Method and system for detecting aging and deterioration of surge protector

Abstract

The invention discloses a surge protector aging and deterioration detection method and system, and the method comprises the steps: injecting a non-destructive composite test signal into a protection line of a surge protector SPD when the surge protector SPD is in an online operation state, and collecting a real-time multi-physical field response signal of the SPD to the composite test signal; performing multi-frequency-band parallel decomposition on the acquired multi-physical-field response signals to obtain a key-frequency-band sub-signal set reflecting different degradation mechanism characteristics; synchronously extracting time domain and frequency domain characteristic parameters from each key frequency band sub-signal, and constructing a characteristic vector representing the SPD multi-dimensional aging state; inputting the feature vector into a pre-trained lightweight depth evaluation model, and outputting a comprehensive aging score reflecting the overall degradation degree of the SPD in real time; and generating a structured detection report in combination with the comprehensive aging score and a preset degradation grade threshold. According to the embodiment of the invention, the method can achieve the precise recognition of the multi-mechanism degradation of the SPD in an online state, and improves the real-time performance and reliability of aging scoring.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of surge protectors, in particular to a surge protector aging degradation detection method and system. BACKGROUND

[0002] Surge protector (SPD) is a key protective device in power systems and electronic equipment for suppressing transient overvoltage. The core components of metal oxide valve (MOV) and electrode structure are prone to aging degradation due to electro-thermal stress in long-term operation, resulting in a decline in protection performance and even failure. Traditional detection methods rely on offline testing or single parameter monitoring (such as leakage current), which is difficult to identify early degradation online and cannot distinguish between multiple degradation mechanisms (such as grain boundary degradation, contact oxidation or thermal damage). In the prior art, composite signal excitation and multi-physical field response analysis have not been effectively combined, resulting in incomplete extraction of degradation characteristics and inadequate adaptability of evaluation models, especially lacking dynamic characterization of the correlation between micro-zone temperature rise and electrical signals. SUMMARY

[0003] The purpose of the present application is to provide a surge protector aging degradation detection method and system to solve the problems in the prior art and achieve accurate identification of multiple mechanism degradation of SPD in online state, improving the real-time and reliability of aging score.

[0004] One embodiment of the present application provides a surge protector aging degradation detection method, which comprises: When the surge protector SPD is in an online running state, a non-destructive composite test signal is injected into its protection line, and real-time multi-physical field response signals of the SPD to the composite test signal are collected, wherein the composite test signal is composed of a triangular wave of a specific frequency range and a step signal of a preset amplitude, and the multi-physical field response signal at least includes a transient voltage response, a transient current response and a SPD surface micro-zone temperature response at a corresponding time point; The collected multi-physical field response signals are subjected to multi-band parallel decomposition to obtain a key frequency band sub-signal set reflecting different degradation mechanism characteristics, wherein the key frequency band sub-signal set at least contains a high-frequency sub-signal reflecting metal oxide valve grain boundary characteristics, a medium-frequency sub-signal reflecting electrode contact degradation and a low-frequency sub-signal reflecting thermal accumulation characteristics; Time domain and frequency domain feature parameters are synchronously extracted from each key frequency band sub-signal, and combined with the micro-zone temperature change rate at the corresponding time point to construct a feature vector representing the multi-dimensional aging state of the SPD, wherein the feature vector contains the energy entropy, waveform distortion factor, specific harmonic component amplitude ratio and correlation coefficient of the temperature change rate of each frequency band; input the feature vector into a pre-trained lightweight deep evaluation model, output a comprehensive aging score reflecting the overall degradation degree of the SPD in real time, wherein the lightweight deep evaluation model is trained based on SPD accelerated aging test data using a transfer learning and attention mechanism fusion algorithm, and can adaptively weight the contribution of different physical field characteristics to degradation; In combination with the comprehensive aging score and a preset degradation level threshold, a specific degradation mode and key degradation site inside the SPD are identified, and a structured detection report containing the degradation level, risk site and maintenance suggestions is generated.

[0005] Optionally, when the surge protector SPD is in an online running state, a non-destructive composite test signal is injected into its protection line, and real-time multi-physical field response signals of the SPD to the composite test signal are collected, wherein the composite test signal is composed of a triangular wave of a specific frequency range and a step signal of a preset amplitude, and the multi-physical field response signal at least includes a transient voltage response, a transient current response and a SPD surface micro-area temperature response at a corresponding time point, including: Based on the SPD online running parameters, a digital signal generator is used to generate a triangular wave sequence of a specific frequency range in real time, and the center frequency of the sequence is dynamically adjusted according to the rated voltage of the SPD to avoid resonance interference, and a frequency adaptive triangular wave signal is output; The step signal of a preset amplitude and the frequency adaptive triangular wave signal are time domain superimposed to form an initial composite test signal, which is injected into the protection line through a current limiter and a voltage clamping circuit to ensure that the signal amplitude is lower than the action threshold of the SPD, and a non-destructive injection signal is output; A high-speed data acquisition card is used to synchronously capture the transient voltage response and the transient current response of the SPD to the non-destructive injection signal, and an infrared thermal imager is used to record the SPD surface micro-area temperature response at a microsecond level sampling rate, so as to ensure that all response signal time stamps are aligned, and an original multi-physical field response data set is output; The original multi-physical field response data set is subjected to time domain interpolation processing to eliminate collection delay, and an adaptive Kalman filter is applied to remove environmental noise, thereby generating a time-synchronized and denoised real-time multi-physical field response signal.

[0006] Optionally, the collected multi-physical field response signal is subjected to multi-band parallel decomposition to obtain a key frequency band sub-signal set reflecting different degradation mechanism characteristics, wherein the key frequency band sub-signal set at least includes a high-frequency sub-signal reflecting metal oxide valve sheet grain boundary characteristics, a medium-frequency sub-signal reflecting electrode contact degradation, and a low-frequency sub-signal reflecting thermal accumulation characteristics, including: The voltage, current and temperature signals in the real-time multi-physical field response signal are subjected to band analysis, the initial frequency band boundary covering high, medium and low frequencies is predefined through fast Fourier transform, and a preliminary frequency band division scheme is output; Based on the preliminary frequency band division scheme, a parallel wavelet packet decomposition algorithm is applied to each physical field signal, a plurality of sub-band signals are extracted, the energy entropy of each sub-band is calculated, candidate sub-bands with energy entropy greater than an energy entropy threshold are screened, and a high-energy sub-band set is output; The high-energy sub-band set is mapped to the SPD degradation physical mechanism, wherein the high-frequency sub-band is associated with the metal oxide valve joint characteristic, the medium-frequency sub-band is associated with the electrode contact degradation, and the low-frequency sub-band is associated with the heat accumulation characteristic. The characteristics are enhanced through mechanism-related filters, and mechanism-marked sub-band signals are output; The mechanism-marked sub-band signals are subjected to cross-physical field frequency band correlation analysis, similar frequency bands are fused through mutual information algorithm, and redundancies are eliminated, thereby generating optimized high-frequency sub-signals, medium-frequency sub-signals and low-frequency sub-signals; The optimized high-frequency sub-signals, medium-frequency sub-signals and low-frequency sub-signals are classified and integrated according to physical mechanisms, thereby forming a key frequency band sub-signal set containing joint characteristics, electrode degradation and heat accumulation characteristics.

[0007] Optionally, time domain and frequency domain feature parameters are synchronously extracted from each key frequency band sub-signal, and a feature vector representing the SPD multi-dimensional aging state is constructed in combination with the micro-area temperature change rate at the corresponding time point, wherein the feature vector contains the energy entropy, waveform distortion factor, specific harmonic component amplitude ratio and correlation coefficient of the temperature change rate of each frequency band, including: For each sub-signal in the key frequency band sub-signal set, a time domain waveform distortion factor is calculated, the factor is obtained by comparing the root mean square error of the actual waveform and the ideal reference waveform, and a time domain distortion factor set of each sub-signal is output; Short-time Fourier transform is applied to the same sub-signal set, and the energy entropy and specific harmonic component amplitude ratio of each frequency band are synchronously extracted, wherein the harmonic component amplitude ratio focuses on the 3rd and 5th harmonics, and a frequency domain feature set of each sub-signal is output; Surface micro-area temperature data at the corresponding time point are extracted from the real-time multi-physical field response signal, the temperature change rate is calculated through first-order differentiation, and the temperature change rate is aligned to the time sequence of each sub-signal through interpolation, and a synchronous temperature change rate vector is output; The time domain distortion factor set, the frequency domain feature set and the synchronous temperature change rate vector are input into a correlation engine, the Pearson correlation coefficient of each feature parameter and the temperature change rate is calculated, the energy entropy and the waveform distortion factor are fused, and a coupled feature matrix is output; The coupling feature matrix is normalized and organized as a structured vector in the frequency band dimension, and the vector elements include the energy entropy, waveform distortion factor, specific harmonic component amplitude ratio, and correlation coefficient with temperature change rate of each frequency band, to generate a feature vector representing the SPD aging state.

[0008] Optionally, the feature vector is input into a pre-trained lightweight deep evaluation model, and a comprehensive aging score reflecting the overall degradation degree of the SPD is output in real time, wherein the lightweight deep evaluation model is trained based on SPD accelerated aging test data, using a transfer learning and attention mechanism fusion algorithm, and can adaptively weight the contribution of different physical field features to degradation, including: Load the pre-trained lightweight deep evaluation model from the embedded storage system, which is trained based on SPD accelerated aging test data, uses transfer learning to initialize parameters, and has a compressed convolutional neural network architecture, and output a model instance; Before inputting the feature vector into the model instance, apply a dynamic standardization module based on Z-score to adjust the feature scale according to the real-time SPD operating environment, ensure input compatibility with the model, and obtain a standardized feature vector; After inputting the standardized feature vector into the model, activate the built-in attention mechanism fusion algorithm, which adaptively weights different physical field features, and outputs the original degradation score through a fully connected layer; Apply a sigmoid function to map the original degradation score to the range of 0-100, and combine it with online running parameters for real-time calibration to generate a comprehensive aging score.

[0009] Optionally, the comprehensive aging score and the pre-set degradation level threshold are combined to identify the specific degradation mode and key degradation site inside the SPD, and a structured detection report containing the degradation level, risk site, and maintenance suggestion is generated, including: Compare the comprehensive aging score with the pre-set degradation level threshold, which sets multiple levels based on historical failure data, and output the specific degradation level label; Based on the degradation level label and the feature vector, map to the internal degradation mode through a decision tree rule engine, and output the degradation mode description; Combine the degradation mode description and the key frequency band sub-signal set, and use the back propagation algorithm to locate the risk site inside the SPD, and output the risk site coordinates and type; Input the degradation level label, degradation mode description, and risk site into the report template engine to automatically generate a structured detection report containing the degradation level, risk site, and maintenance suggestion.

[0010] Another embodiment of the present application provides a surge protector aging degradation detection system, the system comprising: The collection module is configured to inject a non-destructive composite test signal into a protection line of the surge protector (SPD) when the SPD is in an online running state, and collect a real-time multi-physical field response signal of the SPD to the composite test signal, wherein the composite test signal is composed of a triangular wave of a specific frequency range and a step signal of a preset amplitude, and the multi-physical field response signal at least includes a transient voltage response, a transient current response, and a SPD surface micro-area temperature response at a corresponding time point; The decomposition module is configured to perform multi-band parallel decomposition on the collected multi-physical field response signal to obtain a key frequency band sub-signal set reflecting characteristics of different degradation mechanisms, wherein the key frequency band sub-signal set at least includes a high-frequency sub-signal reflecting characteristics of metal oxide valve plate grain boundaries, a medium-frequency sub-signal reflecting electrode contact degradation, and a low-frequency sub-signal reflecting thermal accumulation characteristics. The extraction module is configured to synchronously extract time domain and frequency domain feature parameters from each key frequency band sub-signal, and construct a feature vector representing a multi-dimensional aging state of the SPD by combining a micro-area temperature change rate at a corresponding time point, wherein the feature vector includes energy entropy, waveform distortion factor, specific harmonic component amplitude ratio, and a correlation coefficient with the temperature change rate of each frequency band. The evaluation module is configured to input the feature vector into a pre-trained lightweight deep evaluation model to output a comprehensive aging score reflecting an overall degradation degree of the SPD in real time, wherein the lightweight deep evaluation model is trained based on SPD accelerated aging test data by using a transfer learning and attention mechanism fusion algorithm, and can adaptively weight the contribution of different physical field characteristics to degradation. The generation module is configured to identify a specific degradation mode and a key degradation position inside the SPD by combining the comprehensive aging score and a preset degradation level threshold, and generate a structured detection report including a degradation level, a risk position, and a maintenance suggestion.

