A relay operating state fault risk prediction method and system

Abstract

The present application relates to the technical field of fault diagnosis, in particular to a relay operation state fault risk prediction method and system. In view of the problem that transient micro degradation characteristics are easily covered by systematic baseline distortion of sensing circuit, the present application extracts the original pulse width modulation parasitic harmonic as a probe, calculates the equivalent reactance evolution characteristics of the relay, maps the temperature drift baseline and mechanical vibration vector to the Poincare phase space, extracts the nonlinear topological distortion manifold by calculating the maximum Lyapunov index jump, constructs a structural causal model containing thermal inertia delay hidden nodes and performs counterfactual intervention, cuts off the error transmission path of temperature drift on topological characteristics, and reconstructs the absolute physical degradation index stripped of measurement interference. The present application realizes the absolute decoupling of signal false distortion and real physical ablation, and ensures the accuracy of fault warning under complex extreme conditions.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, specifically to a method and system for predicting fault risks in relay operation. Background Technology

[0002] As new energy vehicles evolve towards 800V and above high-voltage, high-power fast-charging architectures, high-voltage DC relays, as core safety isolation components, directly affect the safety of the entire vehicle. Currently, relay health monitoring mainly relies on static threshold comparisons or conventional data-driven models. However, under extremely complex long-term on-board high-voltage operating conditions, existing technologies reveal serious bottlenecks in perception and prediction.

[0003] On the one hand, existing technologies struggle to address the nonlinear degradation abrupt changes caused by the coupling of multiple physical fields (thermal, mechanical, and electrical). When vehicle bumps and sudden large current changes occur simultaneously, relay contacts are highly susceptible to micro-bounces and high-temperature arcs, causing contact impedance to surge exponentially within milliseconds, creating a degradation "avalanche." Existing algorithms are prone to misinterpreting this high-frequency chaotic signal as external noise and filtering it out, or becoming overly sensitive and causing erroneous vehicle degradation. On the other hand, existing technologies generally overlook the systematic errors introduced by the sensing link. In long-term service and extreme thermal alternation environments, the acquisition circuitry within the battery management system is prone to temperature drift and parameter aging. This systematic baseline distortion of the sensing link severely overlaps with the actual impedance changes of the relays, leading to ambiguities in the state assessment of existing systems.

[0004] More seriously, when the above conditions are combined, long-period measurement baseline distortion can deeply mask the nonlinear transient degradation characteristics. Specifically, thermal drift in the sensing circuit can cause a false rise in the measurement baseline. If the relay experiences micro-bounce or arc avalanche at this time, the true high-frequency impedance spikes will be superimposed on the distorted baseline. When faced with such contaminated composite signals, existing algorithms are prone to misinterpreting fatal degradation abrupt changes as random noise caused by sensor thermal instability and filtering them out, directly leading to the loss of the optimal intervention window to prevent complete contact welding.

[0005] In summary, under the coexistence of nonlinear transient high-frequency interference and long-period systematic measurement reference distortion, how to overcome the masking effect of the distortion reference on transient change characteristics and achieve accurate measurement and reliable evaluation of the true electrical characteristics of relays is a technical problem that urgently needs to be solved in this field.

[0006] To address this, a method and system for predicting relay operation status failure risks are proposed. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for predicting fault risks in relay operation. Addressing the problem that transient micro-degradation characteristics are easily masked by the systematic baseline distortion of sensing circuits, this invention extracts the system's native pulse width modulation parasitic harmonics as probes to calculate the equivalent reactance evolution characteristics of the relay. It combines the temperature drift baseline and mechanical vibration vector mapping to the Poincaré phase space, and extracts the nonlinear topological distortion manifold by calculating the maximum Lyapunov exponential jump. A structural causal model including thermal inertial delay hidden nodes is constructed, and counterfactual intervention is performed to cut off the error propagation path of temperature drift to topological characteristics, reconstructing the absolute physical degradation exponent stripped of measurement interference. This invention achieves absolute decoupling between spurious signal distortion and real physical ablation, ensuring the accuracy of fault warnings under complex and extreme operating conditions.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting fault risk in relay operation includes acquiring voltage waveforms and synchronous pulse width modulation command sequences at both ends of the relay; extracting parasitic harmonics with frequencies that are preset multiples of the fundamental frequency of the pulse width modulation command sequence from the voltage waveforms as measurement probes; and calculating the equivalent reactance evolution characteristics of the relay based on the transfer function dispersion matrix generated by the signals generated by the measurement probes passing through the input and output ends of the relay. The temperature drift baseline of the sensing circuit and the mechanical vibration vector of the relay are collected and combined with the equivalent reactance evolution characteristics as state variables; the state variables are mapped to the Poincaré phase space to reconstruct the trajectory attractor; and the nonlinear topological distortion manifold is extracted by calculating the maximum Lyapunov exponential jump of the trajectory attractor. A structural causal model is constructed, encompassing ambient temperature, the temperature drift baseline, and the nonlinear topological distortion manifold. Based on the structural causal model, counterfactual intervention is performed. By shielding the error transmission effect of the temperature drift baseline on the nonlinear topological distortion manifold, the target state characteristics corresponding to a preset standard temperature reference environment are reconstructed. Based on the target state characteristics, an absolute physical degradation index is extracted for fault risk prediction.

[0009] Preferably, extracting parasitic harmonics with a frequency that is a preset multiple of the fundamental frequency of the pulse width modulation command sequence as a measurement probe includes: extracting the fundamental frequency period of the pulse width modulation command sequence using a synchronous phase-locked loop; truncating the voltage waveform according to the fundamental frequency period to obtain a voltage signal sequence with independent periods; performing a discrete Fourier transform on the voltage signal sequence to obtain a frequency domain amplitude-phase matrix; extracting spectral components in the frequency domain amplitude-phase matrix whose frequency values ​​are equal to the fundamental frequency multiplied by a preset multiple integer; reconstructing the time-domain parasitic harmonic signal by performing an inverse Fourier transform on the spectral components; and using the parasitic harmonic signal as the measurement probe.

[0010] Preferably, calculating the equivalent reactance evolution characteristics of the relay by passing through the transfer function dispersion matrix generated by the signals at the relay input and output terminals includes: calculating the ratio of the cross power spectral density and the self power spectral density of the signals at the relay input and output terminals by the measurement probe, generating a transfer function dispersion vector characterizing the physical response of the relay; continuously extracting the transfer function dispersion vectors of multiple independent observation periods within a preset time sliding window and arranging and splicing them in time sequence to construct a transfer function dispersion matrix; decomposing the transfer function dispersion matrix into a real impedance matrix and an imaginary reactance matrix; extracting the eigenvalue sequence of the first preset number of principal elements of the imaginary reactance matrix, and using the eigenvalue sequence as the equivalent reactance evolution characteristics of the relay.

[0011] Preferably, mapping the state variables to the Poincaré phase space to reconstruct the trajectory attractor includes: aligning and normalizing the temperature drift baseline, the mechanical vibration vector, and the equivalent reactance evolution characteristics using an absolute physical time reference, and concatenating them into a multidimensional temporal state vector; calculating the delay time parameter of the multidimensional temporal state vector using the mutual information method; calculating the embedding dimension parameter of the multidimensional temporal state vector using the spurious nearest neighbor method; performing time delay coordinate mapping on the multidimensional temporal state vector based on the delay time parameter and embedding dimension parameter fixed during system initialization to generate a phase space coordinate point set; and using the phase space coordinate point set to fit and generate a closed trajectory attractor in the Poincaré phase space.

