Corrosion-resistant circuit board column connection structure sliding wire failure prediction method

Abstract

The invention discloses a corrosion-resistant circuit board column connection structure sliding wire failure prediction method, which comprises the following steps: acquiring multi-modal operation data of a circuit board structure under multiple environmental stresses such as temperature, humidity and vibration through a multi-sensor network, and carrying out filtering, normalization and feature extraction to generate a structured multi-physical field coupling feature vector; finite element simulation and a material aging experiment are combined, and a time sequence feature set of multi-physical field information such as structural stress, elastic modulus degradation and friction coefficients is constructed; according to the method, based on a material microscopic damage mechanism, mapping between slippery thread trend characteristics and damage degrees is established by using methods such as support vector regression and the like, and the mapping is input into a long-short term memory network with online self-adaptive capability for dynamic prediction, so that high-precision and online evaluation of the slippery thread failure trend is realized; and a prediction result with spatial positioning and failure path interpretation is output, so that the preventive maintenance decision-making capability and the structure failure early warning level of the electronic equipment are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of corrosion-resistant circuit board column connection structure failure prediction, and particularly relates to a corrosion-resistant circuit board column connection structure slipping thread failure prediction method. BACKGROUND

[0002] At present, the failure prediction of the corrosion-resistant circuit board column connection structure is one of the key technical problems in the field of high-reliability electronic equipment manufacturing and maintenance. With the development of electronic products towards high density and high integration, the column connection structure (such as BGA, CSP and other interconnection methods) is prone to slipping thread and other microscopic failure modes under complex service environment, which seriously affects the long-term reliability and safety of the equipment. In view of this problem, researchers and engineering and technical personnel at home and abroad have proposed various prediction and evaluation schemes, mainly including life prediction methods based on stress and strain monitoring, data-driven state trend analysis technology, and structure reliability evaluation assisted by finite element simulation, etc. The existing mainstream technologies mainly focus on two categories: one is to rely on a large number of structure sensor data, combined with traditional time series analysis or early convolutional / recurrent neural network models, to use the structure response or "slipping thread" change trend for failure warning; the other is to use the finite element method to simulate the mechanics of the circuit board and the column connection structure under typical load, and to fit the column connection failure life model combined with the experience reduction coefficient. In recent years, some scholars have tried to include single physical parameters such as material aging and interface friction into the failure criterion to improve the physical rationality of calculation and prediction. Simple splicing of machine learning and finite element results, multivariate statistical regression and other trend feature extraction methods have appeared in the field. However, these technologies often focus on prediction accuracy in industrial practice, ignoring the consistency of the nature of the failure mechanism; The application scope of typical technical schemes mainly covers: failure warning of high-reliability electronic packaging, life modeling of column connection structure under multi-environmental stress, structure health monitoring based on sensor and simulation data fusion, etc. However, the existing methods mostly have the following limitations: (1) Weak correlation between trend features and physical mechanisms. At present, most feature extraction methods (such as signal statistics, wavelet energy, spectral principal component, etc.) lack clear mapping with the core failure processes such as material micro-damage evolution and interface slipping, resulting in that the prediction model, although showing good short-term accuracy, lacks explainability and is difficult to provide theoretical basis for engineering operation and failure tracing; (2) Lack of multi-physical field coupling analysis capability. Many methods only focus on structural stress or single environmental parameters, and fail to organically integrate temperature, humidity, vibration stress and material aging, interface friction and other multi-physical field factors, resulting in that the model cannot accurately reflect the real damage evolution path of the column connection structure under complex working conditions; (3) The prediction model lacks online adaptability. Most existing machine learning or deep learning models are trained statically, and have limited adjustment capability when facing sudden changes in the field environment or long-term feature drift, which can easily lead to lagging or distorted prediction results; (4) The failure path space has poor interpretability. Traditional methods report the device state based on statistical results such as the overall wire slip amount and key point stress value, which makes it difficult to visually explain the spatial evolution of the failure process and affects the accurate positioning of high-risk areas and subsequent precision maintenance. SUMMARY

[0003] The present application provides a corrosion-resistant circuit board column connection structure wire slip failure prediction method to solve the above technical problems.

