Test and Evaluation Method and System for Multidimensional Coupling Characteristics of Skateboard Chassis Decoupling Interface

Abstract

This invention relates to the field of intelligent electric vehicle chassis testing and health management technology, specifically to a method and system for testing and evaluating the multi-dimensional coupling characteristics of a skateboard chassis decoupling interface. The method involves establishing a normal decoupling fingerprint benchmark; simultaneously applying three types of active excitations under normal load conditions to extract three types of features; inputting these features into a three-layer heterogeneous graph network, iteratively pruning low-contribution connections to converge and obtain the dominant coupling transmission link and the contribution ratio of each path segment; then prioritizing the application of minor active physical interventions at the cross-domain, layer-crossing locations with the highest contribution to identify coupling weaknesses; finally, using contribution as weights to fuse frequency domain, time domain, and power system decay components to calculate the interface's comprehensive health index, and feeding back to drive the incremental update of the normal decoupling fingerprint benchmark. This technical solution achieves accurate evaluation of the health status of the skateboard chassis decoupling interface under multi-physics coupling, accurate location of coupling weaknesses, and adaptive evolution of the benchmark.

Description

Technical Field

[0001] This invention relates to the field of intelligent electric vehicle chassis testing and health management technology, specifically to a method and system for testing and evaluating the multi-dimensional coupling characteristics of a skateboard chassis decoupling interface. Background Technology

[0002] The skateboard chassis standardizes and decouples the upper body and chassis at the physical, electrical, and communication levels, integrating mechanical connections, high-voltage power distribution, low-voltage signals, and real-time communication through a unified decoupling interface. This interface is simultaneously subjected to the coupling effects of multiple physical fields, including mechanical vibration, electrical power surges, thermal stress, and communication loads, during vehicle operation. Its health directly determines the safety and reliability of the entire vehicle. With the increasing reliance on decoupling interfaces in drive-by-wire chassis and autonomous driving, the coupling behavior of the interface under multi-physical field excitation exhibits highly nonlinear and cross-domain transmission characteristics: mechanical fretting wear can cause contact resistance drift, which in turn alters the heat flux distribution through Joule heating and may lead to eye diagram closure in the communication physical layer; further, timing errors can trigger abnormal control commands. Once this multidimensional coupling evolves from a controllable decoupling state to coupling failure, it often appears in a hidden and sudden form. Traditional alarm methods based on single physical quantity thresholds are difficult to provide early warnings in the early stages of failure, and cannot reveal the coupling transmission relationship between different physical domains, let alone locate the key weak path points that lead to coupling failure. As a result, in practical applications, only periodic tightening or replacement of parts can be used for maintenance, which is costly and cannot avoid random failures.

[0003] In existing technologies, the evaluation of interface status is mostly limited to the independent monitoring of bolt preload torque, contact resistance, or communication bit error rate, lacking means to jointly decouple coupling features from multi-physics field signals. While some methods introduce multi-sensor fusion, they only perform simple data layer stitching, failing to analyze deeper coupling patterns such as coupling resonance, coupling chaos, and shared latent variables in the feature space. Furthermore, existing diagnostic models are mostly static thresholds or offline-trained classifiers with fixed benchmarks, unable to adaptively update as interface performance deteriorates and the environment changes, leading to a decline in the reliability of evaluation results over time. Regarding the location of coupling weaknesses, traditional methods heavily rely on manual inspection or identifying weak areas based on experience, lacking an optimization and verification mechanism that can penetrate different physical domains from the source of excitation, accurately quantify the contribution of each coupling path, and automatically converge to the critical link based on contribution. Summary of the Invention

[0004] To overcome the problems of existing technologies, such as difficulty in multi-dimensional analysis of decoupling interface coupling characteristics, inability to locate weak coupling paths, and lack of adaptive benchmark updates, this invention provides a method and system for testing and evaluating the multi-dimensional coupling characteristics of skateboard chassis decoupling interfaces. This system enables accurate evaluation of the health status of skateboard chassis decoupling interfaces under multi-physics field coupling, accurate location of weak coupling points, and adaptive evolution of the benchmark.

[0005] The technical solution of this application specifically includes: According to one aspect of this application, a method for testing and evaluating the multi-dimensional coupling characteristics of a skateboard chassis decoupling interface is provided, comprising: During the early operation phase of the interface, mechanical strain, contact resistance, communication clock deviation and heat flux signals are collected. The decoupling invariant characteristics are obtained by analyzing the slow characteristics of each signal, and a normal decoupling fingerprint benchmark is constructed. When the interface load state is consistent with the normal decoupling fingerprint benchmark, random vibration, pulse power jump and communication broadcast storm are applied simultaneously to collect multiple field responses; three analysis channels are used to extract the deviation of the normal decoupling fingerprint benchmark from different dimensions. The extracted features are three types of features: coupling resonance features, coupling chaos features and coupling latent variable features. Three types of features are input into a three-layer heterogeneous graph network. The mechanical-electrical layer connection weights are initialized with resonance features, the state evolution of each layer is constrained by chaotic features, and the cross-layer message passing is driven by latent variable features. In the network iteration, each round evaluates the contribution of the connection to the resonance and latent variable features by the mutual information of cross-layer nodes, removes low-contribution connections and feeds the weight graph back to the next round, until the cross-domain edge weights converge, and outputs the dominant coupling link and the contribution ratio of each path to the coupling failure. A small amount of physical intervention is applied at the cross-domain layer crossing point with the highest contribution ratio in the dominant coupling link. Three types of features before and after the intervention are extracted and progressively verified: if the chaotic feature regresses the baseline attractor morphology, the resonance feature shows a reduction in the coherence value of a specific frequency band and the direction of the latent variable approaches the baseline center, then it is confirmed as a weak point; otherwise, the location that failed to verify is negatively fed into the graph network to readjust the weights and find the optimal solution until a consistent weak point is confirmed by the joint verification of the three features. After identifying the weak points and links, three types of features and contribution ratios of each physical quantity on the chain are extracted. The decay components of the benchmark in the frequency domain, time domain and dynamic system dimensions are calculated by weighting the contribution ratios and merging them into an interface comprehensive health index. Based on this, the normal decoupling fingerprint benchmark is incrementally updated, forming a complete link from benchmark reference, deviation analysis, path tracing, active verification to correction.

[0006] As a further option of the method of the present invention, the decoupling invariant features obtained from the slow feature analysis of each signal include: Slow feature analysis was performed on mechanical strain, contact resistance, communication clock skew, and heat flux signals respectively. The slow feature analysis aimed to minimize the square mean of the time derivative of the output features. With the objective and under the zero-mean constraint Unit variance constraint and decorrelation constraints with extracted features The following solution yields the slow characteristic components of each physical quantity that change most gradually and are uncorrelated, among which... For the first The first physical quantity One output slow feature component; Components whose statistical stability satisfies the condition that their coefficient of variation is below a preset threshold and their autocorrelation decay time constant is greater than a preset duration threshold are selected from all slow feature components and used as decoupling invariant features.

[0007] As a further option of the method of the present invention, the normal decoupling fingerprint benchmark construction includes: Construct a decoupling-invariant eigenvector from the decoupling-invariant features. Calculate its mean vector during the early operational phase. Covariance Matrix ; Mahalanobis distance Based on, take satisfaction The equiprobable ellipsoids serve as the boundary envelope, where For degrees of freedom The chi-square distribution at confidence level The upper quantile below, This represents the total number of invariant feature vector samples collected during the early operational phase; this forms the normal decoupling fingerprint benchmark, and deviations beyond the boundary envelope indicate a deviation in the interface state.

[0008] As a further option of the method of the present invention, the coupled resonance feature extraction method is as follows: Representative mechanical strain and contact resistance signals from the excitation process are selected, and continuous wavelet transforms are performed to obtain time-frequency complex coefficients. and ; Calculate the force-electric time-frequency coherent locking value in the time-frequency plane. ;in for The complex conjugate; The peak amplitude and peak frequency of the time-averaged coherence spectrum are extracted from the preset key frequency band set to form the coupled resonance characteristics.

[0009] As a further option of the method of the present invention, the coupled chaotic feature extraction method is as follows: The joint response vector of mechanical strain, contact resistance and communication clock deviation during the excitation process is mapped to three-dimensional phase space to reconstruct the attractor orbit; Calculate the attractor correlation dimension Quantify fractal characteristics and estimate the maximum Lyapunov exponent. To quantify initial value-sensitive dependencies, where For time step, and These represent the initial distance and the evolved distance between the reference point and its nearest neighbor, respectively. Number of reference points; Will and The combination results in coupled chaotic features.

[0010] As a further option of the method of the present invention, the method for extracting coupled latent variable features is as follows: A time-frequency autoencoder is constructed, which takes the time-frequency representation of the multi-physics response as input, compresses it into low-dimensional latent variables through the encoder, and then reconstructs it through the decoder. The training loss includes the reconstruction error and the maximum mean difference between the latent variable distribution and the prior distribution. In the latent space, based on the mutual information of each dimension and each physical quantity signal, the latent variables are separated into independent factors that are significantly associated with a single physical domain and shared generative factors that are simultaneously associated with two or more physical domains. The independent factors and shared generative factors are spliced ​​together to form coupled latent variable features, so as to express the cross-domain coupled generative structure in the low-dimensional manifold.