[0011] Still another embodiment of the present application provides a storage medium having a computer program stored therein, wherein the computer program is configured to execute the method described in any one of the above embodiments when running.

[0012] Still another embodiment of the present application provides an electronic device comprising a memory and a processor, wherein the memory has a computer program stored therein, and the processor is configured to execute the computer program to perform the method described in any one of the above embodiments.

[0013] Compared with the prior art, the surge protector aging degradation detection method provided by the application injects a non-destructive composite test signal into the protection line of the surge protector SPD when the SPD is in an online running state, and collects real-time multi-physical field response signals of the SPD to the composite test signal; the collected multi-physical field response signals are subjected to multi-band parallel decomposition to obtain a key frequency band sub-signal set reflecting different degradation mechanism characteristics; time domain and frequency domain characteristic parameters are synchronously extracted from each key frequency band sub-signal to construct a characteristic vector representing the multi-dimensional aging state of the SPD; the characteristic vector is input into a pre-trained lightweight deep evaluation model to output a comprehensive aging score reflecting the overall degradation degree of the SPD in real time; and a structured detection report is generated by combining the comprehensive aging score and a preset degradation level threshold, so that the multi-mechanism degradation of the SPD in the online state can be accurately identified, and the real-time performance and reliability of the aging score are improved. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 A hardware structure block diagram of a computer terminal of a surge protector aging degradation detection method provided by an embodiment of the application is shown in the figure. Figure 2 A flowchart of a surge protector aging degradation detection method provided by an embodiment of the application is shown in the figure. Figure 3 A structural diagram of a surge protector aging degradation detection system provided by an embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0015] The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the application, and cannot be explained as a limitation on the application.

[0016] An embodiment of the application first provides a surge protector aging degradation detection method, which can be applied to electronic equipment such as a computer terminal, specifically, a general computer, etc.

[0017] The following will be described in detail by taking a computer terminal as an example. Figure 1 A hardware structure block diagram of a computer terminal of a surge protector aging degradation detection method provided by an embodiment of the application is shown in the figure. Figure 1 As shown in the figure, the computer device includes a processor, a memory and a network interface connected through a system bus, wherein the memory can include a non-volatile storage medium and an internal memory.

[0018] The non-volatile storage medium can store an operating system and a computer program. The computer program includes program instructions, which, when executed, can make the processor execute any surge protector aging degradation detection method.

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

[0020] The internal memory provides an environment for the running of a computer program in a non-volatile storage medium, which, when executed by the processor, can cause the processor to execute any surge protector aging degradation detection method.

[0021] The network interface is configured to perform network communication, such as sending assigned tasks, etc. Figure 1 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

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

[0023] Referring to Figure 2 The embodiments of the present application provide a surge protector aging degradation detection method, which can include the following steps: S201, when the surge protector SPD is in an online running state, a non-destructive composite test signal is injected into the protection line thereof, and a real-time multi-physical field response signal of the SPD to the composite test signal is collected, wherein the composite test signal is composed of a triangular wave of a specific frequency range and a step signal of a preset amplitude, and the multi-physical field response signal at least includes a transient voltage response, a transient current response and a SPD surface micro-area temperature response at a corresponding time point; Specifically, a triangular wave sequence of a specific frequency range can be generated in real time based on the SPD online running parameters through a digital signal generator, and the center frequency of the sequence is dynamically adjusted according to the SPD rated voltage to avoid resonance interference, and a self-adaptive triangular wave signal is outputted; Running parameter analysis and frequency decision The system first acquires the online running parameters of SPD (Surge Protection Device) through smart meters in real time, including but not limited to line rated voltage (such as 220VAC or 380VAC), load current fluctuation range (such as 0-100A), and historical surge event statistics. These parameters are input into the Central Frequency Dynamic Calculation Module (CFDCM). The module has a frequency mapping algorithm built-in, which automatically calculates the center frequency (CF) according to the rated voltage value. For example, for a 220VAC system, the initial value of CF is set to 5kHz; for a 380VAC system, CF is raised to 8kHz. This design is based on the principle of electromagnetic compatibility (EMC): higher voltage systems need to avoid lower frequency band power grid harmonic interference bands (such as 2-4kHz power frequency harmonic region). The calculation process introduces a voltage-frequency conversion coefficient (VFCC), which is obtained through laboratory calibration, ensuring that the frequency adjustment step (such as 200Hz frequency shift corresponding to every 10V voltage change) accurately matches the electrical characteristics of different SPD models, thereby effectively avoiding resonance interference (Resonance Interference) that may be caused by the internal LC circuit of SPD.

[0024] Real-time generation of triangular wave sequence Digital Signal Generator (DSG) adopts Direct Digital Synthesis (DDS) technology, receives the target center frequency instruction output by CFDCM. Taking CF=5kHz as an example, DSG generates a triangular wave sequence in a specific frequency range (SFR): the lower bound frequency (LBF) is CFx0.8, i.e. 4kHz, and the upper bound frequency (UBF) is CFx1.2, i.e. 6kHz. The slope symmetry (SS) of the triangular wave is set to 50% (the rising edge time is equal to the falling edge time), and the amplitude is temporarily set to the intermediate value 50% Vpp (peak-to-peak voltage). During waveform generation, the phase accumulator (PA) of DSG is driven by a 100MHz clock to ensure that the frequency resolution reaches 0.1Hz. At the same time, the dynamic frequency fine-tuning unit (DFFTU) continuously monitors the line background noise spectrum (through the parallel FFT analyzer), and if a noise peak is detected near 4.5kHz, the CF is automatically shifted to 5.2kHz to realize frequency adaptation.

[0025] Anti-interference optimization and signal output To suppress the influence of high-frequency switching noise on the purity of the triangular wave, a 2nd-Order Butterworth Low-Pass Filter (2B-LPF) is connected in series to the output stage of DSG, and the cutoff frequency is set to 1.5 times the UBF (9kHz in this example). The filtered signal is buffered and isolated by a voltage follower (VF), and finally a frequency-adaptive triangular wave signal (FATWS) is output. The signal characteristics are: voltage amplitude range 0-10Vpp (programmable), frequency error <±0.5%, total harmonic distortion (THD) <1.5%, meeting the accuracy requirements of subsequent signal superposition.

[0026] The step signal with a preset amplitude is time-domain superimposed with the frequency-adaptive triangular wave signal to form an initial composite test signal, and is injected into the protection line through a current limiter and a voltage clamping circuit to ensure that the signal amplitude is lower than the SPD action threshold, and a non-destructive injection signal is output. Step signal parameter configuration and time domain superposition The preset amplitude step signal (PASS) is generated by a dedicated step generator. Its amplitude is dynamically set according to the operation threshold (OT) of the SPD: for voltage-limiting SPD (such as MOV type), OT is usually 1.2 times the maximum continuous operating voltage (MCOV). Taking a 220VAC system as an example, MCOV=275V, then OT=330V, and the PASS amplitude is set to 10% of OT, i.e. 33V (ensuring far below the action point). The step rise time (RT) is fixed at 100ns to cover the high-frequency response characteristics. The superposition process is completed in the time-domain signal synthesizer (TDSS): a high-speed operational amplifier is used inside the TDSS to perform analog addition operation on the FATWS and the PASS. The superposition ratio is adjusted by a precision resistance network, and the typical ratio is 70% for the triangular wave and 30% for the step signal, forming the initial composite test signal (ICTS), whose time-domain characteristics are characterized by a transient jump edge superimposed on the triangular wave slope.

[0027] Non-destructive safety circuit design To ensure the absolute safety of the injected signal, the ICTS first passes through the current limiter (CL). The CL adopts a two-stage protection architecture: The first stage is a self-resetting fuse (polymer positive temperature coefficient device, PPTC), with an action current of 5mA (far below the SPD leakage current alarm value); The second stage is a constant current source circuit (CCS), with a maximum output current of 1mA (a typical safety value).

[0028] The signal then enters the voltage clamping circuit (VCC), which is composed of back-to-back Zener diodes (such as BZT52C33S, clamping voltage 33V) and transient voltage suppressors (TVS, model SMBJ36CA, clamping voltage 36V) in parallel, double protection to limit accidental overvoltage below 40V. According to the test, the equivalent output impedance of the injection end is >1MΩ, and the load effect on the protection circuit can be ignored.

[0029] Signal injection and online verification The processed ICTS is coupled to the SPD parallel protection circuit through a high-frequency isolation transformer (HFIT, transformation ratio 1:1, bandwidth DC-100kHz). The injection point is located at the front end of the SPD terminal, using a non-invasive clamp probe (such as Pomona 6992). The non-destructive verification module (NDVM) monitors the voltage and current of the injection point in real time: if the instantaneous current >0.8mA or the voltage >OTx8% (i.e. 26.4V), the soft shutdown (through MOSFET switch) is triggered immediately, and the alarm log is generated. Under normal conditions, the non-destructive injection signal (NDIS) is output, with typical parameters: voltage peak ≤30V, current peak ≤0.5mA, duration 500ms, and automatic test performed every 15 minutes.

[0030] The high-speed data acquisition card is used to synchronously capture the transient voltage response and transient current response of the SPD to the non-destructive injection signal, and the infrared thermal imager is used to record the SPD surface micro-area temperature response with a microsecond-level sampling rate, ensuring that all response signal timestamps are aligned, and outputting the original multi-physical field response data set; Multi-channel high-speed synchronous acquisition architecture The response signal acquisition system is based on the PXIe platform (PCI eXtensions for Instrumentation Express). The core device is a high-speed data acquisition card (HSDAC, model such as NI PXIe-5162), which is configured with four-channel synchronous sampling: Channel 1 (voltage response): connected to the two ends of the SPD through a high-voltage differential probe (such as Tektronix THDP0200, bandwidth 200MHz), range ±50V, resolution 16bit; Channel 2 (Current Response): A Rogowski Coil (e.g., PEARSON 110A, sensitivity 0.1 V / A) is wrapped around the SPD ground wire, with a range of ±2 A. Channel 3 (Trigger Signal): The NDIS synchronization output is directly collected as the time reference.

[0031] The sampling rate of the acquisition card is set to 10 MS / s (million samples per second), with a storage depth of 256 MB per channel. The trigger mode is Rising Edge Trigger, and the threshold is set to 20% of the NDIS amplitude.

[0032] Microzone Temperature Synchronous Imaging Technology An infrared thermal imager (IRTI, e.g., FLIR X8580SC) is aimed at the surface of the SPD housing, focusing on the area corresponding to the MOV valve plate (about 10×10 mm microzone). To achieve microsecond-level sampling: High-Speed Mode is enabled, with a maximum frame rate of 5000 fps (0.2 ms / frame); The Region of Interest (ROI) is reduced to 32×32 pixels to improve data transmission rate; The built-in Time Synchronization Unit (TSU) receives the trigger signal from the HSDAC and aligns the clock through the IRIG-B (Inter-Range Instrumentation Group-B) time code, with a timestamp accuracy of ±100 ns. The temperature data is output in 14-bit digital format, with a sensitivity of 0.02℃.

[0033] Data Integration and Time Alignment After the acquisition is started, the HSDAC and IRTI capture response data for 500 ms simultaneously. The Signal Alignment Processor (SAP) performs three steps of processing: Coarse Alignment: Based on the rising edge of the trigger signal, the -5 ms to +495 ms data segment of all channels is intercepted; Fine Alignment: The Cross-Correlation (CC) between the voltage response signal and the NDIS reference signal is calculated, with a time shift compensation accuracy of 10 ns; Temperature mapping: interpolate (cubic spline) each frame of IRTI temperature data to HSDAC's 10 MS / s time axis, generating continuous temperature curves. The final output is a Raw Multi-Physics Response Dataset (RMPRD) containing three strictly synchronized time-domain sequences: voltage (unit V), current (unit A), temperature (unit °C), with a total data size of about 150 MB per test.