[0012] Preferably, the nonlinear topological distortion manifold is extracted by calculating the maximum Lyapunov exponent jump of the trajectory attractor, including: selecting an initial reference point on the trajectory attractor and calculating the initial neighboring point with the smallest Euclidean distance to the initial reference point; calculating the L2 norm of the difference vector of adjacent sampling points of the state variable within a time sliding window as a one-dimensional time-domain energy gradient; when the one-dimensional time-domain energy gradient exceeds a preset wake-up threshold, truncating a local time window containing feature mutations; performing local phase space reconstruction on the state variables within the local time window to generate a local trajectory attractor, and calculating the short-time maximum Lyapunov exponent of the local trajectory attractor; when the short-time maximum Lyapunov exponent is greater than a preset chaotic bifurcation threshold, taking the phase space coordinate point set region corresponding to the local time window as the nonlinear topological distortion manifold.

[0013] Preferably, constructing a structural causal model encompassing ambient temperature, the temperature drift baseline, and the nonlinear topological distortion manifold includes: constructing a causal directed graph with the ambient temperature as the root node, the temperature drift baseline as intermediate disturbance nodes, and the nonlinear topological distortion manifold as observation result nodes; defining the path from the root node to the intermediate disturbance node as a first directed edge, and the path from the intermediate disturbance node to the observation result node as a second directed edge; introducing a thermal inertial delay hidden node representing heat conduction delay between the root node and the intermediate disturbance node, and assigning the real-time virtual junction temperature calculated using an equivalent thermal network conduction model to the thermal inertial delay hidden node; training and outputting a first structural mapping function for the temperature drift baseline based on historical data, and fitting and outputting a second structural mapping function for the nonlinear topological distortion manifold using a multilayer perceptron network; and combining the causal directed graph, the first structural mapping function, and the second structural mapping function to generate the structural causal model.

[0014] Preferably, counterfactual intervention is performed based on the structural causal model. By shielding the error propagation effect of the temperature drift baseline on the nonlinear topological distortion manifold, the target state features corresponding to the preset standard temperature reference environment are reconstructed. This includes: performing operator intervention in the structural causal model to lock the value of the intermediate interference node to a constant calibration reference value and cut off the variable propagation relationship of the first directed edge; replacing the real-time acquired temperature drift baseline input with the constant calibration reference value in the propagation path of the second directed edge; combining the real-time acquired relay reactance evolution features as joint input variables, performing forward inference calculation through the second structural mapping function, and outputting a pure topological feature vector stripped of the influence of the intermediate interference node, and using the pure topological feature vector as the target state features corresponding to the preset standard temperature reference environment.

[0015] Preferably, the method of extracting an absolute physical degradation index based on the target state characteristics for fault risk prediction includes: calculating the Mahalanobis distance between the target state characteristics and the cluster center of a preset failure state distribution cluster; performing exponential smoothing filtering on the Mahalanobis distance to generate the absolute physical degradation index; and triggering a fault risk warning command when the absolute physical degradation index is less than a preset safety margin value, and generating a prediction report containing the expected remaining number of safe disconnections.

[0016] A relay operation state fault risk prediction system, comprising: Equivalent reactance extraction module: acquires the voltage waveform across the relay and the synchronous pulse width modulation command sequence, extracts parasitic harmonics with a frequency that is a preset multiple of the fundamental frequency of the pulse width modulation command sequence from the voltage waveform, and uses them as measurement probes; calculates the equivalent reactance evolution characteristics of the relay based on the transfer function dispersion matrix generated by the signal passing through the input and output terminals of the relay by the measurement probes; Topological distortion identification module: Collects the temperature drift baseline of the sensing circuit and the mechanical vibration vector of the relay, and combines them with the equivalent reactance evolution characteristics as state variables; maps the state variables to the Poincaré phase space to reconstruct the trajectory attractor; and extracts the nonlinear topological distortion manifold by calculating the maximum Lyapunov exponential jump of the trajectory attractor. Degradation Index Assessment Module: Constructs a structural causal model covering ambient temperature, the temperature drift baseline, and the nonlinear topological distortion manifold; performs counterfactual intervention based on the structural causal model, and reconstructs the target state characteristics corresponding to the preset standard temperature reference environment by shielding the error transmission effect of the temperature drift baseline on the nonlinear topological distortion manifold; and extracts the absolute physical degradation index based on the target state characteristics to predict failure risk.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention extracts parasitic harmonics synchronously generated by the electric drive system as a measurement probe to calculate the equivalent reactance evolution characteristics of the relay. This design effectively utilizes the inherent high-frequency electrical signals in the vehicle's high-voltage network, eliminating the need for dedicated signal generation and injection hardware in the underlying architecture. While optimizing system hardware costs and reducing physical intrusion, it leverages the physical penetration characteristics of high-frequency signals to more sensitively capture early signs of degradation in the microstructure of the relay contact surface, effectively enhancing the ability to detect hidden physical damage.

[0018] 2. This invention maps multidimensional electrical and environmental state variables to a Poincaré phase space and extracts the topologically distorted manifold by calculating the maximum Lyapunov exponential jump of the trajectory attractor. This achieves a shift in feature extraction methods towards the nonlinear dynamic topological dimension. The monitoring mechanism based on abrupt changes in topological geometry can better adapt to the slow drift and fluctuation of the reference signal and exhibits high sensitivity to transient nonlinear chaotic phenomena caused by contact micro-bounces, thus providing a more ample window for early intervention and warning for defense systems.

[0019] 3. This invention constructs a structural causal model and introduces counterfactual intervention logic to shield the error propagation effect of temperature drift baseline on topologically distorted manifolds. This algorithmic architecture enables the system to perform variable decoupling and feature cleaning at the data computation level, allowing for the targeted removal of spurious interference components caused by the thermal drift of the measuring equipment itself from complex signals coupled with multiple physics fields. The reconstructed physical degradation index more objectively maps the true health evolution trajectory of the relay itself, effectively improving the accuracy and reliability of vehicle condition assessment and predictive decision-making under complex thermal alternating conditions. Attached Figure Description

[0020] Figure 1 A flowchart of a relay operation state fault risk prediction method provided in an embodiment of the present invention; Figure 2 A structural diagram of a relay operation status fault risk prediction system provided in an embodiment of the present invention; Figure 3 A flowchart of counterfactual intervention provided for embodiments of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figures 1 to 3 This invention provides a method for predicting fault risks in relay operation, the technical solution of which is as follows: A method for predicting fault risks in relay operation states includes: The voltage waveform across the relay and the synchronous pulse width modulation command sequence are obtained. Parasitic harmonics with a frequency that is a preset multiple of the fundamental frequency of the pulse width modulation command sequence are extracted from the voltage waveform and used as a measurement probe. The equivalent reactance evolution characteristics of the relay are calculated based on the transfer function dispersion matrix generated by the signal passing through the input and output terminals of the relay by the measurement probe. The temperature drift baseline of the sensing circuit and the mechanical vibration vector of the relay are collected and combined with the equivalent reactance evolution characteristics as state variables; the state variables are mapped to the Poincaré phase space to reconstruct the trajectory attractor; and the nonlinear topological distortion manifold is extracted by calculating the maximum Lyapunov exponential jump of the trajectory attractor. A structural causal model is constructed, encompassing ambient temperature, the temperature drift baseline, and the nonlinear topological distortion manifold. Based on the structural causal model, counterfactual intervention is performed. By shielding the error transmission effect of the temperature drift baseline on the nonlinear topological distortion manifold, the target state characteristics corresponding to a preset standard temperature reference environment are reconstructed. Based on the target state characteristics, an absolute physical degradation index is extracted for fault risk prediction.