[0011] As a further option of the method of the present invention, in the three-layer heterogeneous graph network: The cross-layer connection weight between mechanical and electrical layer nodes is initialized by the time-frequency coherence locking value of the corresponding frequency band in the coupling resonance characteristics; The coupled chaotic features are converted into global coupling indicators, and an upper limit constraint is imposed on the evolution range of the hidden states of each layer of nodes. The dominant receiving direction of cross-layer message passing is determined by the drift direction of the shared generating factor in the latent space of the coupled latent variable features. The hidden state of a node is updated through intra-layer attention aggregation and gating fusion of cross-layer messages.

[0012] As a further option of the method of the present invention, the contribution of connections to resonance and latent variable features is evaluated by cross-layer node attention mutual information in each round, and low-contribution connections are pruned, including: For cross-layer edges Calculate the mutual information contribution of attention ,in To score attention, and For nodes and Hidden state For mutual information, This is the set of all cross-layer edges; The adaptive pruning threshold is obtained by multiplying the median contribution of all current cross-layer edges by a preset ratio coefficient. Cross-layer edges with a contribution below the threshold are permanently removed, forcing subsequent iterations to focus on regions with high coupling strength.

[0013] As a further option of the method of the present invention, the output dominant coupling link and the contribution ratio of each path segment to the coupling failure include: In the graph network after pruning and convergence, candidate paths are enumerated, with the node corresponding to the starting point of the stimulus as the source node and the node corresponding to the response observation point as the sink node. Path-based coupling strength , The contribution of mutual information of attention to cross-layer edge e, To select a candidate path from the source node to the sink node, The largest one serves as the dominant coupling and transmission link; The contribution ratio of each path segment on the link is determined as follows: , For the first The contribution of cross-layer edges corresponding to segment paths. This represents the total number of road segments.

[0014] As a further option of the method of the present invention, the step of applying a small amount of physical intervention at the cross-domain layer crossing point with the highest contribution ratio in the dominant coupling link, and simultaneously extracting three types of features before and after the intervention, including: The cross-domain layer-crossing path segment with the largest contribution ratio is selected, and the intervention method is selected according to the physical domain type to which it belongs: for mechanical connection surfaces, a pre-tightening torque is applied for fine adjustment; for electrical terminals, a nanometer-level contact surface reversal is applied; for communication transceivers, an additional clock bias not exceeding the preset ratio of the protocol jitter tolerance is introduced. During the intervention, the full physical field response of the interface was collected synchronously, and coupled resonance features, coupled chaotic features, and coupled latent variable features were extracted in parallel from the data segments before and after the intervention.

[0015] As a further option of the method of the present invention, the judgment condition for the morphology of the chaotic feature regression benchmark attractor is: Let the baseline coupled chaotic characteristics be: The chaotic characteristics before and after intervention are respectively and When satisfied At that time, it is determined that the coupled chaotic features regress to the normal state.

[0016] As a further option of the method of the present invention, the judgment condition for the reduction of the coherence value of the specific frequency band of the resonance feature is as follows: In the key frequency band associated with the path segment that contributes the highest proportion in the dominant coupling transmission link, the peak value of time-frequency coherence lock-in after intervention is smaller than that before intervention, i.e. ,in, In the frequency band before intervention Peak value of time-frequency coherent lock value, For intervention in the frequency band The peak value of the time-frequency coherent lock value.

[0017] As a further option of the method of the present invention, the judgment condition for the direction of the hidden variable to approach the reference center is: Let the baseline coupling latent variable characteristics be: The characteristics of coupled latent variables before and after intervention are as follows: and When satisfied When the time is right, the direction of the latent variable is determined to shift towards the reference center.

[0018] As a further option of the method of the present invention, the interface-based comprehensive health index acquisition step includes: according to Calculate the frequency domain decay component, according to Calculate the time-domain decay component, according to Calculate the decay component of the dynamic system, where , , Path segments The current coupled resonance features, coupled latent variable features, and coupled chaotic features at the corresponding positions. , , For the corresponding benchmark value; The overall health index of the interface is ; , and These are the fusion weight coefficients for the frequency domain, time domain, and dynamic system dimensions, respectively.

[0019] Another aspect of this application provides a multi-dimensional coupling characteristic testing and evaluation system for the decoupling interface of a skateboard chassis, the system comprising: The benchmark construction module is used to collect mechanical strain, contact resistance, communication clock deviation and heat flux signals in the early operation phase of the interface, analyze the slow characteristics of each signal to obtain decoupling invariant characteristics, and construct a normal decoupling fingerprint benchmark. The multi-dimensional feature extraction module is used to simultaneously apply random vibration, pulse power jump and communication broadcast storm when the interface load state is consistent with the normal decoupling fingerprint benchmark, and collect multi-field responses. It uses three analysis channels to extract the deviance of the normal decoupling fingerprint benchmark from different dimensions. The extracted features are three types of features: coupling resonance features, coupling chaos features and coupling latent variable features. The link optimization module is used to input three types of features into a three-layer heterogeneous graph network. The mechanical-electrical layer connection weights are initialized with resonance features, the state evolution of each layer is constrained by chaotic features, and the cross-layer message passing is driven by latent variable features. In the network iteration, each round evaluates the contribution of the connection to the resonance and latent variable features by the mutual information of cross-layer nodes, removes low-contribution connections and feeds back the weight graph to the next round, until the cross-domain edge weights converge, and outputs the dominant coupling link and the contribution ratio of each path segment to the coupling failure. The weak point identification module is used to apply a small amount of physical intervention at the cross-domain layer crossing point with the highest contribution ratio in the dominant coupling link, and simultaneously extract three types of features before and after the intervention and progressively verify them: if the chaotic feature regresses the baseline attractor shape, the resonance feature shows a reduction in the coherence value of a specific frequency band and the direction of the latent variable approaches the baseline center, then it is identified as a weak point; otherwise, the corresponding position is injected into the graph network to readjust the weights and find the optimal solution until a consistent weak point is jointly confirmed by the three features. The health assessment and correction module is used to identify weak points and links, extract three types of characteristics and contribution ratios of each physical quantity on the chain, and calculate the decay components of the benchmark in the frequency domain, time domain and dynamic system dimensions by weighting the contribution ratios. These components are then integrated into an interface comprehensive health index, which is used to incrementally update the normal decoupled fingerprint benchmark, forming a complete link from benchmark reference, deviation analysis, path tracing, active verification to correction.

[0020] The beneficial effects of this application are as follows: This invention breaks through the limitations of traditional single-physical-quantity threshold alarms and simple data layer splicing. Based on the invariance of slow features in multi-physics fields and cross-domain decoupling theory, it deeply analyzes the hidden coupling patterns from multiple dimensions, including frequency domain resonance, dynamic system chaos, and latent space generation factors, improving the completeness of multi-dimensional coupling feature analysis by more than 40%. Through a closed-loop mechanism of iterative pruning of heterogeneous graph network attention mutual information and progressive consistency verification by active physical intervention, it achieves quantitative optimization and precise weak point location of the dominant coupling links penetrating each physical domain from the source of excitation, completely replacing manual experience-based investigation. The accuracy of weak point location reaches more than 95%, advancing the early warning time of hidden cross-domain coupling failure by more than 30%. At the same time, based on the comprehensive health index driven by the fusion of multi-dimensional decay components, it drives the incremental update of the normal decoupling fingerprint benchmark, endowing the evaluation system with adaptive evolution and anti-drift capabilities. It overcomes the defect of static benchmarks declining credibility due to service environment degradation, reducing the credibility decay of long-term evaluation by more than 50%, thereby realizing a complete link from benchmark reference, deviation analysis, path tracing, active verification to benchmark correction. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall control method; Figure 2 Flowchart of control method S100; Figure 3 Flowchart of control method S200; Figure 4 Flowchart of control method S300; Figure 5 Flowchart of control method S400; Figure 6 Flowchart of control method S500; Figure 7 A bar chart showing the accuracy of locating weak points in coupling; Figure 8 A comparative bar chart showing the average advance warning time for locating weak points in coupling; Figure 9 This is a trend chart of the comprehensive health index of the left-side interface of the front axle. Detailed Implementation

[0022] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0023] The theoretical basis of this invention is built upon three pillars: the theory of slow feature invariance of multiphysics, the theory of multidimensional decoupling of cross-domain coupling features, and the theory of dominant link penetration optimization in heterogeneous graph networks.

[0024] The definitions of the core variables and the derivation of the formulas are as follows: 1. Objective function for slow feature analysis: Let the first... The first physical quantity The slow feature components are: Slow feature analysis extracts the slowest-changing components by minimizing the squared mean of the feature's time derivative. The objective function is... The constraints are , And for any satisfy The decoupled invariant eigenvector composed of the slow characteristics of all physical quantities is denoted as... .

[0025] 2. Standard Decoupled Fingerprint Baseline Boundary Criterion: Decoupled Invariant Feature Vector To the normal center The Mahalanobis distance is defined as ,in Let be the covariance matrix. The normal boundary is an equiprobable ellipsoid. ,when When the interface status deviates significantly.

[0026] 3. Time-frequency coherent lock value: Assume the mechanical strain signal... With contact resistance signal The time-frequency coefficients after wavelet transform are respectively and The time-frequency coherent locking value is defined as Extracting coupled resonance feature vectors from the peak values ​​of a specified frequency band set. .

[0027] 4. Coupled Chaotic Feature Extraction: Reconstructing the joint response of multiple fields into a three-dimensional phase space vector sequence. Attractor correlation dimension Depend on Estimate, where the correlation integral The maximum Lyapunov index was generated by It is estimated that the two constitute coupled chaotic features. .

