 levels (Decomposition Level=4), generating 16 sub-bands. The band division is based on a binary recursive approach: Level 1 (0 Hz to sampling frequency / 2), Level 2 subdivided to 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 Level 4 is 100 kHz / 16 = 6.25 kHz. Key frequency bands are preset based on the characteristics of charging pile faults.

[0090] Electrical characteristic frequency bands: 0 Hz to 1 kHz (DC component and low-frequency ripple), 1 kHz to 10 kHz (switching frequency harmonics);

[0091] Mechanical characteristic frequency bands: 100 Hz to 500 Hz (fan fundamental frequency), 1 kHz to 2 kHz (bearing resonance).

[0092] Energy entropy feature extraction

[0093] The energy entropy (EE) is calculated within each sub-band, and its physical meaning reflects the randomness of the signal energy distribution. Specific steps:

[0094] Step 1: Calculate the sum of squares of the sub-band coefficients to obtain the band energy (BE).

[0095] Step 2: Divide BE by the total signal energy to normalize it into a probability distribution (PD).

[0096] Step 3: Calculate the entropy value using the Shannon Entropy Formula: EE = -Σ(PD ×log2(PD)).

[0097] For example, when a cooling fan bearing wears down, the energy entropy in the 1 kHz to 1.5 kHz frequency band increases from 0.8 bits to 1.2 bits, indicating a more dispersed energy distribution (an anomaly).

[0098] Initial Multidimensional Feature Tensor Construction

[0099] The energy entropy values ​​of all sub-bands of the five signals (current, voltage, triaxial vibration, etc.) are organized into an Initial Multidimensional Feature Tensor (IMFT) according to a time window (WindowLength = 1 second). The tensor dimension is:

[0100] Dimension 1: Number of samples within the time window (e.g., 1000 samples, corresponding to 1 second (s));

[0101] Dimension 2: Number of signal types (current, voltage, X / Y / Z vibration, 5 categories in total);

[0102] Dimension 3: Number of sub-bands (4-layer decomposition, totaling 16 bands).

[0103] 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.

[0104] The initial multidimensional feature tensor is reconstructed, and redundant dimensions are compressed through Tucker decomposition to generate a compact original multidimensional dataset.

[0105] The Necessity and Strategies of Tensor Reconstruction

[0106] IMFT exhibits significant dimensional redundancy:

[0107] High correlation between adjacent time samples (e.g., small changes in electrical parameters within 1 millisecond (ms));

[0108] The energy entropy of some frequency bands is close to zero (e.g., current signals have no energy above 10 kHz).

[0109] Tensor reshaping reassembles the original three-dimensional structure (time × signal type × frequency band) of the IMFT into an optimized new dimension:

[0110] New Mode 1: Merges time and signal type to form a joint "time-signal" dimension;

[0111] New Mode 2: Retains the frequency band dimension, but selects only 8 key frequency bands whose total energy exceeds a threshold (e.g., 1% of total energy);

[0112] New Mode 3: Adds a "Feature Type" dimension, reserving room for future expansion.

[0113] After reconstruction, the tensor size is compressed from 1000×5×16 to 500×8×1 (compression rate of 83%), reducing the computational load.

[0114] Section 2: Core Operations of Tucker Decomposition

[0115] Tucker decomposition decomposes the reconstructed tensor into the product of the core tensor and the factor matrices:

[0116] 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;

[0117] Step 2: Solve iteratively using the Higher-Order Orthogonal Iteration (HOOI) algorithm:

[0118] Fix two dimensions and perform Singular Value Decomposition (SVD) on the remaining dimensions.

[0119] Truncate the singular vectors (keeping the first 50, 3, and 1 principal components) and update the factor matrix;

[0120] Step 3: After convergence, the core tensor size is 50×3×1, and the factor matrix sizes are 500×50 (time-signal), 8×3 (bandwidth), and 1×1 (feature type).

[0121] Generation of compact datasets

[0122] The core tensor and factor matrix together constitute the Compact Raw Multidimensional Dataset (CRMDD). This dataset is stored in binary format and contains:

[0123] Core tensor (50×3×1): Stores coupling relationships across time, frequency band, and feature type;

[0124] Time-signal factor matrix (500×50): Each column represents a time-signal joint mode (such as "current slowly changing mode");

[0125] Frequency band factor matrix (8×3): Each column represents the frequency band clustering result (e.g., "low frequency electrical frequency band group");

[0126] Feature type factor matrix (1×1): scalar value is 1, preserving dimensional consistency.

[0127] For example, a core tensor element (25,2,1) = 0.8 indicates that the correlation strength between the 25th time-signal mode and the 2nd frequency band is 0.8. The dataset size is only 12% of the original IMFT, laying the foundation for subsequent real-time analysis.

[0128] First, key physical quantities of the charging pile during operation are simultaneously collected by multiple sensors, including electrical parameters (current, voltage) and mechanical parameters (temperature, vibration). High-frequency sampling ensures the capture of transient abnormal signals. Heterogeneous sensor data are aligned in the time domain through a preset protocol to form a structured multidimensional dataset, providing a unified data foundation for subsequent analysis and enabling comprehensive perception of the charging pile's operating status. High-frequency sampling can capture transient fault characteristics that are easily missed by traditional methods. Multidimensional data integration solves the problem of sensor data heterogeneity, creating conditions for refined status analysis.

[0129] S202, perform nonlinear feature mapping on the original multidimensional dataset to construct a set of state feature vectors, and decompose each state feature vector into electrical sub-vectors and mechanical sub-vectors to separate and capture electrical characteristics and mechanical vibration characteristics.

[0130] Specifically, the original multidimensional dataset can be input into the kernel function space, and a high-dimensional state feature space can be generated by nonlinear mapping through the radial basis function kernel.

[0131] Kernel function space construction principle

[0132] The original multidimensional dataset contains quantized values ​​of current, voltage, temperature, and vibration signals, with dimensions reaching tens of dimensions (e.g., 12-dimensional current harmonic components, 8-dimensional voltage ripple characteristics, 5-dimensional temperature gradient, and 18-dimensional vibration spectrum energy). These parameters exhibit complex nonlinear coupling relationships (e.g., sudden current changes trigger increased vibration). Directly employing linear analysis would result in the loss of crucial correlation information. The essence of 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 station. This unit receives all parameter values ​​from a time slice (e.g., a 500-millisecond window) in the original dataset, treating it as a multidimensional coordinate point.

[0133] Operating mechanism of radial basis function kernel

[0134] The Radial Basis Function Kernel (RBF Kernel) is a core mapping tool whose mathematical essence is calculating the "similarity distance" between data points. This kernel function contains two key parameters: the kernel width coefficient (Gamma, γ) and the truncation threshold (Epsilon, ε). γ controls the extent of the mapped feature space (typically 0.01-0.1), while ε determines whether small correlations are ignored (typically 1e-5). For example, when dealing with the correlation between current harmonic distortion rate (THD) and chassis vibration acceleration, the RBF kernel calculates the "gravitational strength" between the two in a 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, then these two events will be brought closer together in the mapped feature space to the same cluster region. The mapping process is achieved through matrix operations: a similarity matrix (SM) is constructed on the original data point set. The matrix element SM_ij represents the kernel function output value of the i-th and j-th time slice data. Finally, a high-dimensional state feature space (HDSFS) with a dimension much higher than the original data (e.g., from 43 dimensions to 1000 dimensions) is generated.

[0135] Physical meaning guarantee of feature space

[0136] To prevent high-dimensional spaces from becoming mere mathematical games, the system ensures interpretability through the injection of physical constraints. For example:

[0137] The current-voltage eigenspace is forced to satisfy the deformable constraints of Kirchhoff's laws;

[0138] The vibration characteristic subspace must conform to the mechanical resonance frequency formula (such as the approximate expression f_res=1 / (2π√(k / m)));

[0139] The mapping between temperature gradient and electrical loss must satisfy the first law of thermodynamics.