[0034] Time-domain interpolation is applied to the Raw Multi-Physics Response Dataset to eliminate acquisition delays and an Adaptive Kalman Filter is applied to remove environmental noise, generating time-synchronized and denoised real-time multi-physical field response signals.

[0035] Time-domain interpolation eliminates channel delays Although the hardware is synchronized, there are still differences in the physical delays of each sensor: the voltage probe delay is about 15 ns, the current probe delay is 50 ns, and the thermal imager imaging delay is 200 μs. The Time-Domain Interpolation Engine (TDIE) uses the Piecewise Cubic Hermite Interpolating Polynomial (PCHIP) algorithm: Take the voltage channel as the reference (with the smallest delay), and keep the time axis unchanged; The current signal time axis is shifted forward by 35 ns in total (compensating for the delay difference with the voltage); The temperature signal, due to its large delay, is reconstructed as a continuous curve on the 10 MS / s sampling points based on the original data points spaced 200 μs apart. After interpolation, the deviation of all signals on the time axis is <1 ns, meeting the requirements of strict time synchronization.

[0036] Adaptive Kalman Filter denoising Main sources of environmental noise: Electromagnetic noise (such as switching power supply harmonics): affects voltage / current signals; Air convection disturbance: affects temperature measurement.

[0037] The Adaptive Kalman Filter (AKF) is designed for three-channel parallel processing: State equation: assume that the signal change rate is constant (x k =x k−1 +v k ); Observation equation: direct measurement value (z k =x k +wk Parameter adaptation: dynamically adjust the process noise covariance matrix Q and the measurement noise covariance matrix R according to the local variance of the signal. For example, when the current signal has a sudden jump in variance, Q is automatically expanded by 10 times to track the rapid change; when it is stable, Q is reduced to improve the smoothness. The signal-to-noise ratio (SNR) of the filtered signal is improved by >20 dB.

[0038] Quality check and signal output The denoised data is subjected to signal integrity verification (SIV): Check if the phase difference of the voltage / current is within the MOV theoretical tolerance (±5°); Verify if the temperature change rate conforms to the heat conduction model (e.g., |dT / dt|<10℃ / s). The data that passes the verification is marked as time-synchronized and denoised real-time multi-physics response signals (TSDRT-MPRS) and stored as a structured array: [timestamp, voltage (V), current (A), temperature (℃)]. The timestamp accuracy is 0.1μs, and the data can be used for subsequent frequency band decomposition and feature extraction.

[0039] By injecting a composite test signal (triangle wave and step signal superimposed) into the running SPD, the voltage, current and micro-zone temperature response are synchronously collected without affecting the normal operation of the device. The composite signal design takes into account the excitation requirements of different degradation mechanisms, with the triangle wave covering the wide frequency characteristics, and the step signal used to detect the transient response capability. Multi-physics synchronous acquisition provides a comprehensive data basis for subsequent analysis, realizes non-invasive data acquisition for online detection of SPD, and avoids the limitations of traditional detection which must be powered off. The multi-physics response signal can comprehensively reflect the electrical performance and thermal characteristics of the SPD, providing multi-dimensional data support for accurate evaluation of the degradation state.

[0040] S202, the collected multi-physics response signals are subjected to multi-frequency band parallel decomposition to obtain a key frequency band sub-signal set reflecting the characteristics of different degradation mechanisms, wherein the key frequency band sub-signal set at least contains a high-frequency sub-signal reflecting the metal oxide valve sheet grain boundary characteristics, a medium-frequency sub-signal reflecting the electrode contact degradation, and a low-frequency sub-signal reflecting the thermal accumulation characteristics; ​Specifically, the voltage, current and temperature signals in the real-time multi-physical field response signal can be subjected to band analysis respectively, the initial frequency band boundary covering high frequency, medium frequency and low frequency is predefined through fast Fourier transform, and a preliminary frequency band division scheme is output. Band analysis pre-processing The system first performs signal integrity checking on the collected real-time multi-physical field response signal. The voltage signal (unit: volt V) is collected by a high-voltage differential probe, and the range covers ±6 kilovolts (kV); the current signal (unit: ampere A) is captured by a Rogowski coil sensor, and the bandwidth is 100 kilohertz (kHz); the temperature signal (unit: degree Celsius ℃) is obtained by an infrared thermal imager at a sampling rate of 500 thousand frames per second (500 kfps). The three signals are aligned in time stamp through a time synchronization module, with an error of less than 1 microsecond (μs). The signal pre-processing module performs baseline correction on the original data, eliminates direct current bias (such as power frequency 50 Hz interference), and applies a Hanning Window function to reduce spectral leakage. The pre-processed signal is split into independent voltage, current and temperature data streams, which are respectively input into the band analysis channel.

[0041] Fast Fourier transform frequency band division Each physical field signal channel is built-in with a Fast Fourier Transform (FFT) processor. Taking the voltage signal as an example: the FFT processor converts the time-domain voltage waveform into a frequency-domain energy spectrum, and the analysis range is set to 10 hertz (Hz) to 10 megahertz (MHz). Three significant energy concentration areas are identified in the energy spectrum: Low frequency area (10 Hz-1 kHz): corresponding to the thermal time constant of the heat accumulation process; Medium frequency area (1 kHz-100 kHz): reflecting the mechanical contact vibration frequency of the electrode; High frequency area (100 kHz-10 MHz): matching the grain boundary dielectric relaxation frequency of the metal oxide valve plate.

[0042] The system automatically divides the initial boundary according to the energy peak position: low frequency boundary LF_bound=1 kHz, high frequency boundary HF_bound=100 kHz. The FFT analysis upper limit of the temperature signal is set to 1 kilohertz (kHz) due to its slow response, and only two sub-bands of low frequency (0.01-1 Hz) and ultra-low frequency (<0.01 Hz) are divided. The final output includes the preliminary frequency band division scheme of the voltage, current and temperature three-channel frequency band boundary parameters (such as voltage signal: low frequency 10 Hz-1 kHz, medium frequency 1-100 kHz, high frequency 100 kHz-10 MHz).

[0043] Dynamic boundary calibration To prevent feature omission caused by fixed boundaries, the system introduces an energy accumulation threshold (EAT = 85%) for secondary calibration. Calculate the energy proportion of each frequency band, if the energy of a certain frequency band exceeds EAT, expand its boundary until the energy proportion drops to 80%. For example, when an abnormal energy spike occurs at 120 kHz in the current signal, the high-frequency boundary is expanded from 100 kHz to 150 kHz. The calibrated frequency band boundaries are stored in association with the physical field type to form the final version of the preliminary frequency band division scheme.

[0044] Based on the preliminary frequency band division scheme, apply the parallel wavelet packet decomposition algorithm to each physical field signal, extract multiple sub-band signals, and calculate the energy entropy of each sub-band. Screen candidate sub-bands with energy entropy greater than the energy entropy threshold, and output the high-energy sub-band set; Wavelet packet decomposition engine architecture The system deploys a parallel wavelet packet decomposition (PWPD) engine, which includes three independent processing units corresponding to voltage, current, and temperature signals. Each unit uses a Daubechies 4 (db4) wavelet basis for 5-layer decomposition, dividing the original signal into 32 sub-bands. For example, after 5-layer decomposition of the voltage signal (bandwidth 10 MHz), the sub-band width is 312.5 kHz. The decomposition process uses a zero-phase filter algorithm to avoid signal distortion, and implements streaming processing through double-buffer memory technology with a delay of less than 5 ms.

[0045] Energy entropy dynamic calculation and screening Calculate the energy entropy (EE) of each sub-band signal, which represents the uncertainty of sub-band energy in total energy distribution. The calculation formula is: EE = -Σ(P_i * log2(P_i)) (where P_i is the energy proportion of the i-th sub-band).

[0046] The system sets the energy entropy threshold (EET) to 0.75. When the EE of a sub-band is greater than EET, it indicates that the frequency band carries significant aging feature information. Taking the current signal as an example: under electrode degradation state, the EE value of the 50-70 kHz sub-band (corresponding to the contact point vibration frequency) can reach 0.92, far exceeding EET. The system scans all sub-bands in real time, screens out candidate sub-bands with EE > 0.75, and records their center frequencies (e.g. voltage signal: 125 kHz, 2.3 MHz; current signal: 52 kHz, 880 kHz; temperature signal: 0.05 Hz).

[0047] Cross-physical field band association To improve the reliability of features, the system performs cross-physical field band verification: if a band is simultaneously filtered as a candidate in at least two physical fields (such as the current 52 kHz and voltage 55 kHz bands), a band merging mechanism is triggered to generate a joint candidate band (JCB). The final output contains a high-energy sub-band set including independent candidate bands and joint candidate bands, and marks the physical field and center frequency to which it belongs.

[0048] Map the high-energy sub-band set to the SPD degradation physical mechanism, where high-frequency sub-bands are associated with metal oxide valve sheet grain boundary characteristics, medium-frequency sub-bands are associated with electrode contact degradation, and low-frequency sub-bands are associated with thermal accumulation characteristics. Through mechanism-related filters, enhance features and output mechanism-labeled sub-band signals. Degradation mechanism mapping rule library The system has a built-in degradation mechanism mapping rule library (DMMR), which includes three types of core mapping: High-frequency mapping (> 100 kHz): associated with metal oxide valve sheet grain boundary characteristics. For example, the 125 kHz band corresponds to grain boundary capacitance changes, and bands above 2 MHz reflect grain boundary ion migration; Medium-frequency mapping (1-100 kHz): associated with electrode contact degradation. For example, bands near 50 kHz indicate electrode spring fatigue, and 80 kHz bands correspond to weld point cracks; Low-frequency mapping (< 1 kHz): associated with thermal accumulation characteristics. The 0.1 Hz band reflects heat sink aging, and the 0.01 Hz band indicates moisture accumulation due to seal failure.

[0049] The mapping process uses a band-mechanism matching score (BMMS). When the center frequency of the band and the typical frequency of the mechanism differ by less than 20% of the bandwidth, BMMS = 1 (complete match).

[0050] Mechanism-related filter design For each successfully mapped sub-band, load a dedicated mechanism-related filter (MCF): High-frequency filter: designed as a bandpass + differential enhancement combination to highlight grain boundary nonlinear response (such as the steep change in the resonance peak of the 2 MHz band); Medium-frequency filter: uses a comb filter to suppress power frequency harmonics and enhance electrode vibration features (such as the impact oscillation envelope of the 52 kHz band); Low-frequency filter: Adaptive Kalman smoothing is applied to separate ambient temperature fluctuations (such as the thermal inertia characteristics of the 0.05 Hz band).

[0051] The signal-to-noise ratio (SNR) of the filtered signal is improved by at least 15 dB, for example, the SNR of a current signal in the 52 kHz band is improved from 8 dB to 23 dB.

[0052] Mechanism labeling and feature enhancement The filtered sub-band signal is assigned a mechanism marker code: Grain boundary characteristic marking: Gxxx (xxx is the frequency value, such as G125k indicating a grain boundary characteristic at 125 kHz); Electrode degradation marking: Exxx (e.g., E52k indicates a 52 kHz electrode failure); Thermal accumulation marking: Txxx (e.g., T0.05 indicates thermal characteristics at 0.05 Hz).

[0053] Finally, a mechanism-marked subband signal with time-frequency domain enhancement features is generated and stored as a triplet of {mark code, filtered signal, center frequency}.

[0054] Cross-physical field frequency band correlation analysis is performed on the mechanism-marked sub-band signal. Similar frequency bands are fused through mutual information algorithm to eliminate redundancy and generate optimized high-frequency sub-signals, mid-frequency sub-signals and low-frequency sub-signals. Mutual information correlation analysis The system uses the Mutual Information (MI) algorithm to calculate the cross-physical field frequency band correlation. Taking the high-frequency region as an example: the MI value of the voltage signal G125k and the current signal G130k is calculated as follows: MI(U_G125k, I_G130k) = ΣΣ p(u,i) log2(p(u,i) / (p(u)p(i))).