[0023] Example 1: This embodiment is primarily applied to high-voltage, high-power DC fast charging scenarios for new energy vehicles operating at 800 volts and above. Under such long-cycle on-board conditions, the battery management system not only faces severe temperature drift of the operational amplifiers in the sensing circuit (systematic baseline distortion), but also the relay contacts are highly susceptible to micro-bounce arcing (transient high-frequency interference) under high current surges and road bumps. When faced with this complex operating condition, traditional timing algorithms often mask the true arc avalanche signal with the false baseline raised by temperature drift, leading to system misjudgment or missed detection.

[0024] As one embodiment of the present invention, refer to Figure 1 A flowchart of a relay operation state fault risk prediction method, referring to... Figure 2 A structural diagram of a relay operation status fault risk prediction system, referring to... Figure 3 Counterfactual intervention flowchart.

[0025] Further, extracting parasitic harmonics with frequencies equal to a preset multiple of the fundamental frequency of the pulse width modulation command sequence as a measurement probe includes: extracting the fundamental frequency period timestamp of the pulse width modulation command sequence using a synchronous phase-locked loop; synchronously truncating the voltage waveform according to the fundamental frequency period timestamp to obtain an independent period voltage signal sequence; performing a discrete Fourier transform on the independent period voltage signal sequence to obtain a frequency domain amplitude-phase matrix; extracting spectral components in the frequency domain amplitude-phase matrix whose frequency values ​​are equal to the fundamental frequency multiplied by a preset multiple integer; reconstructing the time-domain parasitic harmonic signal by performing an inverse Fourier transform on the spectral components, and using the parasitic harmonic signal as the measurement probe.

[0026] Specifically, during the acceleration or fast-charging handshake phase of an electric vehicle, the underlying control unit synchronously records waveforms at an extremely high sampling rate (e.g., 1 MHz). First, it retrieves the pulse width modulation (PWM) log from within the inverter, locks the current reference switching frequency (typically between 10 kHz and 30 kHz) using digital synchronous phase-locked loop (PLL) technology, and extracts its period timestamp. Then, using this timestamp as a cursor, it truncates the continuously acquired broadband voltage waveform by an integer cycle. A fast discrete Fourier transform (FFT) is performed on the truncated discrete-time sequence, converting the time-domain waveform into a frequency-domain matrix containing amplitude and phase information. Within this matrix, a bandpass filtering algorithm is used to selectively extract high-order parasitic harmonic spectral components with a frequency center value between 15 and 21 times the fundamental frequency. Finally, by applying an inverse Fourier transform with a Hanning window, this high-frequency component is restored to a continuous time-domain oscillation signal. This clean signal is then identified as the high-frequency measurement probe used to penetrate the microscopic gaps of the relay.

[0027] This invention utilizes the inherent high-frequency switching noise of the electric drive system as a detection signal, eliminating the cost and system intrusion risk associated with adding high-frequency pulse injection hardware as in traditional solutions. By extracting parasitic harmonics at specific multiples, interference from low-frequency mechanical noise and fundamental current is effectively avoided, improving the signal-to-noise ratio of the initial detection data.

[0028] Further, based on the transfer function dispersion matrix generated by the measurement probe passing through the relay input and output terminals, the equivalent reactance evolution characteristics of the relay are calculated, including: calculating the cross power spectral density and self power spectral density of the signals from the measurement probe at the relay input and output terminals respectively; to extract the evolution characteristics, the transfer function dispersion vectors of multiple consecutive observation periods within a preset time sliding window are extracted and stacked column-wise according to the time sequence to construct a two-dimensional transfer function dispersion matrix; the transfer function dispersion matrix is ​​decomposed into a real impedance matrix and an imaginary reactance matrix; principal component analysis or singular value decomposition is performed on the imaginary reactance matrix to extract the first N principal component eigenvalue sequences of the imaginary reactance matrix within the preset time sliding window, and the eigenvalue sequences are used as the equivalent reactance evolution characteristics of the relay.

[0029] Specifically, after confirming the measurement probe, the first voltage sequence at the relay's input terminal and the second voltage sequence at the output terminal are captured synchronously in real time. First, the self-power spectral density of these two time sequences is calculated in the frequency domain to quantify the energy distribution of the signal itself; simultaneously, the cross-power spectral density is calculated to quantify the phase deflection and energy attenuation relationship between the input and output signals at different frequency points. Then, the cross-power spectral density data is divided by the self-power spectral density data at the input terminal to derive a complex form of the transfer function dispersion vector. To reduce transient random errors in a single measurement and to construct a feature analysis space, the transfer function dispersion vectors of multiple observation periods are continuously extracted within a 50-millisecond sliding time window and stacked according to the time sequence to construct a two-dimensional transfer function dispersion matrix. This matrix objectively characterizes the evolution trend of the relay's filtering characteristics for high-frequency electromagnetic waves under the current physical state. Furthermore, this complex matrix is ​​decoupled, separating the real part matrix (characterizing the Joule heating loss of the macroscopic contact resistance) and the imaginary part matrix (characterizing the equivalent capacitive and inductive reactance formed by the microscopic contact gap). Furthermore, principal component analysis is used to reduce the dimensionality of the two-dimensional imaginary reactance matrix and extract the top three principal component eigenvalues ​​of the imaginary reactance matrix. This sequence of eigenvalues ​​constitutes the equivalent reactance feature characterizing the evolution of the micromorphology of the contact surface.

[0030] This invention overcomes the limitation of traditional DC voltage drop methods, which can only assess macroscopic closed states. By utilizing the imaginary reactance characteristics of the high-frequency transfer function, it can keenly detect sub-millimeter-level roughening and air gap changes on the contact surface caused by minute arc erosion, significantly improving the depth of perception of the very early, implicit physical degradation of the relay.

[0031] Before generating the transfer function dispersion matrix by dividing the cross power spectral density by the self power spectral density, the process further includes: during the vehicle manufacturing or high-voltage power-on preheating stage, sending open-circuit and short-circuit calibration commands to the high-voltage wiring harness, measuring and generating a wiring harness scattering parameter matrix characterizing the inherent parasitic parameters of the high-voltage wiring harness; the generation of the transfer function dispersion matrix specifically includes: calculating the initial transfer function dispersion matrix, and performing inverse cascade calculations on the initial transfer function dispersion matrix and the wiring harness scattering parameter matrix to remove the dynamic parasitic impedance fluctuation component generated by the wiring harness heating with temperature, thereby generating the target transfer function dispersion matrix.