[0140] In the final generated HDSFS, each dimension corresponds to a nonlinear interaction mode (such as the "high harmonic excitation structure resonance" mode), 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 usually maintained between 20 and 30 times the original dimensions.

[0141] Orthogonal feature decoupling is performed in a high-dimensional state feature space, and the FastICA algorithm is used to separate electrical and mechanical features.

[0142] The necessity and challenges of orthogonal decoupling

[0143] During the operation of charging piles, electrical signals (such as DC bus pulsation) and mechanical vibrations (such as fan bearing wear) often interfere with each other. For example, a 2000Hz current oscillation caused by IGBT switching may excite resonance in the chassis panel, which in turn affects the heat sink temperature field. In the HDSFS generated in step 1, these coupled features are still mixed in the same hyperplane. The goal of Orthogonal Feature Decoupling is to project the two types of features into mutually perpendicular subspaces, ensuring that the dot product of the electrical feature vector and the mechanical feature vector is zero (orthogonality condition). To achieve this goal, the system uses the FastICA (Fast Independent Component Analysis) algorithm, which excels at separating statistically independent source components from mixed signals.

[0144] Parameterized execution of the FastICA algorithm

[0145] The core of FastICA is maximizing non-Gaussianity, using negative entropy (J) as the optimization objective. The algorithm comprises four key parameters:

[0146] Iteration tolerance error (Tolerance, Tol=1e-6): controls convergence accuracy;

[0147] Maximum number of iterations (MaxIter=200): Prevents infinite loops;

[0148] Nonlinear function type (G(u)=tanh(u)): used to measure non-Gaussianity;

[0149] Whitening preprocessing flag (Whitening=True): Eliminates correlation between features.

[0150] Taking the separation of cooling fan vibration and PFC circuit noise as an example:

[0151] Input: 1000-dimensional hybrid features from HDSFS;

[0152] Process: The algorithm continuously adjusts the unmixing matrix W through fixed-point iteration;

[0153] 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).

[0154] Physical verification of decoupling effect

[0155] To ensure the reliability of the separation results, the system is equipped with a triple verification mechanism:

[0156] Mutual Information Test (MI < 0.05): Verifies electrical / mechanical vector independence;

[0157] Energy backtracking comparison: The total electrical vector energy must have an error of less than 3% compared to the meter reading;

[0158] Fault injection test: When a fan jamming fault is artificially created, the bearing characteristic energy in the mechanical vector should increase by 20dB while the electrical vector remains unchanged.

[0159] The verified decoupling features will be labeled as Electrical Independent Component Set (EICS) and Mechanical Independent Component Set (MICS) for use in subsequent subvector construction.

[0160] Information entropy weighting is applied to the decoupled electrical features to generate physically interpretable electrical sub-vectors;

[0161] The engineering significance of information entropy weighting

[0162] The decoupled electrical characteristics (EICS) may contain dozens of independent components (e.g., switching frequency components, ground return noise, battery reverse current, etc.), but not all components are equally important for fault prediction. Information Entropy Weighting aims to quantify the "diagnostic value" of each characteristic. The entropy (H) value is calculated based on Shannon theory and reflects the degree of disorder of the characteristic. For example:

[0163] Low current harmonic entropy during stable operation (H≈0.2): strong regularity, low diagnostic value;

[0164] When contactor contacts are burned out, the current surge entropy value is high (H≈0.9): This indicates a high level of abnormal information, and its weight should be increased.

[0165] Calculation process of weighting factors

[0166] The system executes a three-step weighted strategy:

[0167] Feature entropy calculation: Calculate the sample entropy (SampEn) for each electrical component over 500 time windows. Parameter settings: embedding dimension m=2, similarity tolerance r=0.2 times the standard deviation;

[0168] Weight allocation: Using the inverse normalization method, the weight of the i-th feature is W_i = (1 - H_i) / Σ(1 - H_j);

[0169] Physical meaning injection: Superimposed expert rule-based correction factor (RCF). For example, the RCF of the peak-to-peak DC output voltage ripple (Vpp) is constant at 1.2 (due to sensitivity to capacitor failure), while the RCF of the common-mode current is only 0.8.

[0170] Structured output of electrical subvectors

[0171] The weighted feature set is organized into Electrical Sub-Vectors (ESVs), whose data structure consists of three layers:

[0172] 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));

[0173] Auxiliary indicator layer: 15-20 features with medium to low weight (e.g., PFC efficiency offset (weight 0.06));

[0174] Metadata header: contains sampling timestamp (64-bit integer), charging pile ID (CP_ID, 16-byte string), and weighted version number (Weight_Ver=2.3).

[0175] Ultimately, the ESV length is controlled to 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 cycle is 200 milliseconds.

[0176] Wavelet coherence analysis is performed on the mechanical characteristics to extract the phase coupling coefficient between the vibration signal and the cooling fan speed, and a mechanical sub-vector is constructed.

[0177] Technical Motivation for Wavelet Coherence Analysis

[0178] The main source of mechanical vibration in charging piles is the cooling fan (contribution rate >60%), but vibrations from the gearbox and contactor may also cause interference. Wavelet coherence analysis (WCA) is used to quantify the correlation between vibration signals and fan speed signals in the time-frequency domain. Compared with traditional FFT, wavelet transform uses Morlet wavelets as basis functions, with the following parameter settings: center frequency ω0 = 6 (optimal time-frequency resolution balance), scales = 64 (covering the 10-2000Hz frequency band).

[0179] The process of extracting phase coupling coefficients

[0180] The analysis process specifically includes:

[0181] Signal preprocessing: Synthesizing vector magnitude (VM = √(x / Y / Z axis acceleration) from the three-dimensional vibration signals (X / Y / Z axis acceleration) in the MICS 2 +y 2 +z 2 ));

[0182] Fan speed synchronization: The pulse frequency of the fan PWM control signal is obtained through the CAN bus (e.g., 2500 RPM corresponds to 41.67Hz).

[0183] Coherence calculation: Based on the continuous wavelet transform (CWT), calculate the time-frequency domain coherence value: Coh(f,t) = |S(fan_cwt * vib_cwt*)| 2 / (S(|fan_cwt| 2 ) * S(|vib_cwt| 2 )), where S is the smoothing operator and * denotes complex conjugation.

[0184] Key coefficient extraction: At the fan fundamental frequency (f0) and blade passing frequency (BPF = f0 × number of blades), the phase coupling coefficient (PCC) is calculated as follows: PCC_f0 = max( |Coh(f0,t)|×cos(Δφ(t)) ) t∈[t-0.5s,t]. Where Δφ(t) is the real-time phase difference between the vibration signal and the rotational speed signal.

[0185] Engineering encapsulation of mechanical subvectors

[0186] The Mechanical Sub-Vector (MSV) constructed based on PCC includes three types of metrics:

[0187] Main coupling parameters: fundamental frequency coupling coefficient PCC_f0 (normal range 0.7-0.9), blade passing frequency coupling coefficient PCC_BPF;

[0188] Fault characteristic index: coherent energy ratio at the bearing fault characteristic frequency (BPFI / BPFO);

[0189] Auxiliary verification values: kurtosis (K>3.5 warning) and envelope spectral entropy of the three-dimensional vibration signal.

[0190] Typical MSV structure example (unit: dimensionless): [PCC_f0=0.83, PCC_BPF=0.76, BPFI_Energy=0.12, Kurtosis_X=3.2, EnvEntropy=0.45].

[0191] This vector is updated every 500 milliseconds, and a "fan out of sync" warning is triggered when PCC_f0 < 0.6.

[0192] The electrical and mechanical sub-vectors are combined into a set of state feature vectors through a feature cascade layer, and a feature validity flag is added.

[0193] Cascaded layer structural design

[0194] The Feature Concatenation Layer (FCL) does not simply concatenate data, but performs intelligent alignment and compression:

[0195] Time axis alignment: Since the ESV update cycle (200ms) is different from that of MSV (500ms), cubic spline interpolation is used to resample the ESV to the 500ms timestamp;

[0196] Dimensionality compression: Principal component analysis (PCA) was used to reduce the ESV from 30 dimensions to 15 dimensions and the MSV from 8 dimensions to 5 dimensions, while retaining 95% of the original information.