[0055] Where p(u,i) is the joint probability distribution. If MI > mutual information threshold (MIT=0.65), the two frequency bands are determined to reflect the same physical mechanism (e.g., both originating from grain boundary defects). The analysis covers three combinations: Voltage and current are in the same frequency band (e.g., U_G125k vs I_G130k); Current-temperature cross-frequency band (e.g., I_E52k vs T_1.2); Voltage-temperature cross-mechanism (e.g., U_G2M vs T_0.01).

[0056] Frequency band fusion and redundancy elimination Perform weighted fusion on frequency bands with high correlation (MI > 0.65): Intra-mechanism fusion: e.g. U_G125k and I_G130k are fused to optimize high-frequency signals, with weight ratio assigned by signal-to-noise ratio (SNR_U:SNR_I=3:2); Cross-mechanism association: e.g. I_E52k and T_0.5 (electrode heating) are fused to generate optimized mid-frequency signals, preserving phase difference characteristics; Low-correlation (MI < 0.3) frequency bands are discarded directly, e.g. T_10Hz (ambient noise) in temperature signals. The number of signals after fusion is reduced by more than 50% (e.g. from 32 sub-bands to 15).

[0057] Generating optimized sub-signals The fusion results are reorganized into three categories by mechanism type: Optimized high-frequency sub-signals: containing only grain boundary characteristic-related frequency bands (e.g. G125k_opt, G2M_opt after fusion); Optimized mid-frequency sub-signals: integrating electrode degradation and associated thermal characteristics (e.g. E52k_opt, E80k_opt); Optimized low-frequency sub-signals: aggregating pure thermal accumulation characteristics (e.g. T0.05_opt, T0.01_opt).

[0058] Each category of sub-signals is attached with a Band Contribution Index (BCI), e.g. BCI of G125k_opt = 0.92 (indicating that this frequency band carries 92% of the grain boundary degradation information).

[0059] The optimized high-frequency, mid-frequency, and low-frequency sub-signals are classified and integrated according to physical mechanisms, forming a key frequency band sub-signal set containing grain boundary characteristics, electrode degradation, and thermal accumulation characteristics.

[0060] Mechanism characteristic integration The system creates a 3D Feature Container, organizing optimized sub-signals in the following dimensions: X-axis: degradation mechanism (grain boundary characteristics / electrode degradation / thermal accumulation); Y-axis: physical field source (voltage / current / temperature); Z-axis: frequency band level (basic band / fused band).

[0061] For example, the optimized high-frequency sub-signal G125k_opt is stored in the container location: X=grain boundary characteristics, Y=voltage, Z=basic band.

[0062] Temporal and spatial feature alignment Four-dimensional alignment (time-frequency-mechanism-physical field) is performed on the integrated signals: Time alignment: resample to uniform time base (10 ns resolution); Frequency alignment: normalize to standard bandwidth (high frequency 100 kHz / bin, medium frequency 10 kHz / bin, low frequency 1 Hz / bin); Mechanism alignment: focus on rising edge response of grain boundary characteristics, pay attention to oscillation decay of electrode degradation, and extract slope features of thermal accumulation; Physical field alignment: voltage-current signals are associated through impedance phase mapping, and temperature signals are bound with power loss integration.

[0063] Generate final set Output Key Band Sub-signal Set (KBSS), for example, the data structure contains: { "Grain boundary characteristics": [G125k_opt, G2M_opt,...], / / High frequency sub-signal set "Electrode degradation": [E52k_opt, E80k_opt,...], / / Medium frequency sub-signal set "Thermal accumulation": [T0.05_opt, T0.01_opt,...] / / Low frequency sub-signal set } Each sub-signal is attached with metadata: center frequency (e.g. 125.3 kHz), bandwidth (±12.5 kHz), signal-to-noise ratio (e.g. 23.7 dB), and mechanism matching degree (e.g. 0.94). This set is directly input into the subsequent feature vector construction module.

[0064] Using frequency band parallel decomposition technology, complex response signals are decoupled into sub-signals of different frequency bands according to physical mechanisms. High frequency band corresponds to changes in valve piece grain boundary microstructure, medium frequency band reflects electrode contact interface state, and low frequency band embodies thermal accumulation effect. This decomposition method realizes independent extraction of degradation characteristics of multiple parts inside SPD, breaks through the limitations of traditional single signal analysis, and accurately locates the characteristic signals corresponding to different degradation modes through frequency band division guided by physical mechanisms, laying a foundation for subsequent targeted evaluation.

[0065] S203, synchronously extract time domain and frequency domain feature parameters from each key frequency band sub-signal, and combine the micro-area temperature change rate at the corresponding time point to construct a feature vector representing the multi-dimensional aging state of SPD, wherein the feature vector contains energy entropy, waveform distortion factor, specific harmonic component amplitude ratio, and correlation coefficient with temperature change rate of each frequency band; Specifically, for each sub-signal in the key frequency band sub-signal set, a time-domain waveform distortion factor is calculated, which is derived by comparing the root mean square error of the actual waveform and the ideal reference waveform, and a set of time-domain distortion factors for each sub-signal is output. Definition and calculation principle of time-domain waveform distortion factor Waveform distortion factor (WDF) is a core indicator for quantifying the deviation of a signal from an ideal state. For high-frequency sub-signals (reflecting metal oxide valve sheet grain boundary characteristics), medium-frequency sub-signals (reflecting electrode contact degradation), and low-frequency sub-signals (reflecting thermal accumulation characteristics), the system establishes a corresponding ideal reference waveform library. This library is generated based on the standard response curve of the SPD in a new state: for example, the ideal waveform in the high-frequency band is a smooth decaying oscillation wave, the medium-frequency band is a linear step response, and the low-frequency band is a constant slope rising wave. During calculation, the actual collected sub-signal (such as the high-frequency current response of a certain detection) is first time-aligned with the ideal waveform of the same frequency band, and the dynamic time warping (DTW) algorithm is used to eliminate micro-time sequence deviations. After alignment, the signal is segmented into time windows of milliseconds (e.g., each 10 milliseconds is a window), and the root mean square error (RMSE) between the actual waveform and the ideal waveform is calculated for each window. The RMSE value reflects the overall deviation of the waveform in the window, and its physical meaning is the square root of the average of the squares of the distances between the actual data points and the ideal curve. Finally, the distortion factor of the window is defined as the normalized RMSE value, i.e., the ratio of the current RMSE to the RMSE reference value under the same conditions of the new SPD.

[0066] Multi-scale distortion feature extraction and anti-interference processing To distinguish between transient interference and real degradation, the system uses a multi-scale sliding window strategy: RMSE is calculated simultaneously in three levels of windows of 1 millisecond (to capture transient distortion), 10 milliseconds (to analyze local anomalies), and 100 milliseconds (to evaluate overall trends). Each sub-signal (such as the medium-frequency voltage signal related to electrode contact degradation) generates three sets of distortion factor sequences. Subsequently, outliers are removed through an outlier suppression algorithm: if the distortion factor of a window suddenly increases to more than 3 times the average value of adjacent windows (threshold adjustable), and no external electromagnetic interference event is detected in that window (verified by an independent sensor), it is determined to be an effective degradation feature; otherwise, it is considered as noise and smoothed. Finally, a representative distortion factor is output for each scale window - the 90th percentile (P90) of all window distortion factors in that scale is taken to avoid the influence of accidental fluctuations. Thus, a single sub-signal produces three time-domain distortion factors (corresponding to 1ms / 10ms / 100ms scales), and the factor set of all sub-signals (voltage, current, and temperature signals in high, medium, and low frequency bands) constitutes the time-domain distortion factor set.

[0067] Dynamic reference calibration and output verification The ideal reference waveform is not static, but dynamically adjusted according to the SPD online working conditions (such as ambient temperature, load current). The system has a built-in adaptive reference engine: when the SPD comprehensive aging score is lower than 5% (considered healthy), the actual waveform of this detection is automatically integrated into the historical reference library with a weight of 0.1, allowing the ideal waveform to slowly track the natural aging of the device. Before output, the set of time domain distortion factors need to pass the cross-signal consistency test: for example, the distortion factors of high-frequency current and voltage signals should meet the preset proportional relationship (such as current distortion is usually greater than voltage). If a signal distortion factor deviates from this relationship abnormally (deviation exceeds 20%), the resampling process is triggered. The set of verified factors is stored in a three-dimensional structure of "frequency band-physical quantity-time scale" for subsequent feature fusion calls.

[0068] Apply short-time Fourier transform to the same sub-signal set, synchronously extract the energy entropy and specific harmonic component amplitude ratio of each frequency band, where the harmonic component amplitude ratio focuses on the 3rd and 5th harmonics, and outputs the frequency domain feature set of each sub-signal; Engineering implementation of short-time Fourier transform Short-Time Fourier Transform (STFT) is used to analyze the frequency domain characteristics of signals while preserving time domain information. The system sets a Hanning Window with a 50% overlap rate, and the window length is dynamically selected according to the frequency band of the sub-signal: 0.5 milliseconds for high frequency band (to analyze MHz-level components), 2 milliseconds for medium frequency band (to cover hundreds of kHz), and 10 milliseconds for low frequency band (to focus on kHz below). Within each time window, perform Fast Fourier Transform (FFT) on the sub-signal (such as low-frequency temperature signal reflecting heat accumulation), generating a time-frequency matrix - rows represent frequency points, columns represent time windows, and element values are complex spectra. To reduce computational load, only key frequency bands are retained: 1-10 MHz for high frequency band, 10 kHz-1 MHz for medium frequency band, and 0.1-10 kHz for low frequency band, and the energy of the remaining frequency bands is set to zero.

[0069] Extraction of physical meaning of energy entropy and harmonic amplitude ratio Energy Entropy (EE) measures the disorder degree of spectral energy distribution, indirectly reflecting the signal disorder caused by degradation. When calculating, first normalize the spectral amplitude in a single time window to a probability distribution, and then substitute it into the Shannon entropy formula. For example, when the electrode contact is degraded, the energy of the medium frequency current signal will leak from the fundamental frequency to the higher harmonics, and the entropy value will increase. The specific harmonic component amplitude ratio focuses on the 3rd harmonic (H3) and the 5th harmonic (H5), as they are sensitive to SPD degradation: H3 amplitude increases when metal oxide grain boundary is damaged, and H5 is significantly amplified when the electrode is sulfided. The specific calculation is: Fundamental amplitude (F1): take the maximum amplitude within the range of ±5% of the center frequency of the sub-signal; H3 / H5 amplitude: extract the amplitude at 3 times / 5 times the center frequency; Amplitude ratio is defined as: H3 / F1 and H5 / F1; Each time window outputs a set of EE, H3 / F1, H5 / F1 values, forming a feature sequence that evolves over time.

[0070] Dimensionality reduction and noise resistance optimization of frequency domain features The original time-frequency matrix data is large, and the system extracts effective information through feature compression pipeline: Energy entropy sequence: take the average (EE_Mean) and variance (EE_Var) of all time window EE values, representing the overall and volatility; Harmonic amplitude ratio: calculate the 75th percentile value (P75) of H3 / F1 and H5 / F1 within the entire signal duration, highlighting abnormal peaks; In order to suppress noise interference, use coherent averaging technique: average the frequency domain features of the same sub-signal detected 3 times in a row, only keep the repeated patterns. Finally, each sub-signal outputs 5 frequency domain features: EE_Mean, EE_Var, H3 / F1_P75, H5 / F1_P75, and the interaction term of H3 and H5 (H3xH5 / F1 2 ). The feature set of all sub-signals constitutes the frequency domain feature set.

[0071] Extract the surface micro-region temperature data at the corresponding time point from the real-time multi-physical field response signal, calculate the temperature change rate by first-order differentiation, and interpolate and align to each sub-signal time sequence to output a synchronous temperature change rate vector; Refinement of micro-region temperature data There are two challenges in the original temperature data collected by the infrared thermal imager: Spatial specificity: the SPD surface is divided into 32 micro-regions (such as the center of the valve piece and the edge of the electrode), and each micro-region is analyzed independently; Time jitter: the temperature sampling rate (usually 100kS / s) is not consistent with the electrical signal sampling rate (possibly 1MS / s).