[0032] Specifically, considering that long-distance high-voltage wiring harnesses in vehicles generate significant heat during 800-volt high-current fast charging, their parasitic inductance and capacitance to ground will dynamically drift. During the high-voltage preheating phase of vehicle startup, the system uses the pre-charge circuit of the battery management system to send a broadband sweep frequency test signal to the high-voltage wiring harness. Response signals are collected under two extreme states: the relay is physically disconnected (open circuit) and the pre-charge is closed (short circuit). Based on these signals, a wiring harness scattering parameter matrix characterizing the intrinsic attenuation and phase deflection of the wiring harness is established. During real-time vehicle operation, after calculating the initial transfer function dispersion matrix containing the overall response of the wiring harness and relay, microwave network de-embedding technology is used to perform an inverse matrix division operation between this initial matrix and the pre-measured wiring harness scattering parameter matrix. Through this mathematical cascaded decoupling, the system completely removes the parasitic impedance variables generated by the wiring harness, and the final extracted target transfer function dispersion matrix represents only the pure electrical response of the physical contacts inside the relay.

[0033] This invention effectively eliminates the interference of dynamic parasitic impedance fluctuations of external high-voltage wiring harnesses under high-power heating conditions on microscopic reactance characteristics by introducing network de-embedding technology. This method filters out measurement reference distortion caused by the thermal effect of external wiring harnesses at the signal source, prevents the weak degradation changes of the relay body from being masked by macroscopic wiring harness noise, and effectively improves the objectivity and purity of high-frequency electrical feature extraction.

[0034] Before concatenating the temperature drift baseline, mechanical vibration vector, and equivalent reactance evolution characteristics into a multidimensional time-series state vector, the process further includes: obtaining the physical geometric distance from the relay contact to the vibration sensor and the fixed acoustic wave velocity of the transmission medium, and calculating the acoustic propagation delay constant; taking the abrupt change edge of the relay electromagnetic coil driving current as the absolute physical time zero point, rigidly feeding forward translating the mechanical vibration vector on the time axis according to the acoustic propagation delay constant; and using a multiphase filtering resampling algorithm to synchronously increase the low-frequency sampling rate of the translated mechanical vibration vector to the same high-frequency sampling rate as the equivalent reactance evolution characteristics, thereby generating a physically causally aligned target mechanical vibration vector.

[0035] Specifically, considering that the propagation speed of mechanical sound waves in a metal casing is much lower than the transmission speed of electrical signals, resulting in a millisecond-level physical time delay misalignment, and that the sampling rate of mechanical vibration sensors is typically much lower than the extraction frequency of electrical features, the system first abandons the nonlinear time stretching algorithm that destroys the original physical characteristics of the waveform. The system retrieves the rising edge of the drive current at the instant the underlying hardware triggers the relay's action and locks it as the absolute physical time zero point of the multi-physics field. Subsequently, the system reads the preset three-dimensional coordinates of the sensor installation and the coordinates of the relay contacts, calculates the linear physical geometric distance between them, and divides this distance by the fixed acoustic wave velocity of the relay base material (such as nylon or metal) to calculate a definite acoustic propagation time delay constant. Based on this time delay constant, the algorithm performs overall rigid feedforward translation compensation on the time axis of the low-frequency acquired mechanical vibration sequence, eliminating the medium propagation time delay without changing the original oscillation period and envelope shape of the waveform. Next, to match the data dimension of the electrical signals, the system introduces a multiphase filtering resampling algorithm. This algorithm performs linear interpolation and anti-aliasing low-pass filtering on the translated mechanical vibration vector, undistortingly increasing its sampling point density to a high-frequency reference that perfectly matches the evolution characteristics of the equivalent reactance. Finally, all state variables are strictly aligned at the same physical timestamp and concatenated into a multidimensional time-series state vector while maintaining their original physical properties.

[0036] This invention strictly follows the objective causal laws of electromechanical multiphysics. By introducing a rigid feedforward translation and multiphase filtering resampling mechanism based on acoustic constants, it eliminates the propagation delay of heterogeneous physical signal media and unifies data dimensions, while fully preserving the true physical duration and frequency domain energy characteristics of mechanical vibration waveforms. This effectively avoids distortion of the original physical characteristics caused by excessive mathematical fitting, and truly ensures the objectivity and scientific nature of subsequent Poincaré phase space multidimensional topological structure mapping.

[0037] Further, mapping the state variables to the Poincaré phase space to reconstruct the trajectory attractor includes: eliminating time delays and aligning the temperature drift baseline, the mechanical vibration vector, and the equivalent reactance evolution characteristics based on the absolute physical time reference of relay action, then normalizing them and splicing them to generate a multidimensional time-series state vector; calculating the delay time parameter of the multidimensional time-series state vector using the mutual information method; calculating the embedding dimension parameter of the multidimensional time-series state vector using the spurious nearest neighbor method; performing time delay coordinate mapping on the multidimensional time-series state vector based on the delay time parameter fixed during system initialization and the embedding dimension parameter to generate a phase space coordinate point set; and using the phase space coordinate point set to fit and generate a closed trajectory attractor in the Poincaré phase space.

[0038] Specifically, after obtaining the equivalent reactance characteristics, the operational amplifier drift voltage baseline (typically fluctuating between 0 and 5 millivolts) fed back from the battery management system's temperature measurement node is simultaneously retrieved, along with the triaxial vibration numerical sequence output by the high-frequency accelerometer (e.g., amplitude range of 10g to 50g). To eliminate dimensional differences, the minimax normalization method is used to scale the three types of heterogeneous data to a unified numerical range, and they are strictly aligned and stitched together into a multidimensional state vector according to timestamps. Based on this, a nonlinear dynamics algorithm is introduced. First, the optimal delay time step (e.g., set to 5 milliseconds) for the chaotic sequence is determined by calculating the first minimum point of the average mutual information. Then, the spurious nearest neighbor method is used to calculate the dimension value when the abrupt change ratio of the distance between neighboring points is lower than the set tolerance, to determine the optimal embedding dimension (e.g., 4-dimensional or 5-dimensional). During the online monitoring phase of the actual vehicle, the optimal delay time step and the optimal embedding dimension will be implemented as fixed constants to ensure that the topological feature vector subsequently fed into the neural network remains absolutely constant in tensor shape. Based on the two fixed key parameters mentioned above, the one-dimensional time-series vector is delayed and expanded, mapping it to a high-dimensional Poincaré geometric phase space. As the vehicle runs, these discrete coordinate points gradually interweave and connect in the phase space, eventually fitting a closed shape with a specific geometric topology, namely the trajectory attractor.

[0039] This invention changes the traditional early warning algorithm's reliance on the absolute amplitude of the signal, transforming the evolution of multi-physics variables into geometric changes in a high-dimensional space. This mapping mechanism has a natural tolerance for the slow thermal drift of sensors, effectively preventing the direct interference of systematic baseline distortion on the feature extraction process.