[0028] 5. Attention Mutual Information Contribution and Path Optimization in Heterogeneous Graph Networks: In a three-layer heterogeneous graph network, cross-layer edges... The contribution of attention mutual information is defined as ,in To score attention, To hide the node For mutual information. The dominant coupling transmission link obtained after convergence pruning. The path-to-path coupling strength is The contribution ratio of each segment is as follows: .

[0029] 6. Progressive consistency verification condition: The coupled chaotic feature regression before and after intervention is determined as follows: ; specific frequency bands before and after intervention The time-frequency coherent peak reduction is determined as follows: Features of coupled latent variables before and after intervention Towards the normal benchmark center Directional migration determination is When all three conditions are met simultaneously, a weak point in the coupling is identified.

[0030] 7. Interface Overall Health Index: Contribution of each path segment on the dominant link Under weighting, frequency domain fading component Time-domain decay component Powertrain degradation component Comprehensive health index , This is the fusion coefficient.

[0031] The specific embodiments of the present invention will be described in detail below.

[0032] Example 1: Please see Figure 1 The method for testing and evaluating the multi-dimensional coupling characteristics of a skateboard chassis decoupling interface provided in this embodiment includes: S100: In the early operation phase of the interface, it synchronously collects signals from multiple physical fields such as mechanical, electrical, communication and thermal fields, extracts decoupling invariant features through slow feature analysis, and jointly constructs a normal decoupling fingerprint benchmark.

[0033] S200: Under normal load conditions, three types of active excitations are applied: mechanical vibration, electrical power jump, and communication broadcast storm. Coupled resonance features, coupled chaos features, and coupled hidden variable features are extracted by three parallel analysis channels.

[0034] S300: Input three sets of coupling features into a three-layer heterogeneous graph network, and optimize the output of cross-domain dominant coupling transmission links and the proportion of coupling failure contribution of each path segment through attention mutual information iterative pruning.

[0035] S400: Apply a small amount of physical intervention to the cross-domain location with the highest contribution, and confirm the weak coupling point by progressively verifying whether the three sets of features before and after the intervention simultaneously return to the normal benchmark. If it fails, it iterates with negative feedback until it is consistent.

[0036] S500: It integrates the decline components of each dimension with the contribution ratio as the weight to form an interface comprehensive health index, and updates the normal decoupled fingerprint benchmark incrementally based on this, realizing the transition from benchmark reference to self-correction.

[0037] The specific plan is as follows: Please refer to Figure 2 , Figure 2 The detailed process of stage S100 in this embodiment is shown, which includes stages S110 to S140.

[0038] The S100 continuously acquires multi-physics field signals during the early operation phase of the interface, extracts decoupling invariant features using slow feature analysis, and constructs a multi-dimensional normal decoupling fingerprint benchmark.

[0039] S110: In the mechanical connection area of ​​the decoupling interface, strain gauges are arranged in a preset array at the critical load-bearing positions on the flange face; a micro-resistance measurement unit is connected in series at the contact terminals of the high-voltage electrical connector; a clock deviation monitoring circuit is deployed on the clock synchronization pin of the communication transceiver; and thin-film thermopile is arranged on the inner and outer surfaces of the interface housing to collect the heat flux through the interface.

[0040] In one possible implementation of this embodiment, the synchronous continuous acquisition of multiphysics field signals includes: The mechanical strain acquisition method involves connecting strain gauges in a full-bridge differential manner, arranging measuring points in multiple quadrants around the interface to simultaneously acquire tangential and normal orthogonal strain components, and covering the nonlinear response frequency band caused by mechanical loosening to obtain multi-channel mechanical strain data.

[0041] The contact resistance acquisition method involves applying a constant small current excitation that does not cause a significant temperature rise using the Kelvin connection method. Measurement points are arranged on the high voltage positive and negative terminal pairs and the communication reference ground loop. The sampling frequency is set according to the dynamic process of electrical contact to acquire multi-channel contact resistance data.

[0042] The communication clock deviation acquisition method is to continuously calculate the deviation between the local clock and the master clock by using a high-precision timestamp capture unit with time synchronization protocol messages as a reference, and the sampling frequency is synchronized with the message sending frequency.

[0043] The heat flux acquisition method involves a calibrated thin-film thermopile combined with a built-in temperature sensor for cold junction compensation. After signal amplification, analog-to-digital conversion is performed at a sampling frequency adapted to the thermal process changes.

[0044] All sensor channels are absolutely time-aligned using a unified time synchronization source before data acquisition, and the synchronization error between channels is controlled within a preset range. Continuous acquisition yields a set of raw multiphysics signals. ,in For the first Road mechanical strain signals, For the first Road contact resistance signal, This is a communication clock offset signal. For the first Road heat flux signal.

[0045] S120: Perform slow feature analysis on each physical quantity signal acquired by S110, extract multiple feature components from the input signal with the rate of change from slow to fast, among which the slowest changing component reflects the most stable and unchanging structural information hidden in the signal.

[0046] In one possible implementation of this embodiment, slow feature analysis includes: For the One-dimensional continuous-time signal of a physical quantity Perform data preprocessing. Use a fixed time window length. The continuous signal is segmented into non-overlapping data segments. The time window length is determined based on the period of the highest effective frequency component in the signal, ensuring that each analysis window contains sufficiently complete dynamic information. The mean and standard deviation are calculated for each data segment within the window, and the data are standardized by normalizing the mean to zero and the standard deviation to one. All standardized data segments are then concatenated in chronological order to construct the input matrix for slow feature analysis.

[0047] The objective function of slow feature analysis is to minimize the time rate of change of each output feature. For the th... The first physical quantity Each output slow feature component The objective function is The constraints imposed when solving this optimization problem include: zero-mean constraint. Unit variance constraint And remove related constraints, for all satisfy .

[0048] The analytical solution for slow feature analysis is obtained through generalized eigenvalue decomposition. The first-order time derivative of the standardized signal is calculated, and the derivative covariance matrix is ​​constructed. Simultaneously, the covariance matrix of the signal itself is calculated. The generalized eigenvalue problem is solved, and the eigenvalues ​​are sorted in ascending order; smaller eigenvalues ​​indicate slower changes in the corresponding feature component. Slow feature components are then extracted sequentially from the eigenvectors corresponding to the eigenvalues ​​in ascending order.

[0049] Slow feature analysis was performed on all channels of the mechanical strain signal, and the slowest changing components were retained for each channel; the slowest changing components were retained for each channel of the contact resistance signal; the slowest changing components were retained for each channel of the communication clock deviation signal; and the slowest changing components were retained for each channel of the heat flux signal. All retained components constituted the set of slow feature components.

[0050] S130: From all the slow feature components extracted in S120, further screen out the decoupling invariant feature components that have stable characterization ability for the decoupling state, and calculate the statistical stability index and time series persistence index for each slow feature component.

[0051] In one possible implementation of this embodiment, the selection and verification of the decoupling invariant feature includes: For each slow feature component Within multiple consecutive time windows during the early operational phase, the mean and standard deviation are calculated for each window, forming a mean series and a standard deviation series, respectively. The coefficient of variation of the mean series is then calculated. Coefficient of variation of the standard deviation series Preset statistical stability threshold ,when and When the slow feature component is determined to be statistically stable in time, it is retained as a decoupling invariant feature component.

[0052] The autocorrelation decay time constant is further determined for the retained decoupling-invariant eigencomponents, i.e., the autocorrelation function decreases from its maximum value to... The duration of the time elapsed. When the autocorrelation decay time constant is greater than the preset minimum duration threshold, the decoupling invariant feature component is determined to have sufficient temporal persistence and is ultimately included in the construction of the normal decoupling fingerprint benchmark. All decoupling invariant feature components that have passed the above two verifications are combined to form the decoupling invariant feature vector. ,in This represents the total dimension of the decoupled invariant features that are ultimately retained.

[0053] S140: Utilizing the decoupling-invariant eigenvector output from S130 Construct the central shape and boundary envelope of a normal decoupled fingerprint benchmark in a multidimensional space.

[0054] In one possible implementation of this embodiment, the process of constructing a normal decoupled fingerprint benchmark includes: During the continuous sampling time span in the early operation phase, the decoupling invariant eigenvector is calculated. mean vector along the time dimension ,in This represents the total number of decoupled invariant feature vector samples collected during the early operational phase. Mean vector. The central shape of the normal decoupling fingerprint benchmark in multidimensional space is defined, which represents the average coupling mode of the interface under healthy decoupling state, where the decoupling characteristics of each physical domain remain unchanged.

[0055] Calculate the covariance matrix of the decoupled invariant eigenvectors The covariance matrix describes the correlation structure among the decoupling-invariant eigencomponents and reflects the joint distribution characteristics of decoupling-invariant eigencomponents from different physical domains under healthy conditions.

[0056] The boundary envelope is defined as an equiprobable ellipsoid based on Mahalanobis distance, and the interface currently uses decoupled invariant feature vectors. The Mahalanobis distance is expressed as The normal boundary is ,in For degrees of freedom The chi-square distribution at confidence level The upper quantile below. The confidence level is determined based on the interface security level. The Mahalanobis distance of the decoupled invariant feature vectors acquired in real time. Exceeding When the interface decoupling state deviates significantly from the specified state, it is determined that a statistically significant deviation has occurred.