[0197] Data standardization: Z-score standardization (mean=0, variance=1) was used for ESV, and Min-Max normalization (range [0,1]) was used for MSV.

[0198] Generation of state feature vector set

[0199] The structure of the concatenated State Feature Vector (SFV) is as follows:

[0200] Electrical section (15 dimensions): includes weighted DC ripple, IGBT temperature trend, insulation resistance estimate, etc.

[0201] Mechanical section (5 dimensions): includes PCC_f0, PCC_BPF, bearing failure indicators, etc.;

[0202] Cross-features (2D): such as the interaction term of “current harmonic distortion rate × fundamental frequency vibration amplitude”.

[0203] Each SFV has a total dimension of 22. Example fragment: [0.32, -1.2, 0.05, ..., 0.83, 0.76, 1.2, 0.45, 0.12].

[0204] Intelligent setting of feature validity flag

[0205] To address sensor failure or communication packet loss, the system adds a FeatureValidity Flag (FVF) to each SFV. This flag is an 8-bit binary number.

[0206] Bit0: Current sampling validity (1 for 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 reliability (H calculated sample number>300=1); Bit7: Wavelet coherence spectrum resolution flag.

[0207] For example, when the vibration sensor falls off, Bit3 is set to 0 and a "mechanical feature failure" alarm is triggered, and the system automatically switches to pure electrical analysis mode.

[0208] The original data is mapped to a high-dimensional feature space using kernel functions, and electrical and mechanical features are decoupled using independent component analysis (ICA). The electrical sub-vectors contain parameters such as current harmonic distortion rate, while the mechanical sub-vectors focus on vibration spectrum features, enabling targeted extraction of different physical characteristics. Feature decoupling avoids mutual interference between electrical and mechanical signals, improving the purity of feature representation. The separated sub-vectors are more conducive to subsequent targeted analysis of specific types of faults, such as electrical overload or mechanical loosening.

[0209] 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 anomaly probability set;

[0210] Specifically, a lightweight graph neural network can be deployed on edge computing nodes, using electrical sub-vectors as node attributes and charging pile topology connections as edge weights to construct an electrical state graph structure.

[0211] Edge computing nodes (ECNs) deployed at charging stations utilize embedded hardware with neural network acceleration capabilities (such as NVIDIA Jetson TX2 modules), typically with 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) uses a simplified version of the GraphSAGE architecture, comprising two aggregation layers and one fully connected layer (FCL). The network input consists of the electrical sub-vector (ESV) of each charging pile within its jurisdiction, which serves as the attribute of the graph node. The ESV dimension is typically 12 (including quantized features such as current harmonic distortion rate (THD), voltage drop depth (SagDepth), and insulation resistance (IR). The connection relationship between nodes (edge) 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 electrical coupling strength: when the rated power difference between the two charging 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).

[0212] The construction of the graph structure is dynamic. When a new charging pile is added or a power distribution cabinet is upgraded, the edge nodes automatically identify the topology change and update the Adjacency Matrix (AM) by scanning the Modbus-TCP protocol broadcast packets of the local area network. The first-layer aggregation operation of L-GNN performs "neighbor-weighted averaging": for the target node (such as charging pile #05), it collects the ESV data of its first-order neighbors (such as piles #03, #04, and #07), and sums them according to edge weights (e.g., pile #03 has a weight of 0.9, and pile #04 has a weight of 0.6). The weighted result is concatenated with the target node's own ESV and input into a 128-dimensional fully connected layer (FCL) for nonlinear transformation. The second-layer aggregation extends to second-order neighbors, and the final output is a 64-dimensional Electrical State Embedding Vector (ESEV). This process is executed every 200 milliseconds (ms) to ensure responsiveness to rapid events such as power grid fluctuations.

[0213] The key optimization lies in reducing computational load. Parameter quantization compresses the weights of the L-GNN from 32-bit floating-point to 8-bit integers (INT8), reducing the model size by 75%. Sparse matrix multiplication is used in the aggregation stage, processing only non-zero edge weights (e.g., if a node is connected to only 3 neighbors, the zero-weight calculation for the remaining nodes is skipped). Experiments show that processing a single graph inference at edge nodes takes no more than 15 milliseconds (ms), meeting real-time requirements. The final constructed Electrical State Graph Structure (ESGS) contains three core data items: node attributes (ESEV), edge connectivity (AM), and edge weights (EW), providing a structured electrical state representation for subsequent fusion analysis.

[0214] Spatiotemporal convolution is performed on the mechanical subvectors to aggregate the vibration features of adjacent charging piles in the spatial dimension, and causal convolution is applied in the temporal dimension to generate a mechanical state evolution tensor.

[0215] The Mechanical Sub-Vector (MSV) contains vibration signal features (such as the root mean square (RMS) values ​​of X / Y / Z triaxial accelerations, the bearing fault characteristic frequency amplitude (BPFO_amp), and the cooling fan fundamental frequency component (Fan_Fundamental). Spatio-Temporal Convolution is composed of concatenated spatial and temporal convolutions. Spatial convolution first aggregates features from adjacent charging piles in the spatial dimension: Centered on the target pile (e.g., #05), physically adjacent piles within a 5-meter radius (e.g., #04, #06) are selected, and their MSV data are weighted according to the inverse of their distance (e.g., #04, at 3 meters, has a weight of 1 / 3 ≈ 0.33; #06, at 4 meters, has a weight of 1 / 4 = 0.25). The weighted aggregated features are then concatenated with the target pile's own MSV to form the Spatial Enhanced Mechanical Vector (SEMV).

[0216] The temporal dimension employs causal convolution to process continuous time series data. The time window length is set to 64 sampling points (corresponding to 6.4 seconds of data, with a sampling rate of 10Hz). The kernel size is 3×1 (width×height), with a stride of 1, ensuring 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 times 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%.

[0217] The final generated Mechanical State Evolution Tensor (MSET) has dimensions of [time steps × feature channels × spatial nodes]. For example, when processing 5 adjacent piles, the output tensor shape is 32×32×5 (32 time steps, 32 feature channels, and 5 spatial nodes). This tensor improves fault sensitivity through three properties:

[0218] Spatial correlation: Compare the vibration of a single pile with that of neighboring piles. If the vibration amplitude of the fan of pile #05 suddenly increases to more than twice the average value of the neighboring piles, it is marked as abnormal.

[0219] Time dependence: Causal convolution identifies a continuous upward trend in vibrational energy entropy (EE) (e.g., EE increases by >5% over 10 consecutive windows).

[0220] Multi-scale characteristics: shallow kernels capture transient shocks (such as a 2kHz transient pulse caused by loose bolts), while deep kernels identify slow degradation (such as 80Hz sideband energy growth caused by bearing wear).

[0221] Design a dual-channel attention fusion mechanism, in which the electrical channel focuses on current surge events and the mechanical channel focuses on abnormal vibration frequency bands, and outputs a cross-modal attention weight matrix;

[0222] The Dual-Channel Attention Fusion Mechanism (DCAFM) comprises two independent branches: the Electrical Channel (EC) and the Mechanical Channel (MC). The Attention Generation Module (AGM) of the Electrical Channel specifically detects Current Mutation Events (CMEs): an event is flagged when the rate of change of current, dI / dt, exceeds a threshold (e.g., 100 A / ms). For such events, three key metrics are calculated:

[0223] Mutation Duration (MD);

[0224] Mutation Amplitude (MA);

[0225] Mutation Phase (MP, distinguishing between the start / end / steady-state phases of charging).

[0226] These three metrics are mapped to Electrical Attention Weights (EAW) using a three-layer perceptron (MLP), with weights ranging from 0 to 1. For example, if a sudden change of 300A / ms occurs during the charging termination period (MA=300) and lasts for 50ms (MD=50), EAW=0.92 (high risk).