[0072] Solution steps: a) Micro-region matching: according to the SPD three-dimensional model coordinates, map the thermal imager pixels to the preset micro-regions, and take the median temperature of the pixels in each micro-region; b) Time stamp alignment: use the synchronous trigger signal (TTL pulse at the time of test signal injection) to align all sensor clocks; c) Missing value filling: use Lagrange interpolation method to fill in the missing intermediate points due to thermal imager frame rate limitation.

[0073] Temperature Change Rate Calculation and Physical Meaning Temperature Change Rate (TCR) is a key indicator to identify local overheating. The calculation uses the central difference method: TCR(t) = [T(t+Δt) - T(t-Δt)] / (2Δt) Where Δt is 1 millisecond (matched with the minimum time window of electrical signals). For example, when the grain boundary of a valve piece deteriorates, high-frequency current can cause a temporary surge in local TCR; poor electrode contact causes a sustained positive TCR in the medium frequency band. To enhance robustness, a moving average filter (window width 5 milliseconds) is applied, and sudden changes caused by environmental airflow are removed (|dT / dt|>10°C / s is considered invalid).

[0074] Cross-domain Time Series Synchronization Technology The temperature change rate sequence (time resolution 1ms) needs to be aligned with the electrical signal sub-signals (possibly 0.1ms resolution): Interpolation method: apply cubic spline interpolation to the TCR sequence to generate a continuous curve with 0.1ms intervals; Key point binding: at the characteristic moments of the test signal's step rising edge, triangular wave peak point, etc., the interpolated curve is forced to pass through the measured TCR value; Effectiveness verification: calculate the correlation coefficient of the interpolated TCR and the original TCR (requires >0.95), otherwise trigger reacquisition; The final output is a synchronized temperature change rate vector, which has the same length and timing as the electrical signal sub-signals, and the vector elements contain the TCR values of each micro-region.

[0075] Input the time-domain distortion factor set, frequency-domain feature set, and synchronized temperature change rate vector into the correlation engine to calculate the Pearson correlation coefficient of each feature parameter and the temperature change rate. Meanwhile, integrate energy entropy and waveform distortion factor to output the coupled feature matrix; Engineering Calculation of Pearson Correlation Coefficient Pearson Correlation Coefficient (PCC) quantifies the linear correlation strength between feature parameters and TCR. The calculation is performed for each feature (such as time-domain distortion factor_10ms) of each type of sub-signal (such as high-frequency current) and the corresponding micro-region TCR (such as valve piece center temperature): Input: time series with length N (N≈1000 sampling points); Calculate covariance and standard deviation: use Welford's online algorithm to avoid floating point overflow; Output PCC range [-1, 1]: positive value indicates the feature and TCR change in the same direction (e.g. TCR increases when distortion factor increases); To accelerate the calculation, deploy the parallel correlation engine: high / medium / low frequency sub-signals are processed synchronously on three FPGA logic units.

[0076] Physical mechanism-driven rules for feature fusion Simple PCC is not enough to describe complex degradation, and time and frequency domain features need to be fused: Energy entropy-distortion factor coupling: calculate the product of EE and WDF (EE x WDF), amplify the contribution when both are abnormal (e.g. EE and WDF increase when the grain boundary breaks); Harmonic-temperature synergy index: when H3 / F1>0.1 and TCR>2°C / s, mark the "grain boundary overheating risk" flag; Cross-band correlation: calculate the mutual information (Mutual Information) of high-frequency energy entropy and low-frequency temperature change rate to detect thermal-electric coupling effect; The fusion process follows the priority of the degradation physical model: electrode contact degradation is dominated by medium frequency features, and grain boundary characteristics depend on high frequency bands.

[0077] Structured construction of coupled feature matrix The output matrix is a two-dimensional table: Row: features of each sub-signal (e.g. 5 frequency domain features of high frequency voltage + 3 time domain distortion factors) Column: four types of data items, including: Original feature value (e.g. H3 / F1_P75=0.15); PCC with TCR (e.g. 0.62); Fusion index (e.g. EE x WDF=8.7); Risk flag (e.g. grain boundary overheating risk=1).

[0078] Matrix is screened by feature importance: rows with PCC absolute value>0.3 or fusion index exceeding threshold are retained, and redundant items are removed.

[0079] The coupled feature matrix is normalized and organized into a structured vector according to the frequency band dimension, and the vector elements include the energy entropy, waveform distortion factor, specific harmonic component amplitude ratio and correlation coefficient with temperature change rate of each frequency band, generating a feature vector representing the SPD aging state.

[0080] Normalization processing adapts to dynamic working conditions Different features have significant dimensional differences (e.g. energy entropy is dimensionless, harmonic amplitude ratio is between 0~1, distortion factor can be >10), which need to be normalized to [-1, 1] interval: X_norm = 2 × (X - X_min) / (X_max - X_min) - 1.

[0081] The key is dynamic boundary setting: X_min / X_max is not fixed, but is taken from the 5th and 95th percentile values (P5 / P95) of the SPD historical data; If it is a new device, the statistical boundary of the same type of SPD cluster is called.

[0082] For existing range features such as PCC, the arctangent transformation is used to compress extreme values: PCC_norm = arctan(PCC) × 2 / π.

[0083] Dimension design of structured vector The vector is organized in a "frequency band-physical quantity-feature type" hierarchy: First level: high frequency, medium frequency, and low frequency; Second level: within each frequency band, there are three physical quantities (voltage, current, and temperature) (temperature only in low frequency band); Third level: each physical quantity contains fixed feature elements (4 items in total): Energy entropy mean (EE_Mean); 10ms scale waveform distortion factor (WDF_10ms); Harmonic amplitude ratio (H3 / F1_P75, H5 / F1_P75) -> take the mean; PCC with TCR.

[0084] Example vector segment (low frequency band-current part): [0.72, 1.35, 0.08, -0.33] -> meaning: EE_Mean=0.72 (entropy value is moderate); WDF_10ms=1.35 (distortion is more serious); harmonic ratio mean=(H3 / F1+H5 / F1) / 2=0.08; PCC=-0.33 (current distortion and temperature change are negatively correlated).

[0085] Verification and compression of feature vector After generation, three verifications are performed: Integrity check: the length of the vector should be a fixed value (3 frequency bands × 3 physical quantities × 4 features - high frequency / medium frequency missing temperature item = 32 dimensions); Range check: each element value should be between [-1.5, 1.5] (allowing moderate over-limit); Logical consistency: the absolute value of high-frequency band PCC is usually > 0.4, otherwise an alarm is raised.

[0086] Finally, the vector is compressed to 16 dimensions through principal component analysis (PCA) to reduce the complexity of the subsequent model while retaining 95% of the variance. This vector represents the characteristic vector of the SPD aging state and is delivered to the lightweight deep assessment model.

[0087] Through time-frequency joint analysis, multi-dimensional feature parameters are extracted. Energy entropy reflects signal complexity, waveform distortion factor represents nonlinearity, and harmonic component ratio reveals material property changes. In particular, temperature change rate correlation analysis is introduced to establish a correlation model between electrical parameters and thermal properties, forming a comprehensive feature space that describes the aging state and constructing a comprehensive feature system that integrates electrical-thermal multi-parameters, overcoming the shortcomings of traditional methods that rely on a single parameter. The introduction of temperature correlation significantly improves the sensitivity of the feature to thermal degradation, making the evaluation results more physically meaningful.

[0088] S204, input the feature vector into the pre-trained lightweight deep assessment model to output a comprehensive aging score reflecting the overall degradation of the SPD in real time. The lightweight deep assessment model is trained based on SPD accelerated aging test data using transfer learning and attention mechanism fusion algorithm, and can adaptively weight the contribution of different physical field features to degradation. Specifically, the pre-trained lightweight deep assessment model can be loaded from the embedded storage system. This model is trained based on SPD accelerated aging test data, uses transfer learning to initialize parameters, has a compressed convolutional neural network architecture, and outputs a model instance. When the detection system starts the aging assessment process, the embedded processor (such as ARM Cortex-A72 architecture) first accesses its on-board embedded storage system (ESS), which usually uses eMMC 5.1 standard flash memory chips with a capacity of 32GB. The stored model file is saved in binary format, and the file name contains the model version identifier (such as "SPD_Model_v2.3.bin"). The loading process is implemented through direct memory access (DMA) technology for high-speed transmission, loading the model weight parameters and structure configuration information into the processor's dynamic random access memory (DRAM). The core architecture of this lightweight deep assessment model is a specially compressed convolutional neural network (CNN), which is compressed in three ways: Channel Pruning: The number of channels in the original CNN's convolutional layer is reduced from 256 to 64, reducing the computational load by 75%. Quantization Compression: 32-bit floating-point weights are converted to 8-bit integers (INT8), reducing the model size to 1.2 MB. Layer Fusion: Combining the convolutional layer (Convolution), batch normalization layer (Batch Normalization), and activation layer (ReLU) into a single computing unit to reduce memory access delay.

[0089] The pre-training basis of the model comes from the SPD accelerated aging test data, which contains three types of key samples: Electrical stress aging group: applying a 1.2 times rated voltage continuous bias to SPD to simulate long-term overvoltage working conditions; Thermal stress aging group: aging in a 85°C constant temperature oven for 500 hours to accelerate the thermal failure process; Composite stress aging group: alternating 8 / 20 μs lightning current impulse (15 kA) and temperature cycle (-40°C to +125°C).

[0090] Each sample collects more than 100,000 groups of multi-physical field response signals, and the real degradation degree is calibrated by physical means such as metallographic section and X-ray photoelectron spectroscopy. The model parameter initialization uses the transfer learning strategy: the backbone network of MobileNetV2 is used as the basis model (pre-trained on ImageNet dataset), and the last classification layer is replaced by a fully connected layer (containing 128 neurons) that adapts to SPD degradation score. During the transfer process, the first three layers of convolutional weights are frozen, and only the high-level feature extraction layer is fine-tuned to avoid small sample overfitting.

[0091] After the model is loaded, the system performs integrity check (Integrity Check): Calculate the cyclic redundancy check (Cyclic Redundancy Check, CRC) of the model file and compare it with the pre-stored check value. Run the test inference task, input the preset verification feature vector (Verification Feature Vector, VFV), and confirm that the output score is within the expected range (such as 85±2).

[0092] After the check, the model instance is activated and resides in memory, waiting to receive real-time feature vectors. At this time, the model is in a low-power standby state, and the processor reduces the frequency to 800 MHz through dynamic voltage and frequency scaling (DVFS) to save energy, and immediately restores to 1.8 GHz full-speed operation when a data input request is detected.

[0093] Before inputting the feature vector into the model instance, a dynamic standardization module based on Z-score is applied to adjust the feature scale according to the real-time SPD operating environment, ensuring compatibility with the input model, and obtaining a standardized feature vector; The original feature vector (Feature Vector, FV) contains four types of heterogeneous parameters: energy entropy of each frequency band (range 0-10), waveform distortion factor (range 1-5), harmonic amplitude ratio (range 0.01-0.5), and temperature-related coefficient (range -1 to +1). To eliminate dimensional differences, the system calls a dynamic standardization module (Dynamic Standardization Module, DSM), whose core algorithm uses an improved Z-score normalization. Unlike traditional Z-score, the mean μ and standard deviation σ of this module are not fixed values, but are dynamically calculated according to the real-time environment: Environmental temperature compensation: Read the current temperature T_env (unit °C) of the SPD installation location through a digital temperature sensor (such as DS18B20), and when T_env> 40°C, apply a temperature compensation coefficient K_temp=1+0.02×(T_env-40) to the energy entropy parameter; Grid voltage fluctuation compensation: Monitor the real-time voltage effective value U_grid (unit V) of the power grid, and if U_grid deviates from the rated value by ±10%, multiply the harmonic amplitude ratio parameter by a correction factor K_volt=220 / U_grid.

[0094] The dynamic standardization calculation process consists of three steps: Sliding window statistics: Take 1000 groups of feature vectors of the same type of SPD in the last 24 hours as the reference set, and roll the mean μ_roll and standard deviation σ_roll of each feature; Environmental parameter weighting: Input the real-time environmental parameters (T_env, U_grid) into the compensation function to generate a dynamic offset Δμ and a scaling factor λ, for example: Δμ = 0.3×K_temp - 0.1×K_volt; λ = 1.2 - 0.05×|U_grid-220|.