[0040] Further, by calculating the maximum Lyapunov exponent jump of the trajectory attractor, the nonlinear topological distortion manifold is extracted, including: selecting an initial reference point on the trajectory attractor and calculating the initial neighboring point with the smallest Euclidean distance to the initial reference point; calculating the difference vector of the state variable within a time sliding window and solving for the L2 norm of the difference vector as a one-dimensional time-domain energy gradient; when the one-dimensional time-domain energy gradient exceeds a preset wake-up threshold, triggering fixed-point data truncation to obtain a local time window containing feature mutations; performing local phase space reconstruction to generate a local trajectory attractor only for the state variable within the local time window, and calculating the short-time maximum Lyapunov exponent of the local trajectory attractor; when the short-time maximum Lyapunov exponent is greater than a preset chaotic bifurcation threshold, taking the phase space coordinate point set region corresponding to the local time window as the nonlinear topological distortion manifold.

[0041] Specifically, considering the computational constraints of the microcontroller in the vehicle battery management system, the system employs a two-layer heterogeneous dynamics evaluation mechanism to extract the topological distortion manifold. In the low-computational-power pre-detection stage, the algorithm temporarily suspends the energy-intensive phase space reconstruction operation, instead utilizing only extremely low system computational power to continuously calculate the one-dimensional time-domain energy gradient of state variables (such as equivalent reactance characteristics) within a real-time sliding time window (e.g., calculating the sum of squared differences of the first-order differences between adjacent sampling points), and uses this as a low-computational-power pre-detection cursor. During normal smooth temperature drift or conventional mechanical aging, this cursor value is extremely small, and the system's high-dimensional matrix operation unit is in a dormant state; however, once a relay experiences a micro-bounce and triggers a transient arc avalanche, the pre-detection cursor instantly exceeds the preset wake-up threshold. Upon detecting this threshold-crossing signal, the system immediately triggers a fixed-point data truncation command, precisely locking and truncating a local time window of 20 milliseconds before and after the abrupt change node. Subsequently, the system activates the high-dimensional core layer, performing high-dimensional Poincaré phase space reconstruction only on the data slice within this extremely short local time window to generate a local trajectory attractor. It then tracks the short-time evolution trajectories of the initial point and its neighbors, rapidly solving for the short-time maximum Lyapunov exponent of this local attractor. If this short-time maximum Lyapunov exponent exceeds a preset chaotic bifurcation threshold, the occurrence of an anomalous nonlinear mutation is rigorously confirmed from a dynamic perspective. The corresponding phase space coordinate point set region is then stripped and encapsulated into a specific topological distortion manifold matrix for in-depth diagnosis of the downstream causal network.

[0042] This invention introduces a two-layer event-driven triggering mechanism based on a one-dimensional energy gradient, enabling low-latency and low-computational-power quantitative assessment of "chaotic mutation" phenomena. It can not only accurately extract the real damage characteristics caused by brief electric arcs from extremely noisy interference backgrounds, but also exponentially compress the system's computational overhead, perfectly matching the hardware computing power ceiling of vehicle-mounted edge computing nodes, and effectively ensuring the real-time response capability and engineering application feasibility of the fault diagnosis and defense system under extreme working conditions.

[0043] Furthermore, a structural causal model encompassing ambient temperature, the temperature drift baseline, and the nonlinear topological distortion manifold is constructed, including: constructing a causal directed graph with the ambient temperature as the root node, the temperature drift baseline as intermediate disturbance nodes, and the nonlinear topological distortion manifold as observation result nodes; introducing a thermal inertial delay hidden node characterizing heat conduction delay between the root node and the intermediate disturbance nodes; using an equivalent thermal network conduction model, calculating the real-time virtual junction temperature of the operational amplifier inside the sensing circuit based on the ambient temperature, and assigning the real-time virtual junction temperature to the thermal inertial delay hidden node; and utilizing historical ambient temperature... Using temperature data and historical sensing circuit calibration data, a first structural mapping function is trained with real-time virtual junction temperature as the independent variable and temperature drift baseline as the dependent variable. A second structural mapping function is fitted using a multilayer perceptron network, with temperature drift baseline and preset ideal relay characteristics as joint input variables and nonlinear topological distortion manifold as output variables. The structural causal model is generated by combining a causal directed graph, the first structural mapping function, and the second structural mapping function. The path from the root node to the intermediate interference node is defined as the first directed edge, and the path from the intermediate interference node to the observation result node is defined as the second directed edge.

[0044] The equivalent heat network conduction model is constructed based on a first-order thermodynamic resistance-capacitance system. Its physical logic is defined as follows: the rate of change of the operational amplifier junction temperature is directly proportional to the temperature difference between the ambient temperature and the junction temperature, and inversely proportional to the product of thermal resistance and thermal capacity.

[0045] In the digital implementation process, to perform real-time calculations within the microcontroller, the aforementioned continuous differential logic is discretized. Its iterative calculation rule is described as follows: the current real-time virtual junction temperature equals the previous real-time virtual junction temperature plus a temperature correction increment. This temperature correction increment is defined as: the difference between the current ambient temperature and the previous real-time virtual junction temperature multiplied by a discretization coefficient. The discretization coefficient is defined as the sampling period divided by the product of thermal resistance and thermal capacity.

[0046] Regarding the establishment of model parameters, the product of thermal resistance and heat capacity (i.e., the thermal time constant) is obtained through discrete step response experiments. Specifically, the relay sensing circuit is placed in a controlled constant temperature chamber. After a 10-degree Celsius step change in ambient temperature, the time taken for the temperature drift baseline to reach its final stable value of 63.2% is recorded. This time value is defined as the product of thermal resistance and heat capacity. By averaging the results of 3 to 5 sets of step tests in different temperature ranges, the core transfer parameters of the model under different operating conditions can be established.

[0047] Specifically, a directed acyclic graph structure containing hidden nodes is established in the software architecture. The measured in-vehicle ambient temperature is defined as the source root node of the causal network; the temperature drift baseline of the sensing link is defined as an intermediate interference node that is affected by the environment and contaminates the measurement results; and the nonlinear topological distortion manifold extracted in the steps is defined as the final observation result node. Considering the significant thermal resistance and thermal capacity characteristics of the sealed metal casing of the battery management system and the printed circuit board itself, changes in ambient temperature will not be instantly reflected as temperature drift inside the chip. Therefore, the algorithm inserts a thermal inertia delay hidden node between the source root node and the intermediate interference node. Information transmission edges are constructed from the root node to the hidden node, from the hidden node to the interference node, and from the interference node to the observation result node. During the parameter learning phase, the algorithm extracts the vehicle thermal management bench test data from the past 12 months. First, it uses a first-order capacitor-resistance equivalent thermal network conduction model, taking the measured external ambient temperature sequence as the heat source input. After damping filtering by thermal resistance and thermal capacitance, it calculates the real-time virtual junction temperature experienced by the operational amplifier silicon chip. Subsequently, it trains a polynomial regression function as the first structure mapping function to quantify the baseline drift millivolts caused by each 1 degree Celsius increase in the real-time virtual junction temperature.