[0227] The AGM of the mechanical channel focuses on the Abnormal Vibration Frequency Band (AVFB). Four typical fault frequency bands are preset:

[0228] Bearing failure frequency band (BPFI / BPFO, typically 500Hz-2kHz);

[0229] Gear meshing frequency band (GMF, typically 1kHz-4kHz);

[0230] Electromagnetic noise frequency band (EM_Noise, >5kHz);

[0231] Structural resonance band (resonance, characteristic frequency to be identified).

[0232] 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 certain frequency band exceeds the historical baseline by 30%, it is judged as anomalous activation. The mechanical attention weight (MAW) is generated by weighting the frequency band anomaly: if the ER of the BPFO band exceeds the standard by 40% and the ER of the GMF band exceeds the standard by 25% simultaneously, then MAW=0.85.

[0233] The final output Cross-Modality Attention Weight Matrix (CMAWM) is the Cartesian product of EAW and MAW. The matrix has N×M dimensions (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 through a learnable parameter matrix. For example:

[0234] The weights of current surge (EC event type 1) and bearing failure frequency band (MC band 1) are initialized to 0.7 (strong correlation).

[0235] The weights of voltage flicker (EC type 2) and electromagnetic noise band (MC band 3) are initialized to 0.4 (weak correlation).

[0236] The matrix is ​​updated in real time every 100 milliseconds (ms), providing dynamic priority guidance for subsequent fusion.

[0237] Based on the attention weight matrix, a gated cyclic unit is used to fuse the electrical state diagram and the mechanical evolution tensor to generate a multi-dimensional state fusion vector;

[0238] The Gated Recurrent Unit (GRU) uses a two-layer stacked structure with a hidden layer of dimension 128. Its input is a joint representation of the Electrical State Diagram Structure (ESGS) and the Mechanical State Evolution Tensor (MSET).

[0239] Electrical section: Flatten the electrical state embedding vector (ESEV) output by the graph neural network into a 64-dimensional vector;

[0240] Mechanical part: Calculate the mean of MSET over time to generate a 32-dimensional feature vector;

[0241] Concatenate the two to obtain a 96-dimensional raw fusion vector (RFV).

[0242] The fusion process incorporates a dynamically modulated attention weight matrix (CMAWM). This involves three steps:

[0243] Step 1 - Electrical Feature Weighting: Based on the currently detected electrical event type (e.g., CME type 3), extract the corresponding row vector W_i* (dimension 1×M) from CMAWM and scale the 64-dimensional features of ESEV.

[0244] Step 2 - Mechanical Feature Filtering: Based on the activation state of mechanical abnormal frequency bands (such as frequency bands 2 and 4 activation), extract the corresponding column vector W_*j (dimension N×1) from CMAWM, and multiply it element-wise with the 32-dimensional features of MSET.

[0245] Step 3 - Gated Fusion: The weighted electrical features and filtered mechanical features are then concatenated again and input into the GRU unit. The GRU's update gate determines how much historical state is retained, while the reset gate controls the proportion of new features written. For example, when mechanical vibration suddenly increases (MSET energy rises by 40%) and there are no sudden electrical changes, the reset gate opens 70% of the channels to write new data.

[0246] The output of the GRU is compressed to 48 dimensions through a fully connected layer, generating a Multi-Dimensional State Fusion Vector (MDSFV). This vector has three main characteristics:

[0247] Temporal memory: The state evolution trajectory of the past 120 seconds is retained through the GRU hidden state;

[0248] Cross-modal correlation: Current abrupt events (such as dI / dt=150A / ms) and high-frequency vibrations of bearings (>1kHz energy) are mapped to similar regions in the vector space;

[0249] Enhanced fault sensitivity: Experiments show that the detection sensitivity of cooling fan jamming faults has increased from 73% in the single mode to 92% after fusion.

[0250] The vector is updated every 200 milliseconds (ms) and serves as the input baseline for anomaly detection.

[0251] Input the multidimensional state fusion vector into the anomaly detection forest, and output a preliminary set of anomaly probabilities for each dimension.

[0252] The Anomaly Detection Forest (ADF) consists of 100 deeply randomized isolation trees (iTrees). Each tree randomly selects four feature dimensions of the MDSFV (e.g., dimensions 12 / 25 / 33 / 41) for node splitting. The splitting threshold is randomly generated, following a rule such as "whether the value of dimension 12 is greater than 0.7". ADF uses unsupervised training, requiring only normal state data (NSD) for training. During training, 5% Gaussian noise is injected into each iTree to improve model robustness.

[0253] The anomaly probability calculation is performed in three steps:

[0254] Step 1 - Path Length Calculation: For the input MDSFV vector, count the number of edges it traverses from the root node to the leaf node in each iTree (i.e., path length, PL).

[0255] Step 2 - Anomaly Score Conversion: Based on the average PL of all iTrees in the forest, calculate the anomaly score (AS) using the formula S=2^(-E(PL) / c(n)), where c(n) is the path length normalization factor (e.g., c=12.8 when n=1000).

[0256] Step 3 - Dimensional Probability Generation: Feature Contribution Decomposition (FCD) is employed. For the target dimension d (e.g., vibration energy dimension 25), its value is perturbed by ±10% to generate two new vectors. The difference in AS before and after the perturbation, ΔAS, is compared. A larger ΔAS indicates a higher anomalous contribution from that dimension, which is ultimately mapped to the Dimensional Anomaly Probability (DAP), with the formula P_d = sigmoid(5×ΔAS).

[0257] The final output Preliminary Anomaly Probability Set (PAPS) contains 48 probability values ​​(aligned with the MDSFV dimension). Each probability value P_d represents the anomaly confidence level (range 0-1) for the corresponding feature dimension. The system sets two levels of thresholds:

[0258] Warning Threshold (WT=0.65): A yellow warning is issued when the single-dimensional P_d > 0.65;

[0259] Alarm Threshold (AT=0.85): A red alarm is triggered when P_d > 0.85 for three or more dimensions or P_d > 0.95 for a single dimension.

[0260] For example, when "current harmonic THD dimension probability = 0.92" and "bearing vibration sideband dimension probability = 0.88" are detected, an alarm event for IGBT aging of the charging module is generated.

[0261] A lightweight analysis model is deployed at edge nodes, using graph neural networks to process electrical features and spatiotemporal convolution to process mechanical features, followed by cross-modal fusion via an attention mechanism. This process fully leverages the low latency of edge computing to achieve millisecond-level anomaly detection. The distributed architecture meets real-time requirements and avoids cloud transmission delays. Multimodal fusion enhances detection comprehensiveness, enabling the simultaneous identification of complex faults such as contactor arcing (electrical) and fan imbalance (mechanical).

[0262] S204, perform multi-objective optimization on the preliminary anomaly probability set, apply reinforcement learning algorithm to optimize real-time responsiveness and fault prediction accuracy, and generate an optimized state index set;

[0263] Specifically, a dual-objective reward function can be constructed, where real-time responsiveness is rewarded for processing latency, and prediction accuracy is rewarded for fault detection F1 score.

[0264] The system first defines a dual-objective reward function as the evaluation benchmark for reinforcement learning. This function contains two core reward terms:

[0265] 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 anomaly probability. RRR is designed as a negative correlation function; for example, when PL ≤ 5ms, the reward is +10 points, and the reward decreases linearly by 2 points for every 1ms increase in PL. When PL > 20ms, the reward is zero. This design forces the system to prioritize low latency.

[0266] Prediction Accuracy Reward (PAR): Its calculation is tied to 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, a reward of +15 points is given when FDF1 ≥ 0.95, the reward decreases by 3 points for every 0.01 decrease in FDF1, and the reward is zero when FDF1 < 0.8.

[0267] The reward function needs to resolve conflicts between multiple objectives: for example, a higher sampling frequency can improve FDF1 (increase PAR) but will increase computational load (reduce RRR). The system introduces a Dynamic Weight Factor (DWF) for arbitration.

[0268] During peak charging periods (identified through historical data), the RRR weight is set to 0.7 and the PAR weight to 0.3 to ensure that response speed is prioritized.