[0095] Standardization: For the i-th element x_i in the feature vector, the standardized value z_i = (x_i - (μ_roll_i + Δμ_i)) / (λ_i × σ_roll_i) is calculated.

[0096] This process ensures that the feature vectors collected at different times are comparable, even under grid fluctuations or high-temperature environments. For example, if a harmonic amplitude is 0.25 of the original value, the compensated z_i falls within the model-compatible interval of [-1.5, +1.5].

[0097] The standardization module adopts a hardware acceleration design: A dedicated pipeline is implemented in an FPGA chip (model Xilinx Zynq-7000), which includes a 32-bit floating-point multiplier (Multiplier Unit, MU) and an adder (Add Unit, AU); The time consumption of a single standardization is less than 50μs, meeting the real-time requirements; The output normalized feature vector (NFV) is a length-48 array (3 frequency bands × 4 types of features × 4 parameters) in IEEE 754 single-precision floating-point format. The NFV is encapsulated as a data structure and directly transmitted to the model inference engine through memory sharing.

[0098] After the normalized feature vector is input into the model, the built-in attention mechanism fusion algorithm is activated. This algorithm adaptively weights different physical field features and outputs the original degradation score through a fully connected layer; The front end of the model is a two-layer convolution module: The first convolution layer uses 16 3×1 kernels (since the feature vector is one-dimensional time series data) with a stride of 2 to extract local feature patterns; The second convolution layer uses 32 5×1 kernels with a stride of 1 to capture long-range dependencies.

[0099] After dimensionality reduction through max pooling, the convolution output enters the core attention mechanism fusion algorithm (AMFA). This algorithm includes two parallel branches: The channel attention branch (CAB) compresses the spatial dimension through global average pooling (GAP) to generate a channel description vector, which is then processed through two fully connected layers (32 neurons → 16 neurons) to generate channel weights W_ch; Physical Field Semantic Branch (PFSB): According to the pre-defined physical field labels in the feature vector (such as "voltage high frequency band", "current medium frequency band", "temperature low frequency band"), a learnable semantic embedding vector (SEV) is assigned to different physical fields, and the cross-field association weight W_sem is calculated by dot product similarity.

[0100] The attention weight fusion formula is: W_att = a x W_ch + (1-a) x W_sem. Where a is the adaptive mixing coefficient, the initial value is 0.5, and it is optimized by back propagation during training. Weight application process: Multiply the feature map F (size 16x24) of the convolution output by W_att in the channel dimension; For high contribution features (W_att_i > 0.8), start feature boosting, and magnify the value by 1.5 times; For low contribution features (W_att_i < 0.2), perform attenuation (Attenuation), with a scaling factor of 0.6.

[0101] This design enables the model to dynamically focus on key degradation signals, for example, when electrode contact degradation is detected, the weight of the current medium frequency band feature is automatically increased to 0.9 or above.

[0102] The weighted features are input into the fully connected layer (FCL) for decision-making: The first fully connected layer: 256 neurons, activation function ReLU (Rectified Linear Unit), dropout rate (Dropout Rate) set to 0.3 to prevent overfitting; The second fully connected layer: 128 neurons, using LeakyReLU (negative slope 0.01) to solve the gradient vanishing problem; Output layer: single neuron linear output, generating the original degradation score (RDS).

[0103] The RDS range is [-10, +10], negative values indicate better performance than the baseline, and positive values indicate degradation. The inference process enables hardware acceleration and executes 8-bit integer quantization inference on the NPU (Neural Processing Unit), with a single time consumption of <15ms.

[0104] The raw degradation score is mapped to the range of 0-100 using a sigmoid function and combined with real-time calibration parameters to generate a comprehensive aging score.

[0105] The raw degradation score (RDS) needs to be converted to an intuitive percentage score. The system uses a sigmoid function for non-linear mapping: Calculate the base score: S_base = 100 / (1 + e^(-k×RDS)); Where k is the curve steepness coefficient, optimized to k=0.25 in training, so that RDS=0 S_base=50 points, RDS=+8 S_base≈90 points.

[0106] This mapping converts RDS from [-10, +10] to continuous values in (0,100), for example, RDS=+3.2 corresponds to S_base=68.7 points. The conversion process is accelerated by a lookup table (LUT) in hardware: 256 sets of input and output corresponding values are pre-stored in the FPGA, and linear interpolation is used to achieve zero-delay calculation.

[0107] Real-time calibration (Real-time Calibration) is aimed at online running environment deviation: Load current compensation: The real-time load current I_load (unit A) of the SPD online circuit is obtained through the Hall current sensor (range 0-100A). When I_load>30A, a compensation term ΔS_current = 0.15×(I_load-30) is introduced; Cumulative current compensation: Read the total amount of lightning current ΣI_imp (unit kA) absorbed by the SPD from the historical database, and the compensation amount ΔS_imp = -0.2×ΣI_imp (because multiple lightning strikes will mask the aging signal); Time decay compensation: For SPDs installed for more than 5 years, add a decay factor ΔS_age = +1.5 points per year.

[0108] The calibration formula is: Comprehensive Aging Score (CAS) = S_base + ΔS_current + ΔS_imp + ΔS_age.

[0109] For example, the S_base of a certain SPD is 75 points, I_load=45A, ΣI_imp=120kA, and it has been installed for 7 years, then CAS=75+0.15×15-0.2×120+1.5×7=75+2.25-24+10.5=63.75 points.

[0110] The final output CAS needs to pass the confidence validation: Calculate the Mahalanobis Distance (MD) of the feature vector, if MD>3.0, trigger low confidence flag; When the CAS rate of change is >5 / min, start the re-detection mechanism (maximum 3 times to take the median); The results are written to the non-volatile memory (NVM) and an event log is generated, including timestamp, score value and calibration parameter details. The score is uploaded to the monitoring system through a 4-20mA analog signal or Modbus TCP protocol, with a refresh period ≤1 second, meeting the real-time requirements.

[0111] A lightweight deep learning model is used to realize real-time evaluation. The model inherits general degradation knowledge through transfer learning and dynamically adjusts the weights of different features through attention mechanism. The model training is based on accelerated aging test data, covering various typical degradation modes, ensuring the accuracy of evaluation and realizing intelligent identification and quantitative evaluation of complex degradation modes. The attention mechanism enables the model to adapt to different SPD types and operating environments, greatly improving the generalization ability of field application.

[0112] S205, in combination with the comprehensive aging score and the preset degradation level threshold, identifies the specific degradation mode and key degradation position inside the SPD, and generates a structured detection report containing the degradation level, risk position and maintenance suggestions.

[0113] Specifically, the comprehensive aging score can be compared with the preset degradation level threshold, which sets multiple levels based on historical failure data, and outputs specific degradation level labels; After the system receives the comprehensive aging score (CAS) output by the lightweight deep assessment model, the score is a value between 0 and 100, and the higher the value, the more serious the overall degradation of the surge protector SPD. The core basis for score comparison is the preset degradation level threshold (DLT). The threshold is not a single value, but a multi-level interval range set based on statistical analysis of massive historical failure data (HFD). Specifically, HFD comes from two aspects: one is the SPD failure cases recorded in the laboratory accelerated aging test (AAT), covering typical failures such as valve cracking, electrode corrosion under different voltage levels and environmental conditions; the other is the teardown analysis report (TAR) of retired SPDs in the field, which details the actual degradation location and degree. Through machine learning clustering algorithm (such as K-means) correlation analysis of aging characteristics and final failure mode in HFD, engineers divide DLT into four key intervals: healthy interval (CAS 0-20, green), mild degradation interval (CAS 21-50, yellow), moderate degradation interval (CAS 51-80, orange), and severe degradation interval (CAS 81-100, red). For example, when the system detects that the CAS of a certain SPD is 65, it is automatically matched to the "moderate degradation" interval, and outputs the level label DL_Moderate (Degradation Level Moderate).

[0114] Threshold dynamic adjustment mechanism: To ensure the adaptability of the threshold, the system is built-in Threshold Self-update Module (TSUM). This module continuously receives two types of new data: one is the abnormal CAS record (ACR) that exceeds the DLT warning range in the current detection; the second is the subsequent actual fault report (FR) of the same batch of SPD. TSUM uses sliding window statistics (SWS) to calculate the relevance of recent ACR and FR in a half-year cycle (window length WL=180 days). If it is found that the actual failure rate of a certain voltage grade SPD CAS in the 50-70 interval (originally defined as moderate degradation) exceeds the preset upper limit of the failure rate (Fault Rate Upper Limit, FRUL=15%), the threshold revision is automatically triggered: the upper limit of this interval is lowered from 70 to 65, and a new "high degradation" interval (CAS 65-75) is added. The revised DLT takes effect after being confirmed by artificial, ensuring that the grade label always reflects the true risk level.

[0115] Environmental parameter compensation: Since the aging rate of SPD is affected by the operating environment, the system introduces an environmental compensation coefficient (ECC) when comparing CAS and DLT. This coefficient is calculated by real-time monitoring of ambient temperature and humidity (ATH) and cumulative surge count (CSC). For example, when it is detected that the ambient temperature is continuously higher than 40 degrees Celsius (threshold T_high=40℃) and the CSC exceeds 1000 times (threshold CSC_high=1000), the ECC is set to 1.2, and the equivalent aging score (EAS) used for comparison at this time is EAS=CAS×ECC. If the original CAS of a certain SPD is 40, the EAS=48 under the condition of high temperature and high CSC, and the device originally belonging to "mild degradation" (CAS 21-50) will be upgraded to "moderate degradation" level (EAS>50), and the output label DL_Moderate_Enhanced. This dynamic compensation mechanism significantly improves the accuracy of degradation level determination.

[0116] Based on the degradation level label and feature vector, the internal degradation mode is mapped through the decision tree rule engine, and the degradation mode description is output. The core component of the degradation pattern mapping (DPM) is the decision tree rule engine (DTRE). The engine is composed of three layers: the first layer filters the possible pattern range according to the degradation level label (such as DL_Moderate), for example, moderate degradation only needs to check "valve plate local grain boundary degradation" or "electrode slight sulfuration"; the second layer combines the key parameters in the feature vector (FV) for deep analysis, the FV includes energy entropy, waveform distortion factor and other indicators of each frequency band; the third layer performs multi-condition combination judgment. Taking the identification of "electrode contact degradation" as an example: if the feature vector shows that the waveform distortion factor (WDF) of the mid-frequency sub-signal (MFS) is greater than 0.35 (threshold WDF_th=0.35), and the correlation coefficient (CC) between it and the temperature change rate (TCR) is less than 0.2 (threshold CC_low=0.2), then the "electrode contact abnormality" rule branch is triggered.

[0117] The construction of the decision tree rule base is based on the failure physics model (FPM) and the expert knowledge base (EKB). The FPM establishes the multi-field coupling equations of electricity-heat-force, for example, when the electrode sulfuration causes the increase of contact resistance, the specific harmonic amplitude ratio (SHAR) of the 3rd harmonic component of the mid-frequency current response will increase (typical value SHAR_3>15%). The EKB includes more than 300 criteria defined by industry experts, such as "when the energy entropy (EE) of the high-frequency sub-signal drops by more than 30% (threshold EE_drop=30%) accompanied by low-frequency temperature fluctuations, it indicates valve plate micro-cracks". These criteria are encoded as node splitting conditions of the decision tree, forming a rule network containing 52 judgment nodes and 18 terminal patterns. The engine uses the depth-first traversal (DFT) algorithm when executing, with an average processing delay of less than 50 milliseconds.

[0118] Pattern description dynamic generation: The output result is not fixed text, but is dynamically combined through a natural language generation template (NLGT). The template contains three variable slots: degradation subject (DS), degradation mechanism (DM), and severity modifier (SM). For example, when the engine identifies that the valve piece grain boundary oxygen diffusion is intensified: DS = "metal oxide valve piece", DM = "grain boundary oxygen ion migration causes barrier height to decrease", and SM = "locally fast". The final generated degradation pattern description is: "metal oxide valve piece is locally fast degraded due to grain boundary oxygen ion migration causing barrier height to decrease". All description words strictly follow the IEC standard terminology database (ISTD), ensuring the professionalism and normativity of the report.