[0057] Mean vector Covariance matrix Confidence level and decoupling invariant feature dimension Encapsulated and stored together, forming a normal decoupled fingerprint benchmark. This serves as a unified reference for subsequent deviation analysis in the S200 stage and regression verification in the S400 stage.

[0058] Please refer to Figure 3 , Figure 3 The detailed process of stage S200 in this embodiment is shown, which includes stages S210 to S250.

[0059] Under normal load conditions, the S200 synchronously applies three types of active excitations and extracts coupled resonance features, coupled chaotic features, and coupled latent variable features through three parallel analysis channels.

[0060] S210: Under the premise that the interface Mahalanobis distance does not exceed the normal boundary, the device synchronously outputs gradually enhanced random vibration, duty cycle jump pulse power and periodic communication broadcast storm by excitation.

[0061] In one possible implementation of this embodiment, the synchronous application process of the three types of active stimuli includes: The first type of excitation is a gradually increasing random vibration, used to induce a nonlinear response to mechanical connection loosening. The vibration excitation is applied via an electromagnetic vibrator mounted on the mechanical connector of the decoupling interface, which is rigidly fastened to the interface flange face via a clamp. The excitation signal is band-limited white noise with a bandwidth covering the modal frequency range of the interface structure. The power spectral density starts from a low initial value and gradually increases with a preset increment. During the enhancement process, the acceleration response at the interface is continuously monitored. Enhancement stops when the response magnitude reaches a certain proportion of the allowable vibration magnitude of the interface design. The entire enhancement process lasts for a sufficient duration to ensure sufficient excitation of various mechanical nonlinear characteristics.

[0062] The second type of excitation is a duty cycle-jumping pulsed power injection, used to induce recoverable fluctuations in contact resistance. The pulsed power injection is achieved through a programmable electronic load connected in parallel to the high-voltage DC bus, which draws current from the bus in pulse mode. The pulse period is fixed, and the duty cycle cycles through multiple values ​​according to a preset jumping sequence, maintaining a sufficient number of pulse cycles for each duty cycle. The pulse current amplitude is set to a small proportion of the interface's rated current to control the temperature rise caused by Joule heating within a safe range.

[0063] The third type of incentive is a periodic communication broadcast storm, used to disrupt the timing order of time-triggered protocols. The communication broadcast storm is generated by test nodes connected to the interface communication bus. These test nodes send high-priority broadcast frames at progressively increasing rates, starting at a low percentage of the bus's nominal bandwidth and increasing at a preset rate, with the maximum sending rate not exceeding half the bus's nominal bandwidth. The data field of the broadcast frame is filled with random data, and the message identifier is set to a priority higher than that of normal control messages.

[0064] The three types of excitations are initiated synchronously under the control of a unified external trigger signal, with the synchronization error controlled within a preset range. During the excitation application, all sensor channels deployed in S110 maintain continuous acquisition, synchronously recording the dynamic response of each physical quantity under the joint excitation of the multi-physics field.

[0065] S220: The first analysis channel targets the mechanical strain signal and contact resistance signal synchronously acquired during the S210 excitation process, and extracts the nonlinear coupling resonance characteristics between force and electrical signal through time-frequency coherent analysis.

[0066] In one possible implementation of this embodiment, the extraction process of coupled resonance features includes: From the mechanical strain signals collected during the excitation period, the strain channel signal closest to the electrical connector was selected as the representative mechanical response. Meanwhile, the positive contact resistance signal was selected as a representative electrical response. Continuous wavelet transforms were performed on both signals. The complex Morlet wavelet was used as the mother wavelet, and the wavelet parameters were set according to the time-frequency resolution requirements. This yielded the time-frequency complex coefficient matrix of the mechanical strain signal. The time-frequency complex coefficient matrix of the contact resistance signal Both have the same time and frequency resolution.

[0067] Calculate the time-frequency coherent lock value on the time-frequency plane. ,in for The complex conjugate, This represents the Gaussian smoothed average within the local time-frequency neighborhood. The width of the smoothing window in both time and frequency directions is set according to the wavelet support interval. The range of the time-frequency coherence lock value is... The higher the value, the stronger the phase-locked relationship between the two signals at that time and frequency point, and the more significant the coupling interaction between the force domain and the electric domain at that frequency.

[0068] Each vibration excitation enhancement step corresponds to a steady-state vibration segment as an analysis unit. Within each analysis unit, the time-averaged coherence lock value along the frequency axis is calculated to obtain the time-averaged coherence spectrum. Within a preset set of key frequency bands, the peak amplitude and peak frequency of the time-averaged coherence spectrum are extracted. The key frequency bands are determined based on the interface structure modes, electrical resonant frequencies, and historical experience.

[0069] The coherent peak amplitude and frequency of each key frequency band are combined into a coupled resonance eigenvector. This serves as the output of the first analysis channel.

[0070] S230: The second analysis channel targets the multi-physics response signals synchronously acquired during the S210 excitation process. It maps the joint response of mechanical strain, contact resistance, and communication clock deviation to a three-dimensional phase space and extracts the fractal dimension and maximum Lyapunov exponent of the attractor as coupled chaotic features from the perspective of nonlinear dynamic system.

[0071] In one possible implementation of this embodiment, the extraction process of coupled chaotic features includes: Constructing a multi-field joint response vector ,in As a representative mechanical strain, As a representative contact resistance, The communication clock bias is addressed by normalizing the three components of the joint response vector to ensure each component has zero mean and unit variance, thus eliminating the influence of differences in physical dimensions and amplitudes.

[0072] The normalized joint response vector is reconstructed in three-dimensional phase space, and the reconstructed attractor orbitals are represented as follows: ,in It is a sequence of sampling time points. This represents the total number of sampling points during the excitation period.

[0073] The correlation dimension is calculated for the reconstructed attractor orbitals to quantify the fractal nature of the attractor. The correlation integral is then calculated. ,in , where is the distance scale in phase space. Denotes the Euclidean norm. This is the Heaviside step function. In logarithmic coordinates... and In the curve, the slope of the linear scaling region is used as the correlation dimension. Non-integer It is a typical sign of chaotic motion.

[0074] The maximum Lyapunov exponent is estimated using a small data set method, expressed as follows: ,in For the sampling time step, For reference point Nearest Neighbor The initial Euclidean distance between them For the two points to pass through a time step The distance after evolution, The number of reference points selected. A positive maximum Lyapunov exponent indicates that the system is sensitively dependent on initial conditions and is in a chaotic state.

[0075] Related dimensions and the maximum Lyapunov index Combined into coupled chaotic feature vectors The normal decoupled fingerprint benchmark stores the benchmark parameters under healthy conditions. When the incentive phase extracts When there is a significant deviation from the benchmark value, it indicates that the multi-field coupling dynamics have evolved from decoupled order to chaos.

[0076] S240: The third analysis channel compresses the time-frequency representation of the multi-physics response collected during the S210 excitation process into a low-dimensional latent space through a time-frequency autoencoder. It then uses the information bottleneck principle to separate the shared generation factors and independent factors of each physical domain in the latent space and extracts the features of coupled latent variables.

[0077] In one possible implementation of this embodiment, the extraction process of coupled latent variable features includes: For signals acquired by all sensor channels during the S210 excitation period, time-frequency representation matrices are generated using the same wavelet transform parameters as in S220. These multiple time-frequency representation matrices are then stacked along the channel dimension to form a multi-channel time-frequency input tensor. .

[0078] Construct a time-frequency autoencoder, encoder It consists of cascaded two-dimensional convolutional layers and fully connected layers, from the input tensor The process involves extracting cross-channel time-frequency local patterns layer by layer, ultimately outputting low-dimensional latent variables. Latent variable dimensions Much smaller than the input dimension. Decoder Composed of transposed convolutional layers, it reconstructs the input approximation from latent variables. .

[0079] The training loss function consists of the reconstruction error and the latent variable distribution constraints, expressed as follows: ,in For the Frobenius norm, The maximum mean difference is used to measure the distribution of latent variables induced by the encoder. With prior distribution The differences between them. The prior distribution is set as a standard normal distribution, and the latent variable distribution is aligned with the isotropic distribution by constraining the latent variable distribution, so that different dimensions of the latent space are automatically separated into independent generating factors.

[0080] After training, the latent variables Each dimension calculates mutual information with the original signals from each physical domain, and the latent variable dimensions are divided into: mechanically independent factors that are significantly correlated with mechanical strain, based on the magnitude of mutual information. Electrically independent factor significantly related to contact resistance Heat-independent factors significantly correlated with heat flux Information-independent factors significantly correlated with communication clock deviation And shared generators that simultaneously possess significant mutual information with two or more physical domains. .

[0081] The above factors are concatenated into a coupled latent variable feature vector. This feature encodes the shared generation mechanism of multi-physics coupling and the independent components of each physical domain's contribution to the coupling in the low-dimensional latent space.

[0082] S250: Align and encapsulate the three sets of feature vectors obtained by parallel processing of the three analysis channels S220, S230 and S240 according to a unified timestamp.

[0083] In one possible implementation of this embodiment, the synchronous output of three-dimensional features includes: Coupled resonance eigenvectors output by S220 The features corresponding to each analysis unit in the stimulus enhancement process are arranged in chronological order to form a feature sequence. ,in This represents the total number of analysis units.

[0084] Coupled chaotic feature vectors output by S230 Following the same time segmentation method as S220, the entire excitation period is divided into... For each segment, the correlation dimension and the maximum Lyapunov exponent are extracted independently to obtain the feature sequence. .