[0269] During the equipment aging phase (determined by cumulative runtime), the PAR weight is set to 0.8 and the RRR weight to 0.2, focusing on the accuracy of fault prediction.

[0270] The formula for calculating the total reward (TR) is: TR = DWF_RRR × RRR + DWF_PAR × PAR. This formula transforms multi-objective optimization into a single-objective maximization problem.

[0271] The reward function is linked to physical constraints: for example, when the internal temperature of the charging pile exceeds a safety threshold (such as 65℃), the RRR weight is forcibly set to 1.0, triggering the Downclocking Operation Mode (DOM), temporarily sacrificing some accuracy to reduce CPU load and heat dissipation pressure. All weight policies and thresholds are stored in a configurable rule base (CRB), allowing maintenance personnel to adjust them as needed.

[0272] Design a deep Q-network architecture, using the set of anomaly probabilities as the state input and the feature sampling frequency as the action space, and estimate the value function through a two-stream network;

[0273] Deep Q-Network (DQN) uses a dual-stream architecture (DSA) to process state inputs:

[0274] The State Input (SI) is a 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].

[0275] Action space (AS) is defined as the combination of feature sampling frequencies:

[0276] Current / voltage sampling frequency (IV-Sampling Frequency, IV-SF): Selectable from 10kHz, 20kHz, and 50kHz.

[0277] Vibration signal sampling frequency (Vib-Sampling Frequency, Vib-SF): 1kHz, 2kHz, and 5kHz are selectable.

[0278] Temperature sampling frequency (Temp-Sampling Frequency, Temp-SF): fixed at 1Hz (because temperature changes slowly).

[0279] The specific structure of the two-stream network is as follows:

[0280] Real-time Response Stream (RSS): Input PAPS, extract features through two fully connected layers (128 → 64 neurons), and output Response Value Estimate (RVE), which represents the expected delayed reward that can be obtained by taking a certain action.

[0281] Accuracy Prediction Stream (APS): Input PAPS is processed by a Long Short-Term Memory (LSTM) network (32 units) to capture temporal dependencies and outputs an Accuracy Value Estimate (AVE), which represents the expected accuracy reward that can be obtained by taking a certain action.

[0282] The dual-stream outputs are spliced ​​in the Fusion Layer (FL) and then generated into the final Q-Value Matrix (QVM) through a fully connected layer (number of neurons = 32). The matrix dimension is 3 (IV-SF option) × 3 (Vib-SF option) = Q values ​​of 9 actions.

[0283] The network training employs 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, new state S_t+1="[0.68,0.18,0.05]") in a replay buffer (RB). Each time, 32 sets of data (batch size=32) are randomly selected to update the network parameters, avoiding training oscillations caused by temporal correlations. The target network synchronizes the main network weights every 1000 steps to improve stability.

[0284] A non-dominated sorting strategy is adopted to select the action strategy at the Pareto optimal front, and the current sampling frequency and vibration analysis granularity are dynamically adjusted.

[0285] Non-dominated Sorting Strategy (NDSS) is used to resolve action selection in multi-objective conflicts:

[0286] Calculate the biobjective score for each of the nine action options in the current Q-value matrix:

[0287] Objective 1: Real-time responsiveness score (mapped from RVE to a percentage reduction in latency, e.g., RVE=15 → 12% reduction in latency).

[0288] Objective 2: Prediction accuracy score (mapped from AVE to F1 score improvement value, e.g., AVE=20 → F1 score improvement of 0.05).

[0289] Perform Pareto Ranking (PR):

[0290] Level 1 (Pareto optimal frontier): All actions that are not dominated by other actions. 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 A.

[0291] Level 2: A collection of actions that are governed by Level 1 actions but govern other lower-level actions, and so on.

[0292] Rules for selecting the best action in the Pareto front:

[0293] Crowding Distance Calculation (CDC): For each action within the frontier, calculate the sum of its Euclidean distances to its neighboring actions in the target space. A larger distance indicates a sparser (better) decision space for that action.

[0294] Dynamic Preference Injection (DPI): Adjusts preference weights based on the runtime scenario.

[0295] If the charging power is > 150kW (fast charging mode), select the high responsiveness action with the greatest crowding distance (e.g., IV-SF=50kHz).

[0296] If mechanical abnormalities are detected three times consecutively (MAP > 0.6), select a high-precision action (e.g., Vib-SF = 5kHz).

[0297] The final output is an optimized action policy (OAP), such as "IV-SF=50kHz, Vib-SF=2kHz".

[0298] Action execution is linked to analysis granularity (AG):

[0299] 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 bands), thereby improving the accuracy of abnormal frequency band identification.

[0300] When the current sampling frequency drops to 10kHz, Dimensionality Reduction Feature Extraction (DRFE) is triggered. Principal Component Analysis (PCA) is used to compress the electrical feature dimension from 128 to 64, reducing computational load. All granular adjustments are recorded in the Policy Execution Log (PEL) for auditing purposes.

[0301] The network parameters are updated using a near-end policy optimization algorithm, and the optimized set of state indicators is output.

[0302] The Proximal Policy Optimization (PPO) algorithm is responsible for network updates, and its core is the Clipped Objective Function (COF):

[0303] Calculate the probability ratio between the old and new strategies:

[0304] Old Policy (OP): The probability of the network choosing historical action A_t before the update (e.g., probability = 0.65).

[0305] New Policy (NP): The probability of the network choosing the same action A_t after the update (e.g., probability = 0.72).

[0306] The probability ratio (PR) = NP / OP = 0.72 / 0.65 ≈ 1.108.

[0307] Construct the basic objective function: Advantage Function (AF) × PR. AF is calculated using Generalized Advantage Estimation (GAE), which combines immediate reward and future value (discount factor γ = 0.99).

[0308] Apply a pruning mechanism to prevent policy mutations:

[0309] Set the clipping threshold (CT) to ±0.2.

[0310] If PR > 1.2 (i.e., NP is more than 20% higher than OP), force the PR in the objective function to be replaced with 1.2.

[0311] If PR < 0.8 (i.e., NP is more than 20% lower than OP), force PR to be replaced with 0.8.

[0312] For example, if PR = 1.108, the threshold is not exceeded, and the objective function value is AF × 1.108; if PR = 1.25, it is calculated as 1.2. This mechanism ensures that the step size of each parameter update is controlled, avoiding training divergence.

[0313] Output Optimized State Indicator Set (OSIS):

[0314] Metric generation: The optimized DQN is applied to the real-time data stream to output the state evaluation results corresponding to the action policy, including:

[0315] Electrical Health Index (EHI): Range 0-100, >80 is normal.

[0316] Mechanical Vibration Severity (MVS): Range 0-10, <2 is normal.

[0317] Composite Risk Score (CRS): range 0-1, <0.3 triggers an alert.

[0318] Dynamic calibration: OSIS is backtested monthly using historical fault data. If FDF1 remains below 0.85 for an extended period, the Network Retraining Procedure (NRP) is automatically initiated to update the DQN weights. Finally, OSIS is pushed to the cloud-based diagnostic platform via a message queue (such as the MQTT protocol).

[0319] A dual-objective reward function is constructed, dynamically balancing detection speed and accuracy through deep reinforcement learning. The algorithm automatically adjusts the sampling frequency and analysis granularity, selecting the optimal strategy from the Pareto optimal solution set to achieve self-optimization of detection performance. This resolves the contradiction between response speed and accuracy in traditional methods and adapts to different operating conditions. The self-learning characteristic reduces reliance on manual parameter tuning, improving system adaptability.

[0320] S205, perform spatiotemporal consistency verification and confidence screening on the optimized state indicator set, determine the final operating state level, and output a real-time monitoring alarm.

[0321] Specifically, a spatiotemporal consistency verification window can be constructed. In the time dimension, an exponentially weighted moving average is used to filter out instantaneous interference, and in the spatial dimension, the Mahalanobis distance of similar indicators of adjacent charging piles is compared.