[0119] In combination with the degradation pattern description and the key band sub-signal set, the back propagation algorithm is used to locate the internal risk positions of the SPD, and the risk position coordinates and types are output. The core of risk location positioning (RLP) is the signal-space mapping model (SSMM). This model divides the internal structure of the SPD into 200 virtual grid units (VGUs), each corresponding to a physical location (such as the upper left area of the valve piece or the electrode connection). The positioning process first determines the key area based on the degradation pattern description: if the description contains "valve piece", focus on the valve piece area grid (numbered VGU-101 to VGU-150); then extract the abnormal features in the key band sub-signal set (KBSS). For example, when locating the electrode sulfurization position, the fluctuation sequence of the time-domain distortion factor (TDF) of the medium frequency sub-signal (MFS) during the signal duration needs to be analyzed, and the peak value time corresponds to the sulfurization point through-flow instant.

[0120] Back Propagation Algorithm (BPA) is used to optimize the positioning accuracy. The inputs of the algorithm include: the eigenvalues of each sub-signal in KBSS at a specific time (e.g. the energy entropy EE_hf=0.82 of the high-frequency sub-signal at t=15 ms), and the 3D Thermo-mechanical Topology Map (3D-TMTM) of SPD. In the initialization stage, a Failure Probability Weight (FPW) is assigned to each VGU, with a default value of 0.01. BPA performs three steps of iteration: forward propagation to calculate the theoretical value of each VGU abnormal signal; error calculation to compare the theoretical value with the actual detection value (e.g. the actual EE_hf=0.82 of VGU-125, the theoretical value 0.75, the error 0.07); and backward weight correction to adjust the FPW according to the error. For example, when the TDF error of electrode area VGU-117 is consistently positive, its FPW will gradually increase from 0.01 to 0.78 (threshold FPW_high=0.75), and it will be determined as a risk site.

[0121] Coordinate and type output rules: The positioning results need to meet the Dual-condition Filter (DCF) of two conditions: 1) Grid FPW > FPW_high (0.75); 2) There are at least 2 FPW > 0.6 (threshold FPW_medium=0.6) in the adjacent grid to form an Abnormal Cluster (AC). The coordinate output uses Hierarchical Encoding (HE): the first letter represents the component type (V=valve, E=electrode), the middle number is the grid row number, and the last number is the column number. For example, "E-09-12" represents the 9th row and 12th column grid in the electrode area. The type description refers to the Standard Part Classification Table (SPCT), such as "V-TypeA" representing A-type zinc oxide valve, and "E-CuSn" indicating copper-tin alloy electrode. The final output format is: "Risk site: E-09-12 (electrode contact surface sulfuration)".

[0122] The deterioration level label, deterioration mode description, and risk site are input into the report template engine to automatically generate a structured detection report containing the deterioration level, risk site, and maintenance recommendations.

[0123] The report template engine (RTE) adopts a dynamic filling mechanism (DFM) based on XML architecture. The engine is pre-installed with four core modules: 1) a header generator (HG) that inserts basic information such as detection time and SPD number; 2) a level visualization module (LVM) that converts the degradation level label into an intuitive icon (e.g., DL_Moderate corresponds to an orange warning icon); 3) a pattern parser (PP) that disassembles the degradation mode description into three fields: cause, phenomenon, and impact; and 4) a location mapper (LM) that labels coordinates such as "E-09-12" on the SPD structure profile. After the engine is started, it first verifies the logical completeness (LC) of the input data to ensure that the level label, mode description, and risk location are not contradictory (e.g., "severe degradation" must correspond to at least one high-risk location).

[0124] Intelligent generation of maintenance recommendations: based on knowledge graph reasoning (KGR), the system associates the degradation mode with the maintenance strategy library (MSL). The MSL contains 42 standard strategies, each consisting of a trigger condition (TC), an action command (AC), and an expected outcome (EO). For example, when the mode description contains "grain boundary oxygen migration" and the level is DL_Moderate, the trigger strategy TC-07 is activated: AC = "72-hour resistance current test", EO = "quantify the valve aging rate". Safety constraints (SC) are automatically added when generating recommendations, such as mandatory "power-off operation" for high-voltage SPDs (> 40 kV).

[0125] The report structured output: the final report adopts a layered foldable PDF (LF-PDF) format, including five parts: 1) the summary page: the degradation level (such as "moderate degradation-orange warning"), the number of risk parts (such as "3 high-risk points"); 2) the detail page: showing the degradation mode description (including mechanism diagram) one by one; 3) the positioning page: the SPD three-dimensional cross-sectional view is marked with risk coordinates, supporting click highlighting; 4) the maintenance page: the suggestion list is sorted by priority (such as "immediate action: clean the electrode contact surface (E-09-12 area)"); 5) the appendix page: the key feature vector values (such as high-frequency energy entropy = 0.92 > threshold 0.85). After the report is generated, a digital signature (DS) and blockchain notarization (BN) are automatically attached to ensure that the data cannot be tampered with. The entire process is completed within 8 seconds, and the output file is less than 2MB.

[0126] According to the scoring threshold, the degradation level is divided, the feature vector is reversely positioned to the specific degradation part (such as the center or edge area of the valve piece), and a customized report containing specific maintenance measures (such as cleaning the contact surface or replacing the whole) is automatically generated. Abstract scoring is converted into specific and executable maintenance decisions to guide on-site personnel to accurately carry out preventive maintenance, effectively prolong the service life of SPD and ensure the safety of the power system.

[0127] It can be seen that when the surge protector SPD is in an online running state, a non-destructive composite test signal is injected into the protection line of the SPD, and real-time multi-physical field response signals of the SPD to the composite test signal are collected; the collected multi-physical field response signals are subjected to multi-band parallel decomposition to obtain a set of key frequency band sub-signals reflecting the characteristics of different degradation mechanisms; time domain and frequency domain feature parameters are synchronously extracted from each key frequency band sub-signal to construct a feature vector representing the multi-dimensional aging state of the SPD; the feature vector is input into a pre-trained lightweight deep evaluation model to output a comprehensive aging score reflecting the overall degradation degree of the SPD in real time; combined with the comprehensive aging score and the preset degradation level threshold, a structured detection report is generated, so that the multi-mechanism degradation of the SPD in the online state can be accurately identified, and the real-time and reliability of the aging score are improved.

[0128] Another embodiment of the present application provides a surge protector aging degradation detection system, as shown in Figure 3 , the system can include: The collection module 301 is configured to inject a non-destructive composite test signal into a protection line of a surge protector (SPD) when the SPD is in an online running state, and collect a real-time multi-physical field response signal of the SPD to the composite test signal, wherein the composite test signal is composed of a triangular wave of a specific frequency range and a step signal of a preset amplitude, and the multi-physical field response signal at least includes a transient voltage response, a transient current response, and a SPD surface micro-area temperature response at a corresponding time point. The decomposition module 302 is configured to perform multi-band parallel decomposition on the collected multi-physical field response signal to obtain a key frequency band sub-signal set reflecting characteristics of different degradation mechanisms, wherein the key frequency band sub-signal set at least includes a high-frequency sub-signal reflecting a metal oxide valve sheet grain boundary characteristic, a medium-frequency sub-signal reflecting an electrode contact degradation, and a low-frequency sub-signal reflecting a thermal accumulation characteristic. The extraction module 303 is configured to synchronously extract time domain and frequency domain feature parameters from each key frequency band sub-signal, and construct a feature vector representing a multi-dimensional aging state of the SPD by combining a micro-area temperature change rate at a corresponding time point, wherein the feature vector includes energy entropy, waveform distortion factor, specific harmonic component amplitude ratio, and a correlation coefficient with the temperature change rate of each frequency band. The evaluation module 304 is configured to input the feature vector into a pre-trained lightweight deep evaluation model to output a comprehensive aging score reflecting an overall degradation degree of the SPD in real time, wherein the lightweight deep evaluation model is trained based on SPD accelerated aging test data by using a transfer learning and attention mechanism fusion algorithm, and can adaptively weight the contribution of different physical field characteristics to degradation. The generation module 305 is configured to identify a specific degradation mode and a key degradation position inside the SPD by combining the comprehensive aging score and a preset degradation level threshold, and generate a structured detection report including a degradation level, a risk position, and a maintenance suggestion.

[0129] The embodiment of the present application also provides a storage medium, wherein the storage medium stores a computer program, and the computer program is set to execute the steps in any one of the method embodiments.

[0130] Specifically, in the present embodiment, the above-mentioned storage medium can be set to store a computer program for executing the following steps: S201, when a surge protector (SPD) is in an online running state, a non-destructive composite test signal is injected into a protection line of the SPD, and a real-time multi-physical field response signal of the SPD to the composite test signal is collected, wherein the composite test signal is composed of a triangular wave of a specific frequency range and a step signal of a preset amplitude, and the multi-physical field response signal at least includes a transient voltage response, a transient current response, and a SPD surface micro-area temperature response at a corresponding time point. S202, the collected multi-physical field response signal is decomposed in multiple frequency bands in parallel to obtain a key frequency band sub-signal set reflecting different degradation mechanism characteristics, wherein the key frequency band sub-signal set at least contains a high-frequency sub-signal reflecting metal oxide valve sheet grain boundary characteristics, a medium-frequency sub-signal reflecting electrode contact degradation and a low-frequency sub-signal reflecting thermal accumulation characteristics; S203, time domain and frequency domain feature parameters are synchronously extracted from each key frequency band sub-signal, and a feature vector representing the multi-dimensional aging state of the SPD is constructed in combination with the micro-area temperature change rate at the corresponding time point, wherein the feature vector contains the energy entropy, waveform distortion factor, specific harmonic component amplitude ratio and correlation coefficient of the temperature change rate of each frequency band; S204, the feature vector is input into a pre-trained lightweight deep evaluation model to output a comprehensive aging score reflecting the overall degradation degree of the SPD in real time, wherein the lightweight deep evaluation model is trained based on SPD accelerated aging test data using a transfer learning and attention mechanism fusion algorithm, and can adaptively weight the contribution of different physical field characteristics to degradation; S205, in combination with the comprehensive aging score and the preset degradation level threshold, the specific degradation mode and key degradation position inside the SPD are identified, and a structured detection report containing the degradation level, risk position and maintenance suggestion is generated.

[0131] The embodiment of the application also provides an electronic device comprising a memory and a processor, the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any of the method embodiments.

[0132] Specifically, the electronic device described above can further include a transmission device and an input-output device, wherein the transmission device is connected to the processor, and the input-output device is connected to the processor.

[0133] Specifically, in the embodiment, the processor can be configured to execute the following steps through the computer program: S201, when the surge protector SPD is in an online running state, a non-destructive composite test signal is injected into the protection line thereof, and a real-time multi-physical field response signal of the SPD to the composite test signal is collected, wherein the composite test signal is composed of a triangular wave of a specific frequency range and a step signal of a preset amplitude, and the multi-physical field response signal at least includes a transient voltage response, a transient current response and a SPD surface micro-area temperature response at the corresponding time point; S202, the collected multi-physical field response signals are subjected to multi-band parallel decomposition to obtain a key frequency band sub-signal set reflecting different degradation mechanism characteristics, wherein the key frequency band sub-signal set at least contains a high-frequency sub-signal reflecting metal oxide valve sheet grain boundary characteristics, a medium-frequency sub-signal reflecting electrode contact degradation and a low-frequency sub-signal reflecting thermal accumulation characteristics; S203, time domain and frequency domain feature parameters are synchronously extracted from each key frequency band sub-signal, and a micro-area temperature change rate at a corresponding time point is combined to construct a feature vector representing a multi-dimensional aging state of the SPD, wherein the feature vector contains energy entropy, waveform distortion factor, specific harmonic component amplitude ratio and a correlation coefficient with the temperature change rate of each frequency band; S204, the feature vector is input into a pre-trained lightweight deep evaluation model to output a comprehensive aging score reflecting the overall degradation degree of the SPD in real time, wherein the lightweight deep evaluation model is trained based on SPD accelerated aging test data using a transfer learning and attention mechanism fusion algorithm, and can adaptively weight the contribution of different physical field characteristics to degradation; S205, in combination with the comprehensive aging score and a preset degradation level threshold, a specific degradation mode and a key degradation position inside the SPD are identified, and a structured detection report containing a degradation level, a risk position and a maintenance suggestion is generated.