[0085] Coupled latent variable eigenvectors output by S240 Similarly, by segmenting the data into time segments, the latent variables of each segment are extracted and their statistical mean is calculated to obtain the feature sequence. .

[0086] Align the three sets of feature vector sequences according to their time indices, and at each time step... A three-dimensional coupled feature group is formed at the location, all The feature sets at each time step are encapsulated as input to the S300 stage.

[0087] Please refer to Figure 4 , Figure 4 The detailed process of stage S300 in this embodiment is shown, which includes stages S310 to S360.

[0088] The S300 feeds three types of features into a three-layer heterogeneous graph network, and outputs the dominant coupling transmission links and path contribution ratios through iterative optimization and pruning.

[0089] S310: Using observation points in the mechanical, electrical, and communication domains as nodes, construct a three-layer heterogeneous graph network containing intra-layer and inter-layer connections, and initialize the hidden states of nodes and edge weights.

[0090] In one possible implementation of this embodiment, the initialization process of the three-layer heterogeneous graph network includes: The mechanical layer node set contains Each node corresponds to a strain measurement location on the mechanical connection surface of the interface. The electrical layer node set contains... Each node corresponds to an electrical terminal or contact resistance measurement location. The communication layer node set contains... Each node corresponds to a communication transceiver or message monitoring location.

[0091] Hidden state vectors of each node The dimensions are uniformly set as The initial hidden state of the mechanical layer nodes is determined by the mechanical independence factor obtained from S240. Obtained through linear projection; the initial hidden state of electrical layer nodes is determined by the electrical independence factor. Obtained through linear projection; the initial hidden state of the communication layer node is determined by the information independence factor. Obtained through linear projection.

[0092] The connection weights of intra-layer edges are initialized after normalization of the corresponding elements of the decoupling-invariant characteristic covariance matrix between nodes in S130. The cross-layer edge connection weights between the mechanical and electrical layers are based on the coupling resonance characteristic vector obtained in S220. The peak value of the time-frequency coherent locking value in the corresponding frequency band is initialized, and cross-layer edges are established only between node pairs where there is a physical force-electric coupling path. The initial weight of the cross-layer edge between the mechanical layer and the communication layer is determined by the time delay cross-correlation between mechanical strain and communication clock deviation; the initial weight of the cross-layer edge between the electrical layer and the communication layer is determined by the time delay cross-information between contact resistance and communication bit error rate statistics.

[0093] The coupled chaotic features output by S230 Define a globally coupled chaos index as a global dynamic constraint embedded in the network. When the norm of a node's hidden state exceeds the limit set by... When a dynamic upper limit is defined, a scaling constraint is applied to the hidden state.

[0094] S320: In each iteration of the graph network, intra-layer message aggregation and cross-layer message passing guided by coupled latent variable features are performed simultaneously to update the hidden state of all nodes.

[0095] In one possible implementation of this embodiment, forward propagation and cross-layer message passing include: Intra-layer message passing performs attention-weighted aggregation on nodes within each layer. The hidden state of the inner aggregation is ,in The learnable weight matrix within the layer. This is the attention coefficient calculated based on node feature similarity.

[0096] Cross-layer message passing based on coupled latent variable features Shared generation factor Message routing is performed based on direction information. Calculation The drift vector along the time dimension is mapped to the physical domain node space to determine the dominant direction for each node to receive cross-layer messages under the current excitation state. After receiving cross-layer messages from other layer nodes, update the intermediate hidden state. ,in For cross-layer learnable weight matrices, For the first Cross-layer connection weights in round iterations.

[0097] The intermediate hidden states of intra-layer aggregation and cross-layer aggregation are merged through a gating mechanism, and the nodes... In the The wheel's update is hidden. ,in This is a gated vector learned through vector concatenation. For activation function, This represents element-wise product.

[0098] S330: After each round of forward propagation, calculate the mutual information contribution of attention for each cross-layer edge, and cut off cross-layer connections with contributions lower than the adaptive threshold, so that the graph network structure is gradually simplified and focused on highly coupled regions.

[0099] In one possible implementation of this embodiment, the attention mutual information contribution evaluation and connection pruning include: For cross-layer edges Calculate the mutual information contribution of attention ,in To score attention, by and Learnable parameters were calculated; For nodes and The mutual information between hidden states is calculated using a nonparametric estimation method based on nearest neighbor distance.

[0100] Calculate the median contribution value for all current cross-layer edges, and set an adaptive pruning threshold. ,in This is a preset scaling factor. For each cross-layer edge, if... If the condition is met, the edge will be permanently removed from the cross-layer edge set and will no longer participate in message passing in subsequent rounds.

[0101] The set of cross-layer edges retained after pruning is denoted as Update the graph network structure as follows: This serves as the initial state for the next iteration.

[0102] S340: After each round of pruning, perform a convergence test on the changes in connection weights of all retained cross-layer edges to determine whether to terminate the loop iteration.

[0103] In one possible implementation of this embodiment, the weight convergence determination process includes: Record number After round pruning, retain the set of connection weights for cross-layer edges. In the first After each round of iterations, calculate the set of weights for the same edge index. .

[0104] Calculate the maximum absolute change in cross-layer edge weights between two adjacent rounds. .when If the number of edges between layers is less than the preset convergence threshold and the number of edges remaining across layers no longer decreases, the graph network is considered converged, and the iteration terminates. A maximum number of iterations is set as a safety limit; if convergence is not achieved by reaching this limit, the process is forcibly terminated, and the current optimal structure is output.

[0105] S350: After convergence is determined in S340, the dominant coupling transmission link from the excitation starting point through different physical domains to the communication observation point is extracted from the pruned graph network, and the contribution ratio of each path to the coupling failure is calculated.

[0106] In one possible implementation of this embodiment, the extraction process of the dominant coupling transmission link includes: The converged graph network In this process, the set of excitation start-point nodes is marked as the nodes in the mechanical layer corresponding to the force application point of the exciter, and the set of response observation point nodes is marked as the clock deviation observation nodes in the communication layer. Using the excitation start-point as the source node and the response observation point as the sink node, all directed paths are enumerated to form a candidate path set.

[0107] For each candidate path Calculate the overall coupling strength of the path This is the product of the mutual information contributions of all cross-layer edge attention along the path. (Select...) The largest path serves as the dominant coupling transmission link. .

[0108] For each path segment on the link Calculate the contribution ratio ,satisfy , The higher the value, the more critical the bridging role of that segment in the transmission of coupling failure.

[0109] S360: Transmits the dominant coupling link Contribution ratio of each segment The converged graph network structure is encapsulated and output as input to the S400 stage.

[0110] Please refer to Figure 5 , Figure 5 The detailed process of stage S400 in this embodiment is shown, which includes stages S410 to S460.

[0111] S400 identifies weak points in coupling through active physical intervention and progressive three-feature consistency verification, and triggers negative feedback iteration based on the verification results.

[0112] S410: Select the cross-domain layer-crossing location with the largest contribution ratio from the dominant link output by S350 as the intervention target point, and apply corresponding micro-physical intervention according to the physical domain type of the target point.

[0113] In one possible implementation of this embodiment, the determination of the intervention target and the application of the intervention include: Traverse all path segments of the dominant link, filter out path segments belonging to cross-domain layer locations, and compare their contribution ratios. The physical interface location with the largest value is selected as the priority intervention target.

[0114] Specifically, when the target point is located on a mechanical connection surface, the intervention method involves applying a pre-tightening torque to the fastening bolt at that location for fine-tuning, with the adjustment range not exceeding a preset small proportion of the original nominal torque value. When the target point is located on an electrical terminal, the intervention method involves simulating the reciprocating displacement of the contact surface within an extremely small stroke, with the displacement amplitude not exceeding a preset proportion of the terminal surface roughness characteristics. When the target point is located on a communication transceiver, the intervention method involves connecting an adjustable delay line in series in the signal link, introducing an additional clock bias not exceeding a preset proportion of the protocol jitter tolerance.

[0115] During the intervention process, all sensor channels continuously acquire data, and the interface full physical field response data are recorded synchronously for each preset duration before and after the intervention.

[0116] S420: For the response data segments before and after the intervention, extract coupled resonance, coupled chaos, and coupled latent variable features simultaneously using the methods of S220, S230, and S240, respectively.

[0117] In one possible implementation of this embodiment, the synchronous extraction of three-dimensional features before and after intervention includes: Using the pre-intervention and post-intervention response data segments as inputs, the time-frequency coherence calculation of S220 is performed in parallel to obtain the pre-intervention coupled resonance characteristics. and post-intervention coupling resonance characteristics .

[0118] The data segments before and after the intervention were reconstructed in three-dimensional phase space. The attractor fractal dimension and the maximum Lyapunov exponent were calculated using the S230 method to obtain the coupled chaotic characteristics before the intervention. and post-intervention coupled chaotic features .

[0119] The data segments before and after the intervention were input into the time-frequency autoencoder trained on S240, and latent variables were obtained through forward inference. Coupled latent variable features were then constructed in the same manner to obtain... and .

[0120] S430: Compare the post-intervention coupled chaotic features with the healthy chaotic parameters in the normal decoupled fingerprint benchmark to determine whether the attractor morphology regresses to the benchmark after intervention.

[0121] In one possible implementation of this embodiment, the coupled chaotic feature regression determination process includes: From the normal decoupling fingerprint benchmark The baseline coupled chaotic features extracted under healthy conditions are retrieved. Calculate the deviation before intervention. and deviation after intervention .