[0004] The technical solution of the present application is as follows: A corrosion-resistant circuit board column connection structure wire slip failure prediction method comprises: S1: Collecting running data of the circuit board column connection structure under different environmental stress conditions, the running data including temperature, humidity, vibration parameters and structure stress response signals, and recording corresponding environmental condition labels; S2: Filtering and normalizing the collected structure stress response signals to eliminate environmental noise interference and unify the signal dimension, and generating a standardized structure stress sequence; S3: Based on the fusion of finite element simulation and standardized structure stress sequence, calculating the stress distribution cloud of each key node of the column connection structure, and extracting the time sequence evolution characteristics of the local maximum stress points; S4: Fitting the elastic modulus degradation curve according to the material aging experiment data, and combining the time-varying relationship of the interface friction coefficient to generate a set of multi-physical field coupling feature vectors; S5: Establishing a mapping function between the wire slip trend characteristics and the material micro-damage mechanism, and mapping the multi-physical field coupling feature vectors into wire slip trend feature vectors with physical consistency based on the function; S6: Inputting the wire slip trend feature vectors into the prediction model with online parameter updating capability, and dynamically adjusting the model weights through the back propagation algorithm to adapt to the changes in the current environmental conditions; S7: Judging whether the model prediction error exceeds the preset threshold, if it does, triggering the model parameter online optimization mechanism, adjusting the learning rate and updating the historical feature database; S8: Outputting the wire slip failure trend curve based on the optimized prediction model, and generating a prediction result containing failure path explanation, which is used to guide the preventive maintenance decision of high-reliability electronic equipment.

[0005] The corrosion-resistant circuit board column connection structure wire slip failure prediction method provided by the present application has the following beneficial effects: (1) The present application extracts the dynamic evolution of stress concentration area, material stiffness degradation and friction coefficient change and other multi-physical field significant characteristics by coupling real working condition environment data through finite element simulation, realizes the mechanism driven mapping of damage degree and wire sliding trend characteristics combined with micro damage evolution equation, enhances the explainability of the prediction result, and improves the credibility of the wire sliding failure warning; (2) The present application innovatively integrates multi-physical field characteristics and material damage mechanism, so that the prediction model not only captures the real-time structural stress and wire sliding amount change, but also dynamically responds to material performance degradation and interface friction evolution, and adapts to variable and complex environmental conditions. The LSTM network with dynamic weight adjustment and online parameter updating mechanism, combined with error adaptive optimization strategy, realizes real-time learning and correction of the model to environmental disturbance; (3) The present application effectively suppresses high-frequency noise and environmental interference in the collected signal through multi-stage filtering and decomposition such as Butterworth low-pass, FIR band-pass and wavelet multi-scale, ensures high signal-to-noise ratio of the structural stress and wire sliding trend signal. The spatial localization and time sequence accurate synchronization of key information are ensured by the space-time mapping and interpolation, feature fusion method, the high consistency of finite element simulation and measured boundary conditions, which breakthroughly improves the spatial resolution and time tracking ability of stress and wire sliding response; (4) The present application does not need to rely on specific hardware or special sensors, and is suitable for columnar structure failure monitoring of multiple types of electronic equipment and various structural processes, especially suitable for full life cycle health management of high reliability electronic products. In addition, the structured input and output and explainable failure path mechanism of the present method facilitate the embedding of existing equipment management system and maintenance decision platform, and the on-site personnel can intuitively obtain failure warning and spatial positioning instructions through the support of the visualization module; (5) The present application realizes real-time capture and rapid recovery of the abnormal state of the model based on the threshold determination and adaptive learning rate adjustment of the prediction error, effectively avoids the long-term influence of data drift, environmental change and other factors on the prediction accuracy. The continuous expansion and retraining mechanism of the historical feature database ensures long-term iterative optimization of the model, and the self-learning ability is outstanding. BRIEF DESCRIPTION OF DRAWINGS

[0006] Fig. 1 is a flow chart of a corrosion-resistant circuit board columnar structure wire sliding failure prediction method of the present application; Fig. 2 is a sub-flow chart of a corrosion-resistant circuit board columnar structure wire sliding failure prediction method of the present application; Fig. 3 is another sub-flow chart of a corrosion-resistant circuit board columnar structure wire sliding failure prediction method of the present application. DETAILED DESCRIPTION

[0007] In order to make the objects and advantages of the present application more clear, the present application will be further described below in conjunction with embodiments; it should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0008] The preferred implementation methods of the present application are described below with reference to the accompanying drawings. Those skilled in the art should understand that these implementation methods are only used to explain the technical principles of the present application and not to limit the protection scope of the present application.