[0134] The above embodiments according to the drawings illustrate the structure, features and effects of the present application. The above description is only a preferred embodiment of the present application, but the present application is not limited by the drawings. Any changes or modifications made in accordance with the concept of the present application, or equivalent embodiments with equivalent changes, are still within the scope of the present application.

Claims

1. A method of detecting aging deterioration of a surge protector, characterized by, The method comprises: When the surge protector SPD is in an online running state, a non-destructive composite test signal is injected into the protection line thereof, and real-time multi-physical field response signals of the SPD to the composite test signal are collected, wherein the composite test signal is composed of a triangular wave of a specific frequency range and a step signal of a preset amplitude, and the multi-physical field response signals at least include transient voltage response, transient current response and SPD surface micro-area temperature response at the corresponding time point; The collected multi-physical field response signals are subjected to multi-band parallel decomposition to obtain a key frequency band signal set reflecting different deterioration mechanism characteristics, wherein the key frequency band signal set at least contains a high-frequency signal reflecting metal oxide valve sheet grain boundary characteristics, a medium-frequency signal reflecting electrode contact deterioration and a low-frequency signal reflecting heat accumulation characteristics; Time domain and frequency domain characteristic parameters are synchronously extracted from each key frequency band signal, and a micro-area temperature change rate at the corresponding time point is combined to construct a characteristic vector representing the multi-dimensional aging state of the SPD, wherein the characteristic vector contains energy entropy, waveform distortion factor, specific harmonic component amplitude ratio and correlation coefficient of the temperature change rate of each frequency band; The characteristic vector is input into a pre-trained lightweight deep evaluation model to output a comprehensive aging score reflecting the overall deterioration degree of the SPD in real time, wherein the lightweight deep evaluation model is trained based on SPD accelerated aging test data by using a transfer learning and attention mechanism fusion algorithm, and can adaptively weight the contribution of different physical field characteristics to deterioration; In combination with the comprehensive aging score and a preset deterioration level threshold, a specific deterioration mode and key deterioration position inside the SPD are identified, and a structured detection report containing a deterioration level, a risk position and a maintenance suggestion is generated.

2. The method of claim 1, wherein, The method comprises: Based on the SPD online running parameters, a triangular wave sequence of a specific frequency range is generated in real time by a digital signal generator, the center frequency of the sequence is dynamically adjusted according to the rated voltage of the SPD to avoid resonance interference, and a frequency-adaptive triangular wave signal is output; A step signal of a preset amplitude is time-domain superimposed with the frequency-adaptive triangular wave signal to form an initial composite test signal, and the initial composite test signal is injected into the protection line through a current limiter and a voltage clamping circuit to ensure that the signal amplitude is lower than the action threshold of the SPD, and a non-destructive injection signal is output; The transient voltage response and the transient current response of the SPD to the non-destructive injection signal are synchronously captured by using a high-speed data acquisition card, and an infrared thermal imager is used to record the SPD surface micro-area temperature response at a microsecond-level sampling rate, so that the time stamps of all response signals are aligned, and an original multi-physical field response data set is output; The original multi-physical field response data set is subjected to time domain interpolation processing to eliminate collection delay, and an adaptive Kalman filter is applied to remove environmental noise, thereby generating time-synchronized and denoised real-time multi-physical field response signals.

3. The method of claim 2, wherein, The collected multi-physical field response signals are subjected to multi-band parallel decomposition to obtain a key frequency band sub-signal set reflecting different degradation mechanism characteristics, wherein the key frequency band sub-signal set at least includes a high-frequency sub-signal reflecting metal oxide valve sheet grain boundary characteristics, a medium-frequency sub-signal reflecting electrode contact degradation, and a low-frequency sub-signal reflecting thermal accumulation characteristics, and the method comprises the following steps: The voltage, current and temperature signals in the real-time multi-physical field response signals are subjected to band analysis, the initial frequency band boundaries covering high, medium and low frequencies are predefined through fast Fourier transform, and a preliminary frequency band division scheme is outputted; Based on the preliminary frequency band division scheme, a parallel wavelet packet decomposition algorithm is applied to each physical field signal to extract multiple sub-band signals, and the energy entropy of each sub-band is calculated to screen candidate sub-bands with energy entropy greater than an energy entropy threshold, thereby outputting a high-energy sub-band set; The high-energy sub-band set is mapped to the SPD degradation physical mechanism, wherein the high-frequency sub-band is associated with the metal oxide valve sheet grain boundary characteristics, the medium-frequency sub-band is associated with the electrode contact degradation, and the low-frequency sub-band is associated with the thermal accumulation characteristics, the characteristics are enhanced through a mechanism-related filter, and a mechanism-labeled sub-band signal is outputted; The mechanism-labeled sub-band signal is subjected to cross-physical field frequency band correlation analysis, similar frequency bands are fused through a mutual information algorithm, and redundancy is eliminated, thereby generating optimized high-frequency, medium-frequency and low-frequency sub-signals; The optimized high-frequency, medium-frequency and low-frequency sub-signals are classified and integrated according to physical mechanisms to form a key frequency band sub-signal set containing grain boundary characteristics, electrode degradation and thermal accumulation characteristics.

4. The method of claim 3, wherein, Synchronous time domain and frequency domain characteristic parameters are extracted from each key frequency band sub-signal, and the micro-area temperature change rate at the corresponding time point is combined to construct a feature vector representing the SPD multi-dimensional aging state, wherein the feature vector contains the energy entropy, waveform distortion factor, specific harmonic component amplitude ratio and correlation coefficient with the temperature change rate of each frequency band, and the method comprises the following steps: For each sub-signal in the key frequency band sub-signal set, a time domain waveform distortion factor is calculated, the factor is obtained by comparing the root mean square error of the actual waveform and the ideal reference waveform, and a time domain distortion factor set of each sub-signal is outputted; Short-time Fourier transform is applied to the same sub-signal set to synchronously extract the energy entropy and specific harmonic component amplitude ratio of each frequency band, wherein the harmonic component amplitude ratio focuses on the 3rd and 5th harmonics, and a frequency domain feature set of each sub-signal is outputted; Surface micro-area temperature data at the corresponding time point are extracted from the real-time multi-physical field response signals, the temperature change rate is calculated through first-order differentiation, and the temperature change rate is aligned to the time sequence of each sub-signal through interpolation, thereby outputting a synchronous temperature change rate vector; The time domain distortion factor set, frequency domain feature set and synchronous temperature change rate vector are inputted into a correlation engine to calculate the Pearson correlation coefficient of each characteristic parameter and the temperature change rate, and the energy entropy and waveform distortion factor are fused, thereby outputting a coupled feature matrix; The coupling feature matrix is normalized and organized as a structured vector in the frequency band dimension, and the vector elements include the energy entropy, waveform distortion factor, specific harmonic component amplitude ratio, and correlation coefficient with temperature change rate of each frequency band, to generate a feature vector representing the SPD aging state.

5. The method of claim 4, wherein, The feature vector is input into a pre-trained lightweight deep evaluation model, which outputs a comprehensive aging score reflecting the overall degradation degree of the SPD in real time, wherein the lightweight deep evaluation model is trained based on SPD accelerated aging test data using a transfer learning and attention mechanism fusion algorithm, and can adaptively weight the contribution of different physical field features to degradation, including: Load the pre-trained lightweight deep evaluation model from the embedded storage system, which is trained based on SPD accelerated aging test data, uses transfer learning to initialize parameters, and has a compressed convolutional neural network architecture, and output a model instance; Before inputting the feature vector into the model instance, apply a dynamic standardization module based on Z-score to adjust the feature scale according to the real-time SPD operating environment, ensure input compatibility with the model, and obtain a standardized feature vector; After inputting the standardized feature vector into the model, activate the built-in attention mechanism fusion algorithm, which adaptively weights different physical field features and outputs the original degradation score through a fully connected layer; Apply a sigmoid function to map the original degradation score to the range of 0-100, and combine it with online operating parameters for real-time calibration to generate a comprehensive aging score.

6. The method of claim 5, wherein, The comprehensive aging score and the pre-set degradation level threshold are combined to identify the specific degradation mode and key degradation location inside the SPD, and a structured detection report containing the degradation level, risk location, and maintenance suggestions is generated, including: Compare the comprehensive aging score with the pre-set degradation level threshold, which sets multiple levels based on historical failure data, and output the specific degradation level label; Based on the degradation level label and the feature vector, map to the internal degradation mode through a decision tree rule engine, and output the degradation mode description; Combine the degradation mode description and the key frequency band sub-signal set, and use the backpropagation algorithm to locate the internal risk location of the SPD, and output the risk location coordinates and type; Input the degradation level label, degradation mode description, and risk location into the report template engine to automatically generate a structured detection report containing the degradation level, risk location, and maintenance suggestions.

7. A surge protector aging degradation detection system characterized by, The system comprises: The acquisition module is used to inject a non-destructive composite test signal into the protection line of the surge protector SPD when it is in an online running state, and to collect the real-time multi-physical field response signal of the SPD to the composite test signal, wherein the composite test signal is composed of a triangular wave of a specific frequency range and a step signal of a pre-set amplitude, and the multi-physical field response signal at least includes a transient voltage response, a transient current response, and a SPD surface micro-area temperature response at the corresponding time point; The decomposition module is configured to perform multi-frequency band parallel decomposition on the collected multi-physical field response signals to obtain a key frequency band sub-signal set reflecting characteristics of different degradation mechanisms, wherein the key frequency band sub-signal set at least includes a high-frequency sub-signal reflecting a metal oxide valve plate grain boundary characteristic, a medium-frequency sub-signal reflecting an electrode contact degradation, and a low-frequency sub-signal reflecting a thermal accumulation characteristic; The extraction module is configured to synchronously extract time domain and frequency domain characteristic parameters from each key frequency band sub-signal, and construct a feature vector representing a SPD multi-dimensional aging state by combining a micro-area temperature change rate at a corresponding time point, wherein the feature vector includes energy entropy, waveform distortion factor, specific harmonic component amplitude ratio, and a correlation coefficient of the temperature change rate of each frequency band; The evaluation module is configured to input the feature vector into a pre-trained lightweight deep evaluation model to output a comprehensive aging score reflecting an overall degradation degree of the SPD in real time, wherein the lightweight deep evaluation model is trained based on SPD accelerated aging test data by using a transfer learning and attention mechanism fusion algorithm, and can adaptively weight the contribution of different physical field characteristics to degradation; The generation module is configured to identify a specific degradation mode and a key degradation position inside the SPD by combining the comprehensive aging score and a preset degradation level threshold, and generate a structured detection report including a degradation level, a risk position, and a maintenance suggestion.

8. The system of claim 7, wherein, The acquisition module is specifically configured to: generate a triangular wave sequence of a specific frequency range in real time based on SPD online operation parameters through a digital signal generator, a center frequency of the sequence is dynamically adjusted according to a rated voltage of the SPD to avoid resonance interference, and a frequency adaptive triangular wave signal is output; superimpose a step signal with a preset amplitude and the frequency adaptive triangular wave signal in a time domain to form an initial composite test signal, and inject the initial composite test signal into a protection circuit through a current limiter and a voltage clamping circuit to ensure that a signal amplitude is lower than an SPD action threshold, and output a non-destructive injection signal; synchronously capture a transient voltage response and a transient current response of the SPD to the non-destructive injection signal by using a high-speed data acquisition card, and record a micro-area temperature response of the SPD surface at a microsecond level sampling rate by using an infrared thermal imager, so that time stamps of all response signals are aligned, and an original multi-physical field response data set is output; perform time domain interpolation processing on the original multi-physical field response data set to eliminate acquisition delay, and apply an adaptive Kalman filter to remove environmental noise, and generate a time-synchronized and denoised real-time multi-physical field response signal.

9. A storage medium, characterized by The storage medium has a computer program stored therein, and the computer program is configured to execute the method in any one of claims 1-6 when running.

10. An electronic device comprising a memory and a processor, characterized in that, The memory has a computer program stored therein, and the processor is configured to execute the computer program to execute the method in any one of claims 1-6.

Images