[0322] After receiving the set of optimized status indicators, the system first constructs a Spatio-Temporal Consistency Check Window (STCCW). This window is a dynamic data buffer that covers the indicator data for the most recent 20 seconds in the time dimension (window length is configurable) and associates the current charging pile with the three physically adjacent charging piles in the spatial dimension (based on a preset topology map). The time dimension processing uses 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 (meaning that new data has a weight of 0.8, and historical data has an exponentially decaying weight), and the EWMA value is calculated independently for each indicator (such as DC bus voltage fluctuation rate). For example, if a current harmonic distortion rate indicator experiences a transient spike within a window (potentially caused by grid interference), EWMA reduces the weight of older data (e.g., data from 10 seconds ago has only 10.7% of its initial weight) to limit the spike's impact locally, outputting a smoothed trend value (STV) to effectively filter out transient noise. Simultaneously, the system records the standard deviation (SD) of each indicator within the window for subsequent confidence assessment.

[0323] The core of spatial dimension verification is calculating the Mahalanobis distance (MD). For the same type of metric (e.g., the "IGBT module temperature rise rate" of all adjacent piles), the system collects the STV data of each pile within the current verification window to form a multidimensional dataset. First, the global covariance matrix (GCM) of this type of metric is calculated. This matrix is ​​obtained through training with historical normal data and represents the correlation between the metrics of each pile (e.g., the correlation strength between the temperature rise of pile A and the cooling fan speed of pile B). For a certain metric value of the current target charging pile (e.g., temperature rise rate STV = 1.5°C / min), the Mahalanobis distance between it and the mean vector (MV) of the metric group of adjacent piles is calculated: MD = √[(target value - group mean) × inverse of covariance matrix × (target value - group mean) transpose].

[0324] This distance value is dimensionless. If the MD is greater than a preset threshold (e.g., 3.0), it indicates that the pile's performance deviates significantly from the behavior of adjacent pile groups (possibly due to its own malfunction). For example, when a pile's cooling fan malfunctions, its temperature rise rate MD value may reach 4.2, while the MD of a normal pile group is usually less than 1.5. The spatial verification result is output as the Spatial Deviation Coefficient (SDC), defined as the ratio of MD to the threshold.

[0325] The spatio-temporal verification results need to be dynamically fused. The system generates a spatio-temporal consistency score (STCS) for each indicator, calculated as follows: STCS = temporal smoothing factor × (1 / intra-window SD) + spatial alignment factor × (1 / SDC).

[0326] The Time Smoothness Factor (TSF) is the reciprocal of the EWMA residual (the difference between the actual value and the STV); the smaller the residual, the larger the TSF. The Spatial Alignment Factor (SAF) is the reciprocal of the standardized SDC. The STCS value ranges from 0 to 1.0; the closer it is to 1, the more consistent the indicator is in the spatiotemporal dimensions (e.g., STCS = 0.92 for a voltage ripple coefficient). Indicators below the threshold (e.g., 0.6) are marked as "spatiotemporal anomaly candidates," triggering subsequent deep verification. All calculations are performed at the edge nodes, taking less than 50 milliseconds (ms) to ensure real-time performance.

[0327] Design a confidence screening funnel, in which the first layer removes high volatility indicators, the second layer retains stable features, and generates a subset of high confidence indicators;

[0328] The Confidence Screening Funnel (CSF) is a two-stage filtering architecture. The first layer, the Volatility Filtering Layer (VFL), calculates the Dynamic Volatility Index (DVI) for all indicators that pass the spatiotemporal validation. The DVI consists of three elements:

[0329] Short-Term Coefficient of Variation (STCV): The ratio of the standard deviation to the mean of the indicator value within the most recent 5 seconds;

[0330] Historical Volatility Deviation (HVD): The deviation of the current STCV from the baseline value under the same operating conditions over the past 24 hours.

[0331] Sensor Health Index (SHI): Quantifies the reliability of a voltage / temperature sensor based on its self-diagnostic signals, such as zero-point drift.

[0332] For example, if a vibration acceleration index has an STCV of 0.35 (normally <0.1), an HVD of 2.8 times (baseline 0.12), and a SHI of 0.7 (out of 1.0), then DVI = STCV × 0.5 + HVD × 0.3 + (1-SHI) × 0.2 = 0.52. If DVI > 0.4 (threshold adjustable), it is judged as a high volatility index and removed.

[0333] The second stable feature preservation layer (SFPL) performs feature stability authentication (FSA) on the remaining indicators. This step introduces sliding window autocorrelation analysis (SWAA): with a step size of 1 second, the autocorrelation coefficient (AC) of the indicator sequence is calculated within a 10-second window. If AC > 0.8 (strong autocorrelation), it indicates that the indicator changes regularly (e.g., the cooling fan speed increases linearly with increasing temperature), and is judged as a stable feature. Simultaneously, combined with the indicator predictability score (PS), the indicator sequence is fitted using an ARIMA model (autoregressive integral moving average model), and the root mean square error (RMSE) of the prediction residuals is calculated. The smaller the RMSE, the higher the PS. For example, the PS of DC output current can reach 0.95 (RMSE = 0.5A), while the PS of an indicator affected by environmental humidity may only be 0.6.

[0334] The final high-confidence indicator subset (HCIS) is generated. Each indicator in the subset is given an integrated confidence tag (ICT), which is generated by the following formula: ICT = 0.4 × STCS + 0.3 × (1 - DVI) + 0.3 × PS.

[0335] The label value ranges from 0 to 1.0, and only metrics with ICT > 0.75 are retained. For example, the metrics retained after filtering may 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 more than 60% compared to the original set, but the reliability is significantly improved, providing clean input for status level determination.

[0336] The system uses a fuzzy inference engine, takes a subset of high-confidence indicators as input, and outputs a state level matrix through a preset rule base.

[0337] The core of a fuzzy inference engine (FIE) is a predefined rule base (PRB), which contains three types of rules:

[0338] Electrical safety rules: For example, "If the probability of DC overvoltage is > 0.7 and the insulation resistance value is < 500kΩ, then the arc risk level is high";

[0339] Mechanical health rules: For example, "If the vibration energy in the 4kHz frequency band is > 0.15g..." 2 If the bearing temperature is > 85°C and the Hz is / Hz, then the mechanical wear level is "severe";

[0340] Thermal management rules: For example, "If the radiator temperature difference is > 15°C and the fan speed deviates from the set value by > 20%, then the cooling failure level is medium."

[0341] Each rule is described using fuzzy logic. The input metric (such as "overpressure probability") is mapped to a fuzzy set (such as "low / medium / high") through a membership function (MF). For example, the MF for overpressure 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).

[0342] The engine performs multi-stage inference (MSI):

[0343] Step 1: Fuzzification: Convert the precise values ​​in HCIS (e.g., "bearing temperature = 88°C") into fuzzy sets. By querying the temperature membership function (e.g., "normal" < 80°C, "warning" 80-90°C, "danger" > 85°C), 88°C is simultaneously assigned a membership degree of 0.6 for "warning" and 0.4 for "danger".

[0344] Step 2: Rule Activation: Traverse the PRB and match the fuzzy set of the condition part (IF part). For example, for the rule "If bearing temperature = dangerous and vibration energy = high, then fault level = emergency", since temperature "dangerous" has a membership degree of 0.4 and vibration energy "high" has a membership degree of 0.8, the minimum value of 0.4 is taken as the activation strength of this rule.

[0345] Step 3: Defuzzification: For all activation rule results of the output variable (such as "fault level"), calculate the precise value using the centroid method. For example, if "emergency" level is activated at 0.4 and "high risk" is activated at 0.3, the output value = (0.4×90 + 0.3×70) / (0.4+0.3) = 81.4 (out of 100).

[0346] The final output is the State Level Matrix (SLM). This matrix is ​​a 6-row, 3-column two-dimensional structure:

[0347] The row represents six core dimensions: electrical safety, mechanical health, thermal management, communication status, efficiency indicators, and environmental adaptability;

[0348] The columns represent three levels: Normal (0-30 points), Warning (31-70 points), and Alarm (71-100 points).