[0009] As used herein, the singular forms "a", "an" and "the" also include the plural forms unless the context clearly indicates otherwise. It should also be understood that the terms "comprise / comprising" or "have / having" specify the presence of stated features, integers, steps, operations, components, parts, or combinations thereof, but do not exclude the presence or addition of one or more other features, integers, steps, operations, components, parts, or combinations thereof. Meanwhile, the term "and / or" used in the specification includes any and all combinations of the related listed items.

[0010] Please refer to Figs. 1-3 As shown in the drawings, a corrosion-resistant circuit board column connection structure slip failure prediction method comprises: S1: Collecting running data of the circuit board column connection structure under different environmental stress conditions, the running data including temperature, humidity, vibration parameters and structure stress response signals, and recording corresponding environmental condition labels; S2: Filtering and normalizing the collected structure stress response signals to eliminate environmental noise interference and unify signal dimensions, and generating a standardized structure stress sequence; S3: Based on the fusion of finite element simulation and the standardized structure stress sequence, calculating the stress distribution cloud of each key node of the column connection structure, and extracting the time sequence evolution characteristics of the local maximum stress points; S4: Fitting the elastic modulus degradation curve according to the material aging experiment data, and combining the relationship between the interface friction coefficient and time to generate a multi-physical field coupling feature vector set; S5: Establishing a mapping function between the slip trend characteristics and the material micro-damage mechanism, and mapping the multi-physical field coupling feature vector into a slip trend feature vector with physical consistency based on the function; S6: Inputting the slip trend feature vector into a prediction model with online parameter updating capability, and dynamically adjusting the model weight through the back propagation algorithm to adapt to the current environmental condition changes; S7: Judging whether the model prediction error exceeds a preset threshold, if yes, triggering the model parameter online optimization mechanism, adjusting the learning rate and updating the historical feature database; S8: Output the gliding failure trend curve based on the optimized prediction model, and generate the prediction result containing the failure path explanation, which is used to guide the preventive maintenance decision of the high-reliability electronic device.