[0048] Simultaneously, a multilayer perceptron network with three hidden layers is used, with temperature drift baseline data and ideal relay reactance simulation data (i.e., a pure equivalent reactance reference sequence pre-calibrated and measured on an ideal test bench at 25 degrees Celsius with constant temperature and humidity and no contactless physical ablation) as joint inputs. Considering that the nonlinear topological distortion manifold is essentially a set of three-dimensional coordinate points in a high-dimensional Poincaré phase space, direct output would cause the network to fail to converge. Therefore, it is first reduced to a fixed-length one-dimensional topological feature vector using spatial flattening or an autoencoder, and then this one-dimensional topological feature vector is used as the output for fitting training. During the training phase, mean squared error combined with L2 regularization penalty is used as the joint loss function, and ReLU activation function is used in the hidden layer nodes to prevent gradient vanishing. The Adam optimization algorithm is used to continuously update the network weight matrix through backpropagation until the loss function value is lower than the preset convergence threshold, and finally convergence is obtained to obtain the second structural mapping function. The above nodes, directed edges, and two mapping functions together form a structural causal model reflecting the objective physical influence logic between variables.

[0049] This invention compensates for the heat transfer delay error between the external environment and internal devices by introducing a thermally inertial delay hidden node that conforms to thermodynamic laws into the causal transmission link. This not only provides a rigorous mathematical structure to support subsequent active variable intervention, but also effectively prevents feature stripping failure caused by intervention lag when the ambient temperature undergoes a drastic step change (such as a vehicle leaving a cold storage), significantly enhancing the interpretability and generalization ability of the model under extreme thermal alternation conditions.

[0050] Furthermore, based on the structural causal model, counterfactual intervention is performed to reconstruct the target state features corresponding to a preset standard temperature reference environment by shielding the error propagation effect of the temperature drift baseline on the nonlinear topological distortion manifold. This includes: performing operator intervention in the structural causal model to lock the value of the intermediate interference node to a constant calibration reference value and severing the variable propagation relationship of the first directed edge; replacing the real-time acquired temperature drift baseline input data with the locked constant calibration reference value in the propagation path of the second directed edge; combining the real-time acquired relay reactance evolution characteristics as joint input variables, performing forward inference calculation through the second structural mapping function, and outputting a pure topological feature vector stripped of the influence of intermediate interference nodes. This pure topological feature vector is then used as the target state features corresponding to the preset standard temperature reference environment. Specifically, the preset standard temperature reference environment refers to an ideal physical state with no temperature drift, where the sensing circuit is at a constant temperature of 25 degrees Celsius and the operational amplifier temperature drift baseline is calibrated to 0 millivolts.

[0051] Specifically, in the real-time prediction phase, the interference quantizer within the structural causal model is activated. The algorithm does not passively receive the current sensor readings but instead forcibly executes a logic blocking operation: in the memory processing unit, the value representing the intermediate interference node is forcibly overwritten and locked to the calibration zero-point value (i.e., 0 mV drift) under a 25°C standard reference environment, thus mathematically severing the variable fluctuation transmission caused by the first directed edge (ambient temperature). Subsequently, the algorithm forcibly replaces the distorted temperature drift baseline data actually collected on the bus with this locked zero-point calibration value, and feeds the replaced baseline data and the real-time extracted equivalent reactance evolution characteristics into the second structural mapping function with fixed parameters. Through forward propagation of the network model, inference calculations are performed directly to calculate the answer to a counterfactual question: "If the current sensing circuit had not experienced any temperature drift, what would the topological features that should have been measured look like?" Through this virtual intervention-based stripping and reconstruction, the algorithm completely eliminates the spurious deformation gain caused by measurement link distortion in the composite signal, ultimately outputting a set of pure target state feature vectors free from any sensor error contamination.

[0052] This invention applies intervention logic to the field of vehicle electrical measurement, resolving the ambiguity in state assessment caused by the self-drift of sensing circuits masking microscopic arc features. By cutting off the causal path of environmental interference, it achieves surgical-like cleaning of the original distorted signal, ensuring the absolute physical objectivity of the feature extraction results.

[0053] Further, based on the target state characteristics, an absolute physical degradation index is extracted for fault risk prediction, including: calculating the Mahalanobis distance between the target state characteristics and the cluster center of the preset failure state distribution cluster; performing exponential smoothing filtering on the Mahalanobis distance to generate the absolute physical degradation index; when the absolute physical degradation index is less than a preset safety margin value, a fault risk warning instruction is triggered, and a prediction report containing the expected remaining number of safe disconnections is generated.

[0054] Specifically, the reconstructed pure target state features are input into a pre-defined Mahalanobis distance discrimination space. This space stores the covariance matrix and cluster centers of the distribution clusters of catastrophic contact welding failure states of relays, established based on destructive physical analysis experiments. The algorithm calculates the spatial covariance-weighted distance between the current target feature vector and the failure cluster center, i.e., the Mahalanobis distance. To eliminate occasional numerical fluctuations during the calculation process, an exponentially weighted moving average filtering algorithm with a smoothing coefficient of 0.1 is used to numerically smooth the continuously output Mahalanobis distances. The smoothed continuous values ​​are then established as the absolute physical degradation index. The system's main control unit presets a safety margin alarm threshold (e.g., an interval 20% away from the danger boundary). When the real-time updated absolute physical degradation index decreases and falls below this safety margin threshold, a high-level fault risk warning message is immediately sent to the vehicle controller via the controller area network bus. Simultaneously, the descent slope is calculated based on historical action frequencies, and a predictive diagnostic report containing the estimated remaining safe disconnection count of the relay is synchronously output on the vehicle screen or cloud backend.

[0055] This invention successfully transforms highly abstract causal topological vectors into engineering metrics that engineers and vehicle control systems can directly utilize. Distance measurement based on clean features avoids false triggers or missed alarms caused by data contamination, effectively guiding the rational planning of maintenance cycles and improving the operational reliability of the vehicle throughout its entire lifecycle.

[0056] Example 2: This embodiment provides a relay operation status fault risk prediction system. In terms of hardware physical architecture, this system can be implemented based on the underlying main control microprocessor and related peripheral high-frequency analog-to-digital converter circuits of the vehicle battery management system, or it can be deployed as an independent intelligent diagnostic computing unit within the vehicle's high-voltage power distribution box. Logically, the system aims to solve the problem of relay degradation characteristics being masked by system distortion under complex operating conditions by integrating data-driven and physical logic approaches. This system mainly consists of three core functional units: an equivalent reactance extraction module, a topology distortion identification module, and a degradation index evaluation module. These modules are interconnected through an internal high-speed data bus to achieve continuous data flow and feature coupling. The specific operation flow and functional implementation methods of each module in this system will be described in detail below.

[0057] As one embodiment of the present invention, refer to Figure 1 A flowchart of a relay operation state fault risk prediction method, referring to... Figure 2 A structural diagram of a relay operation status fault risk prediction system, referring to... Figure 3 Counterfactual intervention flowchart.