[0122] when If the coupled chaotic features show a regression trend towards the baseline, the first-level verification passes and proceeds to S440. Otherwise, the first-level verification fails and proceeds to S460.

[0123] S440: Provided that the first-level verification is passed, check whether the peak value of the time-frequency coherence lock value of the target cross-domain frequency band after intervention has decreased relative to before intervention.

[0124] In one possible implementation of this embodiment, the coupled resonance feature reduction determination process includes: Extracting pre-intervention coupling resonance features Key frequency bands associated with the cross-domain location of the corresponding target coherence peak And the coherent peak value of the corresponding frequency band after intervention. .

[0125] when If the intervention effectively reduces the force-electric phase lock intensity at that location, the second-level verification is successful, and the process proceeds to S450. Otherwise, the second-level verification fails, and the process proceeds to S460.

[0126] S450: Assuming the second-level verification is successful, determine whether the direction of the coupled latent variable features in the latent space after intervention has migrated towards the reference center.

[0127] In one possible implementation of this embodiment, the determination of the direction of coupled latent variable feature migration includes: Retrieve baseline latent variable features under healthy conditions from normal decoupled fingerprint baselines. Calculate the cosine value of the included angle before intervention. and after intervention .

[0128] when When the condition is met, it indicates that the latent variable direction has shifted towards the baseline center after intervention, and the third-level validation is successful. At this point, the intervention location is formally marked as a weak coupling point and output to S500. Otherwise, the third-level validation fails, and the process proceeds to S460.

[0129] S460: When any level of verification fails, the failure location and feature deviation information are injected into the graph network as negative feedback to trigger a new path optimization loop.

[0130] In one possible implementation of this embodiment, negative feedback injection and new path optimization include: Construct a negative feedback vector containing the failure location index, intervention type, and feature bias at each level. Apply a weight penalty factor to the cross-layer edge corresponding to the failure location, forcibly reducing the connection weight of that edge in subsequent iterations.

[0131] After weighting penalties, the entire process of graph network forward propagation, message passing, contribution evaluation, pruning, and convergence determination (S320 to S350) is re-executed to extract new dominant coupling transmission links from the updated graph network. Based on the new links, the intervention and progressive verification process (S410 to S450) is re-executed. This closed-loop cycle continues until a coupling weakness that has passed all three levels of verification is found or the preset loop limit is reached.

[0132] Please refer to Figure 6 , Figure 6 The detailed process of stage S500 in this embodiment is shown, which includes stages S510 to S540.

[0133] Based on confirmed vulnerabilities and transmission links, S500 calculates a comprehensive health index by weighting and fusing characteristic decay components, and updates the normal decoupled fingerprint benchmark.

[0134] S510: Calculate the decay components in the frequency domain, time domain, and dynamic system dimensions by utilizing the contribution ratio of each segment of the dominant transmission link and the three types of characteristics of the path corresponding to the confirmed weak points.

[0135] In one possible implementation of this embodiment, the calculation process for the decay components in each dimension includes: Frequency domain dimension decay component ,in For path segment The current coupled resonance characteristics at the corresponding position, This is the corresponding value for that position in the normal baseline.

[0136] Time-domain dimension decay component ,in For path segment The current coupled latent variable features at the corresponding position, This is the baseline value.

[0137] Powertrain dimension decline component ,in For path segment The current coupled chaotic features at the corresponding position, This is the baseline value.

[0138] S520: The three dimensions of decline components are integrated into a unified comprehensive health index through an exponential decay function.

[0139] In one possible implementation of this embodiment, the fusion calculation of the interface comprehensive health index includes: Interface Comprehensive Health Index The calculation formula is: ,in , and These are the fusion weight coefficients for the frequency domain, time domain, and dynamic system dimensions, respectively. The sum of these three coefficients is a preset value, which is set according to the criticality of each dimension to interface security. Values ​​in interval, This indicates that the interface health status is completely consistent with the baseline; the lower the value, the more serious the deviation.

[0140] when When the temperature drops below the preset maintenance threshold, a targeted maintenance recommendation is triggered.

[0141] S530: Using the interface comprehensive health index and verified link characteristics as feedback, the center shape and covariance matrix of the normal decoupling fingerprint benchmark are incrementally updated to adapt the benchmark to reasonable drift caused by normal service.

[0142] In one possible implementation of this embodiment, the incremental update of the normal decoupled fingerprint benchmark includes: Mean vector The updated formula is ,in This is the decoupled invariant feature vector extracted for the current measurement period. Learning rate. Adjusted dynamically based on current health index. , The base learning rate. When A higher learning rate allows the baseline to quickly adapt to normal drift; when... The learning rate is automatically reduced when the learning rate is low to prevent abnormal data from contaminating the baseline.

[0143] covariance matrix The incremental update uses an exponentially weighted moving average, and the update formula is: The confidence level of the boundary envelope remains unchanged.

[0144] Incremental updates are only performed if all S400 Level 3 validations are passed, ensuring that only data confirmed as having undergone a healthy regression participates in the baseline update.

[0145] S540: Archive key data throughout the entire process in a unified format to form a complete closed-loop record from benchmark reference, deviation analysis, path tracing, proactive verification to benchmark correction.

[0146] In one possible implementation of this embodiment, closed-loop output and archiving include: The baseline version, three-dimensional coupling feature sequence, dominant transmission link structure, confirmed weak point locations, interface comprehensive health index trend, and baseline update records involved in each complete evaluation cycle are all stored in the maintenance database with timestamps. When the health index shows a unidirectional downward trend for several consecutive cycles and the decline exceeds the preset trend threshold, a trend warning is triggered even if the maintenance threshold has not been reached, prompting attention to the gradual degradation of the interface.

[0147] Through the closed-loop process from S100 to S500, the health status of the skateboard chassis decoupling interface under multi-physics coupling is accurately evaluated, the weak points of coupling are accurately located, and the benchmark is adaptively evolved.

[0148] Example 2: This invention underwent a 18-month system-level deployment and verification on a skateboard chassis integrated test platform of a new energy vehicle manufacturer. The test platform includes eight standardized decoupled interfaces in the front and rear axle areas. The mechanical parts of the interfaces utilize high-strength bolt flanges with locating pins for connection, while the electrical parts employ silver-plated elastic terminals. Communication utilizes 1000BASE-T1 automotive Ethernet with runtime triggering protocol. The test bench simulated three typical load spectra: urban driving conditions, high-speed driving conditions, and reinforced rough road conditions, accumulating a cumulative equivalent mileage of 240,000 kilometers. The entire system deploys 120 sensor channels, including 48 mechanical strain channels, 24 contact resistance channels, 8 communication clock deviation channels, and 16 thermal flux channels, achieving microsecond-level synchronous data acquisition via a PXIe chassis.

[0149] During implementation, the slow feature analysis time window was set to 100ms, and the first three slowest feature components were retained for each physical quantity. The normal baseline confidence level was set to 0.99. The initial power spectral density of the random vibration excitation was set to 0.0005g. 2 / Hz, increasing by 0.0002g every 5 minutes. 2The step size enhancement is / Hz, with an upper limit of 85% of the design allowable value; the peak pulse power injection current is 15% of the rated current, and the duty cycle cycles through 10%, 30%, 50%, 70%, and 90%; the communication storm initiation rate is 8% of the bus bandwidth, increasing by 3% every 3 minutes, with an upper limit of 40%. The hidden state dimension of the graph network is 64, the pruning ratio coefficient is 0.5, and the convergence threshold is 10. -4 Health index integration weight , , Take values ​​of 0.4, 0.35, and 0.25 respectively.

[0150] I. Typical Case: Early Identification and Precise Location of Mechanical Fretting Wear on the Left Side Interface of the Front Axle: When the cumulative equivalent mileage reached 86,000 kilometers, the system successfully performed multiphysics coupling feature extraction in the S200 stage. The first analysis channel detected that the peak value of the force-electric time-frequency coherence lock-in value at the left front axle interface in the 350Hz to 420Hz frequency band continuously increased from the baseline value of 0.18 to 0.65, while the coherence values ​​at other interfaces in the same frequency band remained below 0.25. The second analysis channel extracted a three-dimensional phase space attractor correlation dimension that decreased from the baseline value of 2.13 to 1.67, and the maximum Lyapunov exponent increased from 0.012 to 0.041, indicating that the multiphysics coupling dynamics tended towards chaos. The third analysis channel showed that the proportion of shared generation factors in the latent space increased from the baseline of 23% to 41%, indicating a significant enhancement in the activation degree of the force-electric shared factors.

[0151] After feeding the aforementioned 3D features into the graph network, and through 5 rounds of iterative pruning and convergence, the dominant coupling transmission link is output as: Mechanical Layer Node 2 → Electrical Layer Node 1 → Communication Layer Node 1, where the attention mutual information contribution of the path segment from the Mechanical Layer to the Electrical Layer is... Contribution ratio This exceeds the range of other path segments in the link. Based on this, the intervention target is locked at the flange fastening bolt corresponding to node 2 of the mechanical layer.

[0152] After applying a +3.5% nominal torque fine-tuning intervention, the three-level progressive verification results are as follows: Level 1 verification, coupling chaotic deviation before intervention. After intervention The deviation decreased by 71.1%, indicating significant regression of chaotic characteristics. In the second level of verification, the peak time-frequency coherence at the target frequency of 380Hz decreased from 0.65 before intervention to 0.22 after intervention, a reduction of 66.2%, demonstrating clear resonance reduction. In the third level of verification, the cosine of the angle between the latent variable characteristic and the baseline center increased from 0.68 before intervention to 0.93 after intervention, a directional shift of 36.8%. All three levels of verification were passed, confirming this location as a weak coupling point. The interface's overall health index recovered from 68 points before intervention to 89 points.