[0011] The step S1: Collecting the operation data of the circuit board column connection structure under different environmental stress conditions, the operation data including temperature, humidity, vibration parameters and structure stress response signals, and recording the corresponding environmental working condition labels. Specifically, it includes: S1.1: Based on the industrial sensor network, obtain the multi-modal operation data of the circuit board column connection structure under the action of various typical environmental stresses, the operation data including the thermal field distribution signal output by the temperature sensor, the environmental humidity time series data collected by the humidity sensor, the vibration frequency spectrum information recorded by the three-axis accelerometer, and the structure stress response signal collected by the strain gauge, to construct a multi-dimensional environmental-structure coupling data set; During the operation data collection stage of the circuit board column connection structure under multi-environmental stress conditions, an industrial sensor network is used as the data acquisition framework to realize the synchronous acquisition and spatial calibration of multi-modal and multi-source signals; A distributed temperature collection module (parameters: sensor accuracy ±0.1℃, sampling frequency 1Hz) is used to collect the thermal field distribution signals of the surface and key nodes of the column connection structure, and a two-dimensional thermal field signal matrix is generated through the multi-point temperature measurement matrix; Further, through the industrial-grade humidity sensor array (parameters: relative humidity accuracy ±1%, sampling frequency 0.5Hz), the environmental humidity time series data is collected, and the humidity sequence synchronized with the temperature data in time is generated; Further, through the three-axis acceleration sensor (parameters: range ±16g, sampling frequency 5kHz, frequency response range 0.5Hz-2kHz), vibration acceleration signal collection is performed, and Fourier transform is used to extract vibration frequency spectrum information to form a frequency domain power spectrum density curve; Further, through the high-sensitivity strain gauge (parameters: sensitivity coefficient 2.0±0.1, bridge resistance 120Ω, sampling frequency 2kHz), structure stress response signal collection is realized, and the original voltage output is converted to strain value , and then the elastic modulus is calculated , and the formula is: Wherein is the structure stress, is the material elastic modulus, is the strain value. Through the above multi-source signals, a multi-dimensional environmental-structure coupling data set containing four physical field dimensions of temperature, humidity, vibration and stress is constructed; By using multi-channel synchronous acquisition and timestamp calibration, the output signals of various types of sensors are aligned to a unified time axis, and the sensor deployment coordinates are mapped to the corresponding nodes in the geometric model of the column connection structure to achieve spatial positioning and data association. By using data integrity verification and outlier removal algorithms (parameter: 3σ principle threshold setting), impulse noise and sensor distortion data are removed to ensure that the constructed dataset meets the accuracy requirements of subsequent simulation and feature extraction. By using a sensor network and time synchronization system, the thermal field, humidity spectrum, vibration spectrum and stress response signals from the previous step are integrated into the same data structure to achieve the goal of acquiring full, multi-physics field information. For example, on a circuit board pillar structure equipped with 8 temperature acquisition nodes, 2 humidity sensors, 4 triaxial accelerometers, and 12 strain gauges, the temperature sensors (accuracy ±0.1℃, sampling frequency 1Hz) form an 8×T two-dimensional thermal field matrix under a step change in ambient temperature (25℃→80℃, gradient 5℃ / min). The humidity sensors (accuracy ±1%, sampling frequency 0.5Hz) output a time-series humidity vector under periodic fluctuations in relative humidity of 40%-90%. The triaxial accelerometers (range ±16g, sampling frequency 5kHz) obtain the power spectral density curve through FFT under a 5Hz-1.5kHz sweep frequency vibration loading condition, with its main peak frequency shifting from 200Hz to 210Hz. The strain gauges (sensitivity coefficient 2.0) acquire the peak strain value under the superposition of static load and random vibration conditions. Combined with elastic modulus Pa, through The calculated peak stress is Pa. After all data were time-synchronized and aligned using the NTP protocol, a multimodal coupled original dataset was constructed to meet the requirements of subsequent S1.2 filtering. This dataset can effectively reflect the spatiotemporal correlation between environmental stress and structural response in finite element simulation verification. S1.2: Perform frequency domain filtering processing based on the Butterworth low-pass filtering algorithm on the acquired structural stress response signal to suppress high-frequency noise interference and retain the low-frequency response characteristics of the structure, and obtain the filtered structural stress time series signal. For the original structural stress response signal synchronously acquired and timestamped by the S1.1 multi-sensor network, a Butterworth low-pass filtering algorithm (filter order n=4, cutoff frequency) is used. Frequency domain filtering (Hz) is used to suppress interference components introduced by high-frequency components of sensor thermal noise and environmental mechanical vibration in the signal; Furthermore, by performing a Fast Fourier Transform (FFT) operation on the original stress response signal, its full-band amplitude-frequency response spectrum is obtained. and according to the filter transfer function The filter response output is calculated, and the transfer function is defined as: wherein, is the frequency, is the set cutoff frequency, is the filter order; Further, by multiplying the spectral amplitude by , the amplitude-frequency characteristic after filtering is obtained, and after inverse FFT transformation, the preliminary noise-reduced structural stress time series signal is generated ; Further, the time-domain waveform smoothness of the filtered signal is detected by envelope analysis method, and the root mean square (RMS) value and the crest factor are calculated as the filter effect evaluation index, to ensure that the high-frequency noise component is effectively suppressed while the low-frequency structural response waveform characteristics are not distorted; Through the above Butterworth low-pass filtering process, the original stress response signal is converted into a filtered time series signal with attenuated high-frequency spectral components and preserved low-frequency components, achieving the technical effect of reducing high-frequency noise interference and preserving the true low-frequency mechanical response characteristics of the PCB column structure; For example, in a stress response original signal sequence with a sampling frequency of Hz, the FFT operation obtains the main peak distributed at Hz and Hz, with corresponding amplitudes of MPa and MPa. With a cutoff frequency of Hz and an order of , the Butterworth low-pass filter is configured The amplitude of the spectrum is modified, and the amplitude of the component above Hz is attenuated by more than dB. After inverse FFT, the RMS value of the time-domain signal is reduced from MPa to MPa, and the peak ratio is reduced from to , while the Hz main peak is not significantly attenuated, verifying that the high-frequency interference is effectively filtered out while the low-frequency mechanical response is preserved, providing a high signal-to-noise ratio input signal for the normalization processing of S1.3; S1.3: Adopting Z-score standardization method to normalize the filtered structural stress time series signal to eliminate the dimensional differences of different sensors, generating a standardized structural stress sequence as the input data for subsequent finite element simulation; S1.4: Based on the control parameters of the environmental stress test box and the vibration bench, record the environmental condition label corresponding to each set of collected data, including temperature gradient level, humidity interval, vibration frequency bandwidth and acceleration amplitude range, to establish the mapping relationship between structure response and environmental excitation; S1.5: Time stamp alignment and data packaging of the standardized structure stress sequence and the corresponding environmental condition label are performed to generate a structured running data sample set as the input data source of the finite element simulation module and the feature extraction module.