[0349] Each cell is filled with the deblurred score for that dimension (e.g., thermal management = 65 points, falling into the "Warning" column). The matrix is ​​supplemented with a Global Health Index (GHI), which is derived from the weighted average of the scores of each dimension (electrical weight 0.3, mechanical weight 0.25, thermal management weight 0.2, etc.). For example, GHI = 58 means that the system as a whole is in a warning state.

[0350] The system optimizes alarm distribution by adding a timestamp hash chain to high-risk alarms and pushing them to the operation and maintenance terminal via a priority queue, while simultaneously triggering the device power reduction protection protocol.

[0351] 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 then added to the HRA.

[0352] Generate the current alarm timestamp (e.g., 2025-04-11 14:23:05.876);

[0353] Extract the hash value of the preceding alarm (the first alarm uses the initial value 0x0000);

[0354] Concatenate the current timestamp, alarm content, and previous hash value, and input the SHA-256 algorithm to generate a new hash (e.g., 0x9f86d081...).

[0355] The new hash and timestamp are embedded in the alarm message. THC ensures that the alarm sequence is immutable (e.g., maintenance personnel cannot delete the "IGBT overheat" record) and that the event chain is traceable (e.g., "fan failure → temperature surge → power reduction").

[0356] Alerts are pushed via a Priority Queue (PQ). The queue defines four priority levels:

[0357] P0 (Emergency): Electrical safety / mechanical health alarm (such as arc risk, bearing breakage), delay requirement < 1 second;

[0358] P1 (High Risk): Thermal management alarm or dual-dimensional warning (such as cooling failure + sudden efficiency drop), delay < 3 seconds;

[0359] P2 (Warning): Single-dimensional early warning or environmental anomaly (such as excessive humidity), delay < 10 seconds;

[0360] P3 (Notification): Performance degradation or maintenance reminder (e.g., contactor operation count nearing the end of its lifespan), delay < 60 seconds.

[0361] The queue uses a weighted fair queuing (WFQ) algorithm, allocating 50% of the bandwidth to P0, 30% to P1, and 20% to the rest. Push targets include maintenance personnel's mobile apps (via MQTT protocol), cloud monitoring platforms (HTTPS API), and local industrial control screens (WebSocket).

[0362] The Derating Power Protection Protocol (DP) is used to synchronously trigger P0 / P1 level alarms. 3 ):

[0363] Gradual current reduction: If it is an over-temperature alarm, reduce the output current at a rate of 50A per second (e.g., from 200A to 150A) until the temperature drops back to within the threshold.

[0364] Voltage clamping: If an overvoltage alarm is detected, the charging pile output voltage will be forced to not exceed 105% of the vehicle's requested value;

[0365] Safety lockout: If an insulation fault (<100kΩ) is detected, immediately disconnect the DC contactor and lock out the restart mechanism;

[0366] Power reduction record: The power reduction magnitude (e.g., 30%) and duration are recorded on the blockchain for evidence storage and used for accident liability tracing. SLM is continuously monitored during protocol execution. If GHI drops below 60, power is gradually restored (incrementing by 20A per second).

[0367] Spatiotemporal filtering is employed to eliminate transient interference, and the status of adjacent devices is compared using Mahalanobis distance. A fuzzy inference engine is then used to comprehensively assess the risk level. Confirmed high-risk faults are marked with tamper-proof timestamps, triggering a tiered response mechanism. Spatiotemporal verification effectively suppresses false alarms, and confidence level filtering ensures alarm reliability. The tiered response mechanism balances operational efficiency with device safety, avoiding over-protection that could negatively impact the user's charging experience.

[0368] As can be seen, the system receives real-time operating parameter data from DC charging piles for automobiles and integrates it into a raw multidimensional dataset. Nonlinear feature mapping is then applied to the raw multidimensional dataset to construct a set of state feature vectors, decomposing each state feature vector into electrical and mechanical sub-vectors. A distributed edge computing framework is used to perform real-time fusion analysis of the electrical and mechanical sub-vectors, generating a preliminary anomaly probability set. This preliminary anomaly probability set is then optimized using multi-objective methods to generate an optimized state index set. Finally, the optimized state index set undergoes spatiotemporal consistency verification and confidence level filtering to determine the final operating state level and output real-time monitoring alarms, thereby improving the accuracy and real-time performance of fault detection.

[0369] Another embodiment of the present invention provides a real-time monitoring system for the operating status of a DC charging pile for automobiles, see [link to relevant documentation]. Figure 3 The system may include:

[0370] The receiving module 301 is used to receive real-time operating parameter data from the DC charging pile for automobiles, collect current, voltage, temperature and vibration signals based on a preset high-frequency sampling protocol, and integrate them into a raw multidimensional dataset.

[0371] The mapping module 302 is used to perform nonlinear feature mapping on the original multidimensional dataset, construct a set of state feature vectors, and decompose each state feature vector into electrical sub-vectors and mechanical sub-vectors to separate and capture electrical characteristics and mechanical vibration characteristics.

[0372] Analysis module 303 is used to perform real-time fusion analysis on the electrical sub-vector and the mechanical sub-vector using a distributed edge computing framework to generate a preliminary set of anomaly probabilities;

[0373] Optimization module 304 is used to perform multi-objective optimization on the preliminary anomaly probability set, apply reinforcement learning algorithm to optimize real-time responsiveness and fault prediction accuracy, and generate an optimized state index set.

[0374] The filtering module 305 is used to perform spatiotemporal consistency verification and confidence filtering on the optimized state index set, determine the final operating state level, and output real-time monitoring alarms.

[0375] This invention also provides a storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0376] Specifically, in this embodiment, the storage medium can be configured to store a computer program for performing the following steps:

[0377] S201 receives real-time operating parameter data from the DC charging pile for automobiles, collects current, voltage, temperature and vibration signals based on a preset high-frequency sampling protocol, and integrates them into a raw multidimensional dataset.

[0378] S202, perform nonlinear feature mapping on the original multidimensional dataset to construct a set of state feature vectors, and decompose each state feature vector into electrical sub-vectors and mechanical sub-vectors to separate and capture electrical characteristics and mechanical vibration characteristics.

[0379] 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 anomaly probability set;

[0380] S204, perform multi-objective optimization on the preliminary anomaly probability set, apply reinforcement learning algorithm to optimize real-time responsiveness and fault prediction accuracy, and generate an optimized state index set;

[0381] S205, perform spatiotemporal consistency verification and confidence screening on the optimized state indicator set, determine the final operating state level, and output a real-time monitoring alarm.

[0382] This invention also provides an electronic device, including 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 of the above method embodiments.

[0383] Specifically, the aforementioned electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the aforementioned processor, and the input / output device is connected to the aforementioned processor.

[0384] Specifically, in this embodiment, the processor can be configured to perform the following steps via a computer program:

[0385] S201 receives real-time operating parameter data from the DC charging pile for automobiles, collects current, voltage, temperature and vibration signals based on a preset high-frequency sampling protocol, and integrates them into a raw multidimensional dataset.

[0386] S202, perform nonlinear feature mapping on the original multidimensional dataset to construct a set of state feature vectors, and decompose each state feature vector into electrical sub-vectors and mechanical sub-vectors to separate and capture electrical characteristics and mechanical vibration characteristics.

[0387] 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 anomaly probability set;

[0388] S204, perform multi-objective optimization on the preliminary anomaly probability set, apply reinforcement learning algorithm to optimize real-time responsiveness and fault prediction accuracy, and generate an optimized state index set;

[0389] S205, perform spatiotemporal consistency verification and confidence screening on the optimized state indicator set, determine the final operating state level, and output a real-time monitoring alarm.

[0390] The above description, based on the embodiments shown in the figures, details the structure, features, and effects of the present invention. The above description is only a preferred embodiment of the present invention, but the present invention is not limited to the scope of implementation shown in the figures. Any changes made in accordance with the concept of the present invention, or equivalent embodiments modified to have equivalent changes, that do not exceed the spirit covered by the specification and figures, should be within the protection scope of the present invention.