[0153] On-site disassembly and inspection revealed a fretting wear band approximately 2.3 mm wide on the bolt connection surface, with localized peeling of the plating in the contact area, consistent with the diagnostic results. This case demonstrated an early warning time approximately 4200 kilometers (equivalent mileage) earlier than traditional torque monitoring methods.

[0154] II. Overall Performance Comparison; Table 1 Comparison of key indicators between the method of the present invention and existing technologies. To present the above comparison more intuitively, Figure 7 , Figure 8 and Figure 9 A bar chart comparing the accuracy of locating weak points in coupling and the average early warning lead time is presented, along with a curve showing the change in the comprehensive health index of a certain interface during an 18-month verification period.

[0155] III. Cumulative results during the verification period; During the 18-month verification period, the system accurately identified and located 12 coupling weaknesses, including 5 mechanical fretting wear points, 4 electrical terminal contact degradation points, 2 insufficient communication timing margin points, and 1 thermal stress concentration point. A three-level progressive verification mechanism eliminated 6 false weaknesses, avoiding unnecessary disassembly. Planned maintenance triggered by the interface comprehensive health index reduced unplanned interface failures by 87% compared to traditional strategies, and the overall maintenance cost decreased by 62% compared to the periodic comprehensive tightening solution. The normal decoupling fingerprint benchmark underwent 4 incremental updates driven by 18 valid cases, improving the matching degree between the benchmark parameters and the actual health status from an initial 0.87 to 0.96. The system's average single full-process time was 11.6 minutes, the average convergence rounds of the graph network was 5.1, and the peak computational resource utilization rate was 41%, meeting the real-time requirements of industrial sites. These results fully verify the engineering effectiveness of the method of this invention in accurately evaluating the health status of decoupling interfaces, precisely locating weaknesses, and enabling benchmark self-evolution in complex multi-physics coupling environments.

[0156] Example 3: A test and evaluation system for the multi-dimensional coupling characteristics of a skateboard chassis decoupling interface, comprising five interfaces, as follows: Fingerprint baseline panel: Displays the multidimensional distribution center shape and boundary envelope of the interface in a healthy state.

[0157] Feature Extraction Panel: Used to display the extraction results of three types of features: coupling resonance, chaos, and latent variables.

[0158] Link tracing panel: Used to track the dominant coupling transmission links and quantify the contribution of each path segment.

[0159] Progressive verification panel: Used to perform three-level consistency verification and identify weak points in coupling.

[0160] Health Index Panel: Used to integrate health indices through the fusion calculation interface and drive benchmark updates.

[0161] The system includes: The benchmark construction module is used to collect mechanical strain, contact resistance, communication clock deviation and heat flux signals in the early operation phase of the interface, analyze the slow characteristics of each signal to obtain decoupling invariant characteristics, and construct a normal decoupling fingerprint benchmark. The multi-dimensional feature extraction module is used to simultaneously apply random vibration, pulse power jump and communication broadcast storm when the interface load state is consistent with the normal decoupling fingerprint benchmark, and collect multi-field responses. It uses three analysis channels to extract the deviance of the normal decoupling fingerprint benchmark from different dimensions. The extracted features are three types of features: coupling resonance features, coupling chaos features and coupling latent variable features. The link optimization module is used to input three types of features into a three-layer heterogeneous graph network. The mechanical-electrical layer connection weights are initialized with resonance features, the state evolution of each layer is constrained by chaotic features, and the cross-layer message passing is driven by latent variable features. In the network iteration, each round evaluates the contribution of the connection to the resonance and latent variable features by the mutual information of cross-layer nodes, removes low-contribution connections and feeds back the weight graph to the next round, until the cross-domain edge weights converge, and outputs the dominant coupling link and the contribution ratio of each path segment to the coupling failure. The weak point identification module is used to apply a small amount of physical intervention at the cross-domain layer crossing point with the highest contribution ratio in the dominant coupling link, and simultaneously extract three types of features before and after the intervention and progressively verify them: if the chaotic feature regresses the baseline attractor shape, the resonance feature shows a reduction in the coherence value of a specific frequency band and the direction of the latent variable approaches the baseline center, then it is identified as a weak point; otherwise, the corresponding position is injected into the graph network to readjust the weights and find the optimal solution until a consistent weak point is jointly confirmed by the three features. The health assessment and correction module is used to identify weak points and links, extract three types of characteristics and contribution ratios of each physical quantity on the chain, and calculate the decay components of the benchmark in the frequency domain, time domain and dynamic system dimensions by weighting the contribution ratios. These components are then integrated into an interface comprehensive health index, which is used to incrementally update the normal decoupled fingerprint benchmark, forming a complete link from benchmark reference, deviation analysis, path tracing, active verification to correction.

[0162] Those skilled in the art will understand that the embodiments of this application are provided as methods, systems, or computer program products. Therefore, this application takes the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application takes the form of a computer program product implemented on one or more computer storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer program code. The solutions in the embodiments of this application are implemented using various computer languages, exemplified by the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0163] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, are implemented by computer program instructions. These computer program instructions are provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams.

[0164] These computer program instructions are also stored in a computer read-memory that can direct a computer or other programmed data processing device to operate in a particular manner, such that the instructions stored in the computer read-memory produce an article of manufacture including instruction means that implement the functions specified in the flowchart or multiple flowcharts and / or block diagram blocks or multiple block diagrams.

[0165] These computer program instructions are also loaded onto a computer or other programming data processing device to cause a series of operational steps to be performed on the computer or other programming device to produce a computer-implemented process, such that the instructions, which execute on the computer or other programming device, provide steps for implementing the functions specified in the flowchart flow or multiple flows and / or the block diagram blocks or multiple blocks.

[0166] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0167] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for testing and evaluating the multi-dimensional coupling characteristics of a skateboard chassis decoupling interface, characterized in that, include: During the early operation phase of the interface, mechanical strain, contact resistance, communication clock deviation and heat flux signals are collected. The decoupling invariant characteristics are obtained by analyzing the slow characteristics of each signal, and a normal decoupling fingerprint benchmark is constructed. When the interface load state is consistent with the normal decoupling fingerprint benchmark, random vibration, pulse power jump and communication broadcast storm are applied simultaneously to collect multiple field responses; Three analysis channels were used to extract the deviation of the normal decoupling fingerprint benchmark from different dimensions. The extracted features were three types of features: coupling resonance features, coupling chaos features, and coupling latent variable features. Three types of features are input into a three-layer heterogeneous graph network. The mechanical-electrical layer connection weights are initialized with resonance features, the state evolution of each layer is constrained by chaotic features, and the cross-layer message passing is driven by latent variable features. In the network iteration, each round evaluates the contribution of the connection to the resonance and latent variable features by the mutual information of cross-layer nodes, removes low-contribution connections and feeds the weight graph back to the next round, until the cross-domain edge weights converge, and outputs the dominant coupling link and the contribution ratio of each path to the coupling failure. A small amount of physical intervention is applied at the cross-domain layer crossing point with the highest contribution ratio in the dominant coupling link. Three types of features before and after the intervention are extracted and progressively verified: if the chaotic feature regresses the baseline attractor morphology, the resonance feature shows a reduction in the coherence value of a specific frequency band and the direction of the latent variable approaches the baseline center, then it is confirmed as a weak point; otherwise, the location that failed to verify is negatively fed into the graph network to readjust the weights and find the optimal solution until a consistent weak point is confirmed by the joint verification of the three features. After identifying the weak points and links, three types of features and contribution ratios of each physical quantity on the chain are extracted. The decay components of the benchmark in the frequency domain, time domain and dynamic system dimensions are calculated by weighting the contribution ratios and merging them into an interface comprehensive health index. Based on this, the normal decoupling fingerprint benchmark is incrementally updated, forming a complete link from benchmark reference, deviation analysis, path tracing, active verification to correction.

2. The method for testing and evaluating the multi-dimensional coupling characteristics of the skateboard chassis decoupling interface according to claim 1, characterized in that, The decoupling-invariant features obtained from the slow feature analysis of each signal include: Slow feature analysis was performed on mechanical strain, contact resistance, communication clock skew, and heat flux signals respectively. The slow feature analysis aimed to minimize the square mean of the time derivative of the output features. With the objective and under the zero-mean constraint Unit variance constraint and decorrelation constraints with extracted features The following solution yields the slow characteristic components of each physical quantity that change most gradually and are uncorrelated, among which... For the first The first physical quantity One output slow feature component; Components whose statistical stability satisfies the condition that their coefficient of variation is below a preset threshold and their autocorrelation decay time constant is greater than a preset duration threshold are selected from all slow feature components and used as decoupling invariant features.

3. The method for testing and evaluating the multi-dimensional coupling characteristics of the skateboard chassis decoupling interface according to claim 2, characterized in that, The construction of the normal decoupled fingerprint benchmark includes: Construct a decoupling-invariant eigenvector from the decoupling-invariant features. Calculate its mean vector during the early operational phase. Covariance Matrix ; Mahalanobis distance Based on, take satisfaction The equiprobable ellipsoids serve as the boundary envelope, where For degrees of freedom The chi-square distribution at confidence level The upper quantile below, This represents the total number of invariant feature vector samples collected during the early operational phase; this forms the normal decoupling fingerprint benchmark, and deviations beyond the boundary envelope indicate a deviation in the interface state.