[0058] First, the system uses an equivalent reactance extraction module to perceive and calculate the underlying high-frequency physical characteristics. This module is the data sensing front-end of the entire prediction system. During system operation, the module's hardware data acquisition interface receives continuous voltage waveforms from the high-voltage input and output terminals of the relay in real time, and simultaneously receives pulse width modulation (PWM) command sequences from the underlying communication bus of the electric drive controller. The module's built-in digital signal processing unit first initiates a synchronous phase-locked loop (PLL) mechanism to lock the reference switching frequency of the current PWM command sequence and generates an alignment timestamp. Subsequently, the module truncates the voltage waveform to an integer period according to this timestamp and calls a fast Fourier transform (FFT) algorithm to convert the time-domain signal into a frequency-domain feature matrix. In this matrix, the extraction module is configured with bandpass filtering logic to specifically extract spectral components whose frequency values ​​are equal to integer multiples of the fundamental frequency of the command sequence (such as higher harmonic bands). Through inverse Fourier transform, the system restores this specific high-frequency component to a time-domain oscillation waveform, formally establishing it as a measurement probe that penetrates the microscopic gaps of the relay contacts.

[0059] After acquiring the measurement probe, the matrix solving unit within the module further extracts the first state sequence of the probe on the relay input side and the second state sequence on the relay output side. The module calculates the self-power spectral density and cross-power spectral density of both, and generates a transfer function dispersion vector characterizing the high-frequency physical response of the relay body through a division operation. Then, multiple consecutive dispersion vectors are stacked sequentially within a preset sliding window to construct a two-dimensional transfer function dispersion matrix. Finally, the module performs matrix decomposition and principal component eigenvalue extraction on the imaginary part of this complex matrix to generate the equivalent reactance evolution characteristics of the relay, and sends them to the next-level processing module. The operating logic of this module completely eliminates the need for additional high-frequency generation hardware, realizing zero-invasive micro-damage detection.

[0060] Secondly, the system uses a topological distortion recognition module to perform dimensionality upscaling and distortion capture from one-dimensional time-series signals to high-dimensional spatial features. This module is primarily responsible for accurately identifying transient abnormal degradation behavior under severe baseline drift interference. The data fusion interface of this module receives in parallel the equivalent reactance evolution characteristics output from the previous module, as well as temperature drift baseline data from the sensing circuit and a three-axis mechanical vibration vector from the shell acceleration sensor. To eliminate dimensional differences and propagation delays in heterogeneous data, this module first performs rigid feedforward translation compensation on the mechanical vibration vector based on a preset acoustic propagation delay constant, and then calls a multiphase filtering resampling algorithm to synchronously increase its low-frequency sampling rate to a high-frequency reference consistent with the equivalent reactance evolution characteristics. Subsequently, normalization processing is performed to concatenate the above three types of physical quantities to generate a unified, physically causally aligned multidimensional time-series state vector.

[0061] During the feature extraction stage, the module invokes the built-in nonlinear dynamics algorithm library to calculate the delay time parameter and embedding dimension parameter of the multidimensional temporal state vector. Based on these two parameters, the module configures a two-layer heterogeneous dynamics evaluation mechanism. In the low-computational-power pre-detection stage at the bottom layer, the module's high-dimensional computing unit is dormant, continuously calculating only the one-dimensional temporal energy gradient of the state vector within the time sliding window as a pre-detection cursor. When the pre-detection cursor momentarily crosses the preset wake-up threshold due to relay micro-bounce, the module immediately triggers a fixed-point data truncation command, locking the local time window containing feature mutations. Subsequently, the module wakes up the high-dimensional kernel layer, and only for the data slice within this extremely short local time window, folds and maps the one-dimensional vector to the high-dimensional Poincaré phase space to reconstruct the local trajectory attractor, and quickly solves for the short-time maximum Lyapunov exponent of the local attractor. When the short-time exponent is greater than the preset chaotic bifurcation threshold, the module rigorously confirms the occurrence of the abnormal mutation from a dynamic perspective, and extracts the phase space coordinate point set corresponding to the local time node, encapsulates it into a nonlinear topological distortion manifold, and outputs it to the subsequent module as the core observation feature.

[0062] Finally, the system performs causal logic cleaning through the degradation index evaluation module and outputs the final prediction decision. This module is the brain of the entire prediction system, responsible for removing false interference and quantifying real risks. Internally, the module initializes a structural causal model engine, which maintains a directed acyclic physical logic graph in memory. This graph explicitly defines the hierarchical relationships between system variables: the measured in-vehicle ambient temperature is configured as the root node, which is not controlled by the system; the temperature drift baseline of the sensing link is configured as an intermediate interference node affected by the environment and contaminating the measurement channel; and the nonlinear topological distortion manifold passed from the previous module is configured as the final observation result node. The engine internally embeds a first structural mapping function (quantifying the physical impact of ambient temperature on baseline drift) and a second structural mapping function (quantifying the joint impact of baseline drift and real degradation on the final dimensionality-reduced topological feature vector), trained with massive amounts of historical data. The second structural mapping function converges and is fixed based on a loss function with a regularization penalty term.

[0063] During the real-time evaluation phase, the module does not passively fit the data but actively performs counterfactual intervention. The module's main control logic sends operator intervention commands to the structural causal model, forcibly rewriting and locking the values ​​of the "intermediate interference nodes" in the algorithm's memory to a preset zero-point calibration reference value. This completely shields and cuts off the causal path of error transmission from ambient temperature to the temperature drift baseline at the logical level. Next, the module uses this locked constant reference value to replace the contaminated drift data actually collected on the bus and inputs it into the second structural mapping function for inverse differentiation. Through this "counterfactual" logical reconstruction, the module deduces from the mixed and distorted signal the target state characteristics that the sensing circuit should exhibit when it is in a perfectly isothermal state, i.e., the pure characteristics corresponding to a preset standard temperature reference environment (such as 25 degrees Celsius and 0 millivolt temperature drift). Finally, the module projects this pure target state characteristic onto a preset failure boundary hyperplane to calculate the Mahalanobis distance. After smoothing and filtering, it outputs an absolute physical degradation index with unique physical causality. When the index exceeds the preset safety threshold, the module immediately issues a fault risk warning command via the bus.

[0064] In summary, this prediction system, through the tight integration of three modules, sequentially completes the extraction of high-frequency parasitic probes, the reconstruction of multidimensional chaotic topology, and the removal of counterfactual causal errors. The system's data flow logic is clear, and its physical coupling is tight. Without increasing the burden on onboard hardware, it achieves reliable assessment and proactive intervention of the true health status of relays under complex and compound interference conditions.

[0065] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for predicting fault risks in relay operation, characterized in that, include: The voltage waveform across the relay and the synchronous pulse width modulation command sequence are obtained. Parasitic harmonics with a frequency that is a preset multiple of the fundamental frequency of the pulse width modulation command sequence are extracted from the voltage waveform and used as a measurement probe. The equivalent reactance evolution characteristics of the relay are calculated based on the transfer function dispersion matrix generated by the signal passing through the input and output terminals of the relay by the measurement probe. The temperature drift baseline of the sensing circuit and the mechanical vibration vector of the relay are collected and combined with the equivalent reactance evolution characteristics as state variables; the state variables are mapped to the Poincaré phase space to reconstruct the trajectory attractor; and the nonlinear topological distortion manifold is extracted by calculating the maximum Lyapunov exponential jump of the trajectory attractor. A structural causal model is constructed, encompassing ambient temperature, the temperature drift baseline, and the nonlinear topological distortion manifold. Based on the structural causal model, counterfactual intervention is performed. By shielding the error transmission effect of the temperature drift baseline on the nonlinear topological distortion manifold, the target state characteristics corresponding to a preset standard temperature reference environment are reconstructed. Based on the target state characteristics, an absolute physical degradation index is extracted for fault risk prediction.