[0012] The step S2: filtering and normalization processing of the collected structure stress response signal is performed to eliminate environmental noise interference and unify the signal dimension, and a standardized structure stress sequence is generated. Specifically, it includes: S2.1: The original structure stress response signal collected is subjected to band-pass filtering processing to remove the high-frequency interference components introduced by environmental vibration and sensor noise in the signal, and a denoised structure stress time series signal is obtained; The original structure stress response signal collected and time-synchronized in the S1 stage is subjected to FIR type digital band-pass filtering algorithm (parameters: lower limit frequency of passband is 1-10 Hz, preferably 5 Hz, upper limit frequency of passband is 100-500 Hz, preferably 300 Hz, filter order is 64-256, preferably 128) to realize the function of isolating ultra-low frequency trend items and high-frequency interference components; Further, the filter impulse response is generated by filter coefficient design based on window function method (window function type: Hanning window); The convolution operation is performed on the original stress time series signal ; Further, the power spectral density (PSD) analysis of the signals before and after filtering is performed by FFT, the frequency energy distribution change is calculated, and the high-frequency noise energy attenuation ratio is evaluated; Further, for the residual peak interference components found based on PSD analysis, the neighborhood mean smoothing algorithm (window length points) is performed on the frequency domain amplitude curve to suppress the high-amplitude noise of isolated frequency points; Through band-pass filtering and smoothing processing, the original stress response signal is converted into a denoised structure stress time series signal that retains the target frequency band structure dynamic response and removes low-frequency drift and high-frequency interference, achieving the technical effects of improving signal signal-to-noise ratio and enhancing the accuracy of subsequent wavelet multi-scale analysis; For example, in a stress response original signal with a sampling frequency of Hz and a sampling point number of , the passband range of the FIR band-pass filter is set to ~ Hz, the window type is Hanning window, and the peak-to-peak value of the filtered signal obtained after convolution operation is reduced by MPa to MPa, which is higher than Hz, the energy attenuation of the frequency band is more than dB, which is lower than Hz, the drift component is weakened to of the original energy. After PSD smoothing, the amplitude of the residual interference peak is reduced by about %. The signal is analyzed by subsequent S2.2 wavelet decomposition, and the energy concentration of the sub-band component is significantly improved, verifying the effective denoising and spectral purification effect realized by this step; S2.2: Based on the wavelet transform algorithm, the denoised structural stress time series signal is subjected to multi-scale decomposition, and the wavelet coefficients at each scale are extracted to identify and suppress the non-stationary components in the residual noise; S2.3: The structural stress signal reconstructed by the wavelet coefficient is subjected to mean normalization operation, and the signal amplitude is mapped to the [0, 1] interval to eliminate the influence of the dimension difference between different sensors, forming a normalized stress sequence with a unified dimension; S2.4: Based on the sliding window mechanism, the normalized stress sequence is segmented and processed, and the mean, variance and kurtosis characteristics of each window are calculated to enhance the representation ability of the local statistical characteristics of the signal, and a structural stress feature sequence with time local characteristics is generated; S2.5: The time local feature sequence is time-aligned with the corresponding environmental condition label to generate a joint dataset of structural stress response and environmental stress condition, which is used as input data for subsequent finite element simulation and multi-physical field feature fusion analysis.