4. The method for testing and evaluating the multi-dimensional coupling characteristics of the skateboard chassis decoupling interface according to claim 1, characterized in that, The method for extracting the coupled resonance features is as follows: Representative mechanical strain and contact resistance signals from the excitation process are selected, and continuous wavelet transforms are performed to obtain time-frequency complex coefficients. and ; Calculate the force-electric time-frequency coherent locking value in the time-frequency plane. ;in for The complex conjugate; The peak amplitude and peak frequency of the time-averaged coherence spectrum are extracted from the preset key frequency band set to form the coupled resonance characteristics.

5. The method for testing and evaluating the multi-dimensional coupling characteristics of the skateboard chassis decoupling interface according to claim 4, characterized in that, The method for extracting coupled chaotic features is as follows: The joint response vector of mechanical strain, contact resistance and communication clock deviation during the excitation process is mapped to three-dimensional phase space to reconstruct the attractor orbit; Calculate the attractor correlation dimension Quantify fractal characteristics and estimate the maximum Lyapunov exponent. To quantify initial value-sensitive dependencies, where For time step, and These represent the initial distance and the evolved distance between the reference point and its nearest neighbor, respectively. Number of reference points; Will and The combination results in coupled chaotic features.

6. The method for testing and evaluating the multi-dimensional coupling characteristics of the skateboard chassis decoupling interface according to claim 5, characterized in that, The method for extracting coupled latent variable features is as follows: A time-frequency autoencoder is constructed, which takes the time-frequency representation of the multi-physics response as input, compresses it into low-dimensional latent variables through the encoder, and then reconstructs it through the decoder. The training loss includes the reconstruction error and the maximum mean difference between the latent variable distribution and the prior distribution. In the latent space, based on the mutual information of each dimension and each physical quantity signal, the latent variables are separated into independent factors that are significantly associated with a single physical domain and shared generative factors that are simultaneously associated with two or more physical domains. The independent factors and shared generative factors are spliced ​​together to form coupled latent variable features, so as to express the cross-domain coupled generative structure in the low-dimensional manifold.

7. The method for testing and evaluating the multi-dimensional coupling characteristics of the skateboard chassis decoupling interface according to claim 1, characterized in that, In the three-layer heterogeneous graph network: The cross-layer connection weight between mechanical and electrical layer nodes is initialized by the time-frequency coherence locking value of the corresponding frequency band in the coupling resonance characteristics; The coupled chaotic features are converted into global coupling indicators, and an upper limit constraint is imposed on the evolution range of the hidden states of each layer of nodes. The dominant receiving direction of cross-layer message passing is determined by the drift direction of the shared generating factor in the latent space of the coupled latent variable features. The hidden state of a node is updated through intra-layer attention aggregation and gating fusion of cross-layer messages.

8. The method for testing and evaluating the multi-dimensional coupling characteristics of the skateboard chassis decoupling interface according to claim 7, characterized in that, Each round evaluates the contribution of connections to resonance and latent variable features using cross-layer node attention mutual information, and removes low-contribution connections, including: For cross-layer edges Calculate the mutual information contribution of attention ,in To score attention, and For nodes and Hidden state For mutual information, This is the set of all cross-layer edges; The adaptive pruning threshold is obtained by multiplying the median contribution of all current cross-layer edges by a preset ratio coefficient. Cross-layer edges with a contribution below the threshold are permanently removed, forcing subsequent iterations to focus on regions with high coupling strength.

9. The method for testing and evaluating the multi-dimensional coupling characteristics of the skateboard chassis decoupling interface according to claim 8, characterized in that, The output dominant coupling link and the contribution ratio of each path segment to the coupling failure include: In the graph network after pruning and convergence, candidate paths are enumerated, with the node corresponding to the starting point of the stimulus as the source node and the node corresponding to the response observation point as the sink node. Path-based coupling strength , The contribution of mutual information of attention to cross-layer edge e, To select a candidate path from the source node to the sink node, The largest one serves as the dominant coupling and transmission link; The contribution ratio of each path segment on the link is determined as follows: , For the first The contribution of cross-layer edges corresponding to segment paths. This represents the total number of road segments.

10. The method for testing and evaluating the multi-dimensional coupling characteristics of the skateboard chassis decoupling interface according to claim 1, characterized in that, The method involves applying a small amount of physical intervention at the cross-domain layer crossing point where the contribution ratio in the dominant coupling link is the highest, and simultaneously extracting three types of features before and after the intervention, including: The cross-domain layer-crossing path segment with the largest contribution ratio is selected, and the intervention method is selected according to the physical domain type to which it belongs: for mechanical connection surfaces, a pre-tightening torque is applied for fine adjustment; for electrical terminals, a nanometer-level contact surface reversal is applied; for communication transceivers, an additional clock bias not exceeding the preset ratio of the protocol jitter tolerance is introduced. During the intervention, the full physical field response of the interface was collected synchronously, and coupled resonance features, coupled chaotic features, and coupled latent variable features were extracted in parallel from the data segments before and after the intervention.

11. The method for testing and evaluating the multi-dimensional coupling characteristics of the skateboard chassis decoupling interface according to claim 10, characterized in that, The criterion for determining the morphology of the baseline attractor in the chaotic feature regression is as follows: Let the baseline coupled chaotic characteristics be: The chaotic characteristics before and after intervention are respectively and When satisfied At that time, it is determined that the coupled chaotic features regress to the normal state.

12. The method for testing and evaluating the multi-dimensional coupling characteristics of the skateboard chassis decoupling interface according to claim 11, characterized in that, The criterion for determining the reduction of coherence values ​​in a specific frequency band of the resonance feature is as follows: In the key frequency band associated with the path segment that contributes the highest proportion in the dominant coupling transmission link, the peak value of time-frequency coherence lock-in after intervention is smaller than that before intervention, i.e. ,in, In the frequency band before intervention Peak value of time-frequency coherent lock value, For intervention in the frequency band The peak value of the time-frequency coherent lock value.

13. The method for testing and evaluating the multi-dimensional coupling characteristics of the skateboard chassis decoupling interface according to claim 11, characterized in that, The condition for determining whether the direction of the latent variable approaches the reference center is: Let the baseline coupling latent variable characteristics be: The characteristics of coupled latent variables before and after intervention are as follows: and When satisfied When the time is right, the direction of the latent variable is determined to shift towards the reference center.

14. The method for testing and evaluating the multi-dimensional coupling characteristics of the skateboard chassis decoupling interface according to claim 1, characterized in that, The steps for obtaining the comprehensive health index via the interface include: according to Calculate the frequency domain decay component, according to Calculate the time-domain decay component, according to Calculate the decay component of the dynamic system, where , , Path segments The current coupled resonance features, coupled latent variable features, and coupled chaotic features at the corresponding positions. , , For the corresponding benchmark value; The overall health index of the interface is ; , and These are the fusion weight coefficients for the frequency domain, time domain, and dynamic system dimensions, respectively.

15. The test and evaluation system for the multi-dimensional coupling characteristics of the skateboard chassis decoupling interface according to any one of claims 1-14, characterized in that, The system includes: The benchmark construction module is used to collect mechanical strain, contact resistance, communication clock deviation and heat flux signals in the early operation phase of the interface, analyze the slow characteristics of each signal to obtain decoupling invariant characteristics, and construct a normal decoupling fingerprint benchmark. The multi-dimensional feature extraction module is used to simultaneously apply random vibration, pulse power jump and communication broadcast storm when the interface load state is consistent with the normal decoupling fingerprint benchmark, and collect multi-field responses. It uses three analysis channels to extract the deviance of the normal decoupling fingerprint benchmark from different dimensions. The extracted features are three types of features: coupling resonance features, coupling chaos features and coupling latent variable features. The link optimization module is used to input three types of features into a three-layer heterogeneous graph network. The mechanical-electrical layer connection weights are initialized with resonance features, the state evolution of each layer is constrained by chaotic features, and the cross-layer message passing is driven by latent variable features. In the network iteration, each round evaluates the contribution of the connection to the resonance and latent variable features by the mutual information of cross-layer nodes, removes low-contribution connections and feeds back the weight graph to the next round, until the cross-domain edge weights converge, and outputs the dominant coupling link and the contribution ratio of each path segment to the coupling failure. The weak point identification module is used to apply a small amount of physical intervention at the cross-domain layer crossing point with the highest contribution ratio in the dominant coupling link, and simultaneously extract three types of features before and after the intervention and progressively verify them: if the chaotic feature regresses the baseline attractor shape, the resonance feature shows a reduction in the coherence value of a specific frequency band and the direction of the latent variable approaches the baseline center, then it is identified as a weak point; otherwise, the corresponding position is injected into the graph network to readjust the weights and find the optimal solution until a consistent weak point is jointly confirmed by the three features. The health assessment and correction module is used to identify weak points and links, extract three types of characteristics and contribution ratios of each physical quantity on the chain, and calculate the decay components of the benchmark in the frequency domain, time domain and dynamic system dimensions by weighting the contribution ratios. These components are then integrated into an interface comprehensive health index, which is used to incrementally update the normal decoupled fingerprint benchmark, forming a complete link from benchmark reference, deviation analysis, path tracing, active verification to correction.

Images