2. The method for predicting relay operating state fault risk according to claim 1, characterized in that, Extracting parasitic harmonics with frequencies equal to a preset multiple of the fundamental frequency of a pulse width modulation (PWM) command sequence as a measurement probe includes: extracting the fundamental frequency period of the PWM command sequence using a synchronous phase-locked loop; truncating the voltage waveform according to the fundamental frequency period to obtain a voltage signal sequence with independent periods; performing a discrete Fourier transform on the voltage signal sequence to obtain a frequency domain amplitude-phase matrix; extracting spectral components from the frequency domain amplitude-phase matrix whose frequency values ​​are equal to the fundamental frequency multiplied by a preset multiple; reconstructing the time-domain parasitic harmonic signal by performing an inverse Fourier transform on the spectral components; and using the parasitic harmonic signal as the measurement probe.

3. The method for predicting relay operating state fault risk according to claim 1, characterized in that, The equivalent reactance evolution characteristics of the relay are calculated by passing through the transfer function dispersion matrix generated by the signals at the input and output terminals of the relay. This includes: calculating the ratio of the cross power spectral density and the self power spectral density of the signals at the input and output terminals of the relay by the measurement probe, generating a transfer function dispersion vector characterizing the physical response of the relay; continuously extracting the transfer function dispersion vectors of multiple independent observation periods within a preset time sliding window and arranging and splicing them in time sequence to construct a transfer function dispersion matrix; decomposing the transfer function dispersion matrix into a real impedance matrix and an imaginary reactance matrix; extracting the eigenvalue sequence of the first preset number of principal elements of the imaginary reactance matrix, and using the eigenvalue sequence as the equivalent reactance evolution characteristics of the relay.

4. The method for predicting relay operating state fault risk according to claim 1, characterized in that, Mapping the state variables to the Poincaré phase space to reconstruct the trajectory attractor includes: aligning and normalizing the temperature drift baseline, the mechanical vibration vector, and the equivalent reactance evolution characteristics using an absolute physical time reference, and concatenating them into a multidimensional temporal state vector; calculating the delay time parameter of the multidimensional temporal state vector using the mutual information method; calculating the embedding dimension parameter of the multidimensional temporal state vector using the spurious nearest neighbor method; performing time delay coordinate mapping on the multidimensional temporal state vector based on the delay time parameter and embedding dimension parameter fixed during system initialization to generate a phase space coordinate point set; and using the phase space coordinate point set to fit and generate a closed trajectory attractor in the Poincaré phase space.

5. The method for predicting relay operating state fault risk according to claim 1, characterized in that, Extracting a nonlinear topologically distorted manifold by calculating the maximum Lyapunov exponent jump of the trajectory attractor includes: selecting an initial reference point on the trajectory attractor and calculating the initial neighboring point with the smallest Euclidean distance to the initial reference point; calculating the L2 norm of the difference vector of adjacent sampling points of the state variable within a time sliding window as a one-dimensional time-domain energy gradient; when the one-dimensional time-domain energy gradient exceeds a preset wake-up threshold, truncating a local time window containing a feature mutation; performing local phase space reconstruction on the state variables within the local time window to generate a local trajectory attractor, and calculating the short-time maximum Lyapunov exponent of the local trajectory attractor; when the short-time maximum Lyapunov exponent is greater than a preset chaotic bifurcation threshold, using the phase space coordinate point set region corresponding to the local time window as the nonlinear topologically distorted manifold.

6. The method for predicting relay operating state fault risk according to claim 1, characterized in that, A structural causal model encompassing ambient temperature, the temperature drift baseline, and the nonlinear topologically distorted manifold is constructed, including: constructing a causal directed graph with the ambient temperature as the root node, the temperature drift baseline as intermediate disturbance nodes, and the nonlinear topologically distorted manifold as the observation result node; defining the path from the root node to the intermediate disturbance node as a first directed edge, and the path from the intermediate disturbance node to the observation result node as a second directed edge; introducing a thermal inertial delay hidden node representing heat conduction delay between the root node and the intermediate disturbance node, and assigning the real-time virtual junction temperature calculated using an equivalent thermal network conduction model to the thermal inertial delay hidden node; training and outputting a first structural mapping function of the temperature drift baseline based on historical data, and fitting and outputting a second structural mapping function of the nonlinear topologically distorted manifold using a multilayer perceptron network; and combining the causal directed graph, the first structural mapping function, and the second structural mapping function to generate the structural causal model.

7. The method for predicting relay operating state fault risk according to claim 1, characterized in that, Based on the structural causal model, counterfactual intervention is performed to reconstruct the target state features corresponding to a preset standard temperature reference environment by shielding the error propagation effect of the temperature drift baseline on the nonlinear topological distortion manifold. This includes: performing operator intervention in the structural causal model to lock the value of the intermediate interference node to a constant calibration reference value and severing the variable propagation relationship of the first directed edge; replacing the real-time acquired temperature drift baseline input with the constant calibration reference value in the propagation path of the second directed edge; combining the real-time acquired relay reactance evolution features as joint input variables, performing forward inference calculation through the second structural mapping function, and outputting a pure topological feature vector stripped of the influence of intermediate interference nodes, using the pure topological feature vector as the target state features corresponding to the preset standard temperature reference environment.

8. The method for predicting relay operating state fault risk according to claim 1, characterized in that, Based on the target state characteristics, an absolute physical degradation index is extracted for fault risk prediction, including: calculating the Mahalanobis distance between the target state characteristics and the cluster center of a preset failure state distribution cluster; performing exponential smoothing filtering on the Mahalanobis distance to generate the absolute physical degradation index; when the absolute physical degradation index is less than a preset safety margin value, a fault risk warning command is triggered, and a prediction report containing the expected remaining number of safe interruptions is generated.

9. A relay operation state fault risk prediction system, characterized in that, include: Equivalent reactance extraction module: acquires the voltage waveform across the relay and the synchronous pulse width modulation command sequence, and extracts parasitic harmonics from the voltage waveform at a frequency that is a preset multiple of the fundamental frequency of the pulse width modulation command sequence, as a measurement probe; The equivalent reactance evolution characteristics of the relay are calculated based on the transfer function dispersion matrix generated by the signal passing through the input and output terminals of the relay by the measuring probe. Topological distortion identification module: Collects the temperature drift baseline of the sensing circuit and the mechanical vibration vector of the relay, and combines them with the equivalent reactance evolution characteristics as state variables; maps the state variables to the Poincaré phase space to reconstruct the trajectory attractor; and extracts the nonlinear topological distortion manifold by calculating the maximum Lyapunov exponential jump of the trajectory attractor. Degradation Index Assessment Module: Constructs a structural causal model covering ambient temperature, the temperature drift baseline, and the nonlinear topological distortion manifold; performs counterfactual intervention based on the structural causal model, and reconstructs the target state characteristics corresponding to the preset standard temperature reference environment by shielding the error transmission effect of the temperature drift baseline on the nonlinear topological distortion manifold; and extracts the absolute physical degradation index based on the target state characteristics to predict failure risk.

Images