[0013] The step S3: based on the fusion of finite element simulation and standardized structural stress sequence, the stress distribution cloud of each key node of the column-connected structure is calculated, and the time evolution characteristics of the local maximum stress point are extracted. As shown in Fig. 2 , specifically comprising: S3.1: The standardized structural stress sequence is subjected to spatial mapping processing to establish the corresponding relationship between the structural stress data and the geometric model of the column-connected structure, thereby generating stress boundary condition input suitable for finite element simulation; S3.2: Based on the three-dimensional geometric model and material attribute parameters of the column-connected structure, a finite element analysis model is constructed, and the standardized stress boundary condition is imported, and static and dynamic coupled structural mechanics simulation is performed to obtain the stress response distribution of each key node under different environmental conditions; Based on the three-dimensional geometric model and material property parameters of the column-connected structure, the input conditions for constructing the finite element analysis model include standardized stress boundary condition data processed by S3.1 spatial mapping, three-dimensional structural geometric dimension information, and material constitutive parameter set; Using the geometric modeling tools embedded in the 3D modeling software (parameters: length, width, height and aperture are set according to the actual circuit board column connection structure survey data), a 3D CAD model of the column connection structure with the same scale as the physical object is generated, realizing the construction of the consistency between the geometric integrity of the structure and the simulation model; Furthermore, material mechanical property parameters (including elastic modulus) are imported and corrected through a materials database. Poisson's ratio Yield strength and density Furthermore, nonlinear constitutive relations are established in the simulation platform to ensure that the model can accurately simulate material behavior under high strain rate and high temperature and humidity conditions. Furthermore, a mesh generation algorithm is adopted (parameters: element type is ten-node tetrahedral element, average element side length is 0.2-1 mm, preferably 0.5 mm, key areas are refined to...). The structure is discretized by using a mesh quality check and shape optimization algorithm to ensure that the element distortion rate is less than 5%, thereby improving simulation convergence and computational accuracy. Furthermore, the standardized stress boundary conditions generated in S3.1 are applied to the corresponding nodes or faces of the finite element model. A node position mapping matching algorithm is used to realize the correspondence between the boundary data and the model constraints, and the applied load is calculated using the following node force conversion formula: in, For the first The equivalent effect of each node This represents the normal stress component at that node. The area affected by the node; Furthermore, the static-dynamic coupled analysis module is enabled in the solution settings. The static part is used to simulate the stress distribution field under constant working conditions, and the dynamic part is based on the explicit integration algorithm and the Newmark-β method (parameter: β is set to...). γ take This allows for the solution of transient response under dynamic loads and outputs an array of stress distribution data for each time step. Through the above operations of the finite element solution module, the three-dimensional geometric model, accurate material properties and measured boundary condition data are combined to output the stress response distribution matrix of each key node under different environmental conditions, thus achieving the expected technical effect of quantitative analysis of stress field. For example, in a certain dimension of length mm, width mm, height mm, the circuit board column connection structure is simulated, the elastic modulus GPa, Poisson's ratio , yield strength MPa, density kg / m³ alloy material is selected as the model input. The constitutive relation adopts a multi-linear hardening mode, and the grid element edge length is refined to mm in the through hole and the welding area.