Claims

1. A method for real-time monitoring of the operating status of a DC charging pile for automobiles, characterized in that, The method includes: It receives real-time operating parameter data from DC charging piles for automobiles, collects current, voltage, temperature and vibration signals based on a preset high-frequency sampling protocol, and integrates them into a raw multidimensional dataset. The original multidimensional dataset is subjected to nonlinear feature mapping to construct a set of state feature vectors. Each state feature vector is decomposed into electrical sub-vectors and mechanical sub-vectors to separate and capture electrical characteristics and mechanical vibration characteristics. A distributed edge computing framework is used to perform real-time fusion analysis on the electrical sub-vectors and the mechanical sub-vectors to generate a preliminary set of anomaly probabilities. Specifically, a lightweight graph neural network is deployed on the edge computing nodes, with the electrical sub-vectors as node attributes and the charging pile topology connection relationship as edge weights to construct an electrical state graph structure. Spatiotemporal convolution operations are performed on the mechanical sub-vectors to aggregate the vibration features of adjacent charging piles in the spatial dimension and apply causal convolution in the temporal dimension to generate a mechanical state evolution tensor. A dual-channel attention fusion mechanism is designed, in which the electrical channel focuses on current surge events and the mechanical channel focuses on abnormal vibration frequency bands, outputting a cross-modal attention weight matrix. Based on the attention weight matrix, a gated cyclic unit is used to fuse the electrical state diagram 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 to output a preliminary set of anomaly probabilities for each dimension. The preliminary anomaly probability set is optimized using a multi-objective approach, and a reinforcement learning algorithm is applied to optimize real-time responsiveness and fault prediction accuracy, thereby generating an optimized set of state indicators. The set of optimized status indicators is subjected to spatiotemporal consistency verification and confidence screening to determine the final operating status level and output real-time monitoring alarms.

2. The method according to claim 1, characterized in that, The system receives real-time operating parameter data from the DC charging pile for automobiles, collects current, voltage, temperature, and vibration signals based on a preset high-frequency sampling protocol, and integrates them into a raw multidimensional dataset, including: High-frequency waveforms of current and voltage are simultaneously acquired using a multi-channel anti-aliasing filter, while three-dimensional vibration signals are captured using a MEMS accelerometer to generate a heterogeneous sensing flow with timestamps. An adaptive Kalman filter is used to align the heterogeneous sensor streams in the time domain, eliminating timing deviations caused by differences in the physical locations of the sensors and outputting a synchronized sensing matrix. Wavelet packet decomposition is performed on the synchronized sensing matrix to extract energy entropy features in a specific frequency band and construct an initial multidimensional feature tensor. The initial multidimensional feature tensor is reconstructed, and redundant dimensions are compressed through Tucker decomposition to generate a compact original multidimensional dataset.

3. The method according to claim 2, characterized in that, The step of performing nonlinear feature mapping on the original multidimensional dataset to construct a set of state feature vectors, and decomposing each state feature vector into electrical sub-vectors and mechanical sub-vectors to separate and capture electrical characteristics and mechanical vibration characteristics, includes: The original multidimensional dataset 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. Orthogonal feature decoupling is performed in a high-dimensional state feature space, and the FastICA algorithm is used to separate electrical and mechanical features. Information entropy weighting is applied to the decoupled electrical features to generate physically interpretable electrical sub-vectors; Wavelet coherence analysis is performed on the mechanical characteristics to extract the phase coupling coefficient between the vibration signal and the cooling fan speed, and a mechanical sub-vector is constructed. The electrical and mechanical sub-vectors are combined into a set of state feature vectors through a feature cascade layer, and a feature validity flag is added.

4. The method according to claim 3, characterized in that, The step of performing multi-objective optimization on the preliminary anomaly probability set, applying reinforcement learning algorithms to optimize real-time responsiveness and fault prediction accuracy, and generating an optimized state index set includes: Construct a dual-objective reward function, in which real-time responsiveness is rewarded for processing latency, and prediction accuracy is rewarded for fault detection F1 score. Design a deep Q-network architecture, using the set of anomaly probabilities as the state input and the feature sampling frequency as the action space, and estimate the value function through a two-stream network; A non-dominated sorting strategy is adopted to select the action strategy at the Pareto optimal front, and the current sampling frequency and vibration analysis granularity are dynamically adjusted. The network parameters are updated using a near-end policy optimization algorithm, and the optimized set of state indicators is output.

5. The method according to claim 4, characterized in that, The process of performing spatiotemporal consistency verification and confidence screening on the optimized state indicator set to determine the final operating state level and outputting real-time monitoring alarms includes: A spatiotemporal consistency verification window is constructed. In the time dimension, an exponentially weighted moving average is used to filter out instantaneous interference, and in the spatial dimension, the Mahalanobis distance of the same indicators of adjacent charging piles is compared. Design a confidence screening funnel, in which the first layer removes high volatility indicators, the second layer retains stable features, and generates a subset of high confidence indicators; The system uses a fuzzy inference engine, takes a subset of high-confidence indicators as input, and outputs a state level matrix through a preset rule base. The system optimizes alarm distribution by adding a timestamp hash chain to high-risk alarms and pushing them to the operation and maintenance terminal via a priority queue, while simultaneously triggering the device power reduction protection protocol.

6. A real-time monitoring system for the operating status of a DC charging pile for automobiles, characterized in that, The system includes: The receiving module is used to receive real-time operating parameter data from the DC charging pile for automobiles, collect current, voltage, temperature and vibration signals based on a preset high-frequency sampling protocol, and integrate them into a raw multidimensional dataset. The mapping module is used to perform nonlinear feature mapping on the original multidimensional dataset, construct a set of state feature vectors, and decompose each state feature vector into electrical sub-vectors and mechanical sub-vectors to separate and capture electrical characteristics and mechanical vibration characteristics. The analysis module is used to perform real-time fusion analysis of the electrical sub-vectors and the mechanical sub-vectors using a distributed edge computing framework to generate a preliminary set of anomaly probabilities. Specifically, a lightweight graph neural network is deployed on the edge computing nodes, using the electrical sub-vectors as node attributes and the charging pile topology connections as edge weights to construct an electrical state graph structure. Spatiotemporal convolution operations are performed on the mechanical sub-vectors to aggregate vibration features of adjacent charging piles in the spatial dimension and apply causal convolution in the temporal dimension to generate a mechanical state evolution tensor. A dual-channel attention fusion mechanism is designed, in which the electrical channel focuses on current surge events and the mechanical channel focuses on abnormal vibration frequency bands, outputting a cross-modal attention weight matrix. Based on the attention weight matrix, a gated cyclic unit is used to fuse the electrical state diagram 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 to output a preliminary set of anomaly probabilities for each dimension. The optimization module is used to perform multi-objective optimization on the preliminary anomaly probability set, apply reinforcement learning algorithms to optimize real-time responsiveness and fault prediction accuracy, and generate an optimized state index set. The filtering module is used to perform spatiotemporal consistency verification and confidence filtering on the set of optimized status indicators, determine the final operating status level, and output real-time monitoring alarms.

7. The system according to claim 6, characterized in that, The receiving module is specifically used for: High-frequency waveforms of current and voltage are simultaneously acquired using a multi-channel anti-aliasing filter, while three-dimensional vibration signals are captured using a MEMS accelerometer to generate a heterogeneous sensing flow with timestamps. An adaptive Kalman filter is used to align the heterogeneous sensor streams in the time domain, eliminating timing deviations caused by differences in the physical locations of the sensors and outputting a synchronized sensing matrix. Wavelet packet decomposition is performed on the synchronized sensing matrix to extract energy entropy features in a specific frequency band and construct an initial multidimensional feature tensor. The initial multidimensional feature tensor is reconstructed, and redundant dimensions are compressed through Tucker decomposition to generate a compact original multidimensional dataset.

8. A storage medium, characterized in that, The storage medium stores a computer program, wherein the computer program is configured to execute the method of any one of claims 1-5 when it is run.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the method of any one of claims 1-5.

Images