[0014] S3.1 In the boundary condition data after spatial mapping, the normal stress of a surface node is MPa, the influence area is mm², and the node force is calculated according to the formula N. The above load is applied to the model node set, and a 10 Hz harmonic stress disturbance is superimposed in the dynamic working condition. After static-dynamic coupling analysis, the maximum principal stress is concentrated in the transition area of the solder pad, and the peak stress is MPa, which is increased by about % compared with the static load working condition, providing high-precision input data for subsequent stress contour generation and local maximum stress point time series extraction; S3.3: Visualize the stress distribution data output by the finite element simulation to generate stress contour maps of the column connection structure at different time steps, to intuitively present the stress concentration areas and their evolution trends inside the structure; For the stress distribution data matrix of each key node obtained by the finite element simulation calculation in S3.2, a three-dimensional data visualization algorithm based on isosurface rendering (parameters: color scale range covering the minimum to maximum interval of the absolute value of the node stress, color mapping scheme using the gamma-corrected pseudo- rainbow mode) is used to realize the intuitive presentation of the spatial distribution of the stress field at different time steps; Further, through a triple interpolation algorithm (spatial dimension using trilinear interpolation, time dimension using linear interpolation, parameters: spatial element resolution 0.1 mm, time resolution 0.5 ms), the finite element results are interpolated and resampled in space-time domain, realizing the spatial continuity and time smoothing of the stress distribution data, and obtaining the smoothed stress tensor field sequence; Further, by stress component decomposition and principal stress calculation, the maximum principal stress distribution field at each time step is extracted as the visualization carrier, and the spatial distribution matrix containing the stress concentration area is generated; Further, for the principal stress distribution matrix, a Gaussian kernel convolution smoothing algorithm (parameters: standard deviation σ takes 0.8 element length) is used to suppress local discrete numerical oscillation, and the stress value is mapped to the unified color scale range [0, 1] through normalization processing, to ensure the comparison consistency of the stress contour maps at different time steps; Further, the processed stress distribution matrix is imported into a three-dimensional visualization engine, a rendering mode of Phong lighting model is set, a horizontal sectioning interval is 0.5 mm, a vertical sectioning interval is 0.3 mm, and an output resolution is 4096*4096 pixels, so that a stress nephogram sequence of the columnar structure at different time steps is generated; Through the equivalent surface rendering, interpolation smoothing, principal stress extraction, color scale normalization and high-definition rendering processing, the numerical matrix output by the finite element simulation is converted into analyzable image data capable of directly reflecting the stress concentration area and the time evolution trend, so that the dynamic change process of the complex stress field is visually expressed. For example, when the simulation result visualization of a group of columnar structures with a length of 50 mm, a width of 10 mm, a height of 5 mm and a hole diameter of 1.5 mm is performed, a pseudo- rainbow color mapping scheme with gamma = 2.2 is selected, the maximum principal stress range output by the simulation is linearly mapped to the color scale value [0, 1] from 50 MPa to 320 MPa, and after trilinear interpolation and time 0.5 ms step resampling, the spatial distribution surface of the principal stress field at each time step changes continuously and has no obvious crack-like color band. In the maximum principal stress result obtained by stress component decomposition and principal stress calculation and extraction, the stress normalization value of the pad transition area at the time of 0.25 s is 0.97 (corresponding to the actual value of 310 MPa), which is obviously higher than 0.65 (208 MPa) of other areas at the same time, after Gaussian kernel σ = 0.8 mm smoothing, the image visual noise is reduced by about 42%, and the stress concentration area boundary definition is improved by about 35%, which effectively supports the identification of local maximum stress points and the extraction of stress change sequence in the time dimension of S3.4; S3.4: Identify the local maximum stress points in the columnar structure based on the stress nephogram, and extract the stress change sequence in the time dimension to form the time evolution curve of the local maximum stress points; For the columnar structure stress nephogram sequence at different time steps generated by S3.3, a region growing algorithm based on normalized principal stress threshold segmentation (parameters: the initial seed point is set to the node with a local stress intensity greater than 0.95, the connectivity rule is 26-neighborhood, and the upper limit of iteration number is 50) is used to realize the automatic region identification of high stress concentration areas; Further, the spatial positioning of the candidate maximum stress points in each high stress region is realized by a method based on three-dimensional centroid calculation and spatial connected domain analysis (parameters: the centroid positioning adopts the weighted average method, and the weight is the node stress value), so that the local extreme coordinate set at each time step is obtained; Further, for each candidate extreme point, the maximum principal stress value of the neighborhood thereof is calculated, and the formula: wherein, The maximum principal stress component of the nth node in the neighborhood of the candidate point is determined by eliminating local noise interference and identifying a globally unique local maximum stress point. Furthermore, using a time index matching algorithm (parameters: time step resampling interval 0.5 ms, linear interpolation to fill in missing moments), the stress values ​​of the local maximum stress points at each time step are assembled into a stress change sequence in chronological order. ; Furthermore, a bidirectional Savitzky-Golay smoothing filter (parameters: window length of 9, fitting polynomial order of 3) was used to... Smoothing is performed to suppress high-frequency oscillations and enhance the trend characteristics of the curve, forming a complete time-series evolution curve of the local maximum stress point; Through the above-mentioned region segmentation, extreme value extraction, time series assembly and smoothing, the visualized stress cloud map is quantitatively transformed into stress change curves of key structural risk locations in the entire time domain, realizing accurate tracking of the stress concentration evolution process and preparation of input for subsequent feature extraction. For example, in a length of mm, width is mm, height is In the simulation data of the column-connected structure with a diameter of mm, the stress cloud map color scale normalization range is 0~1. In the region growing algorithm, the initial seed points are two neighboring nodes with a stress normalization value greater than 0.95, which are merged into a single region after connected component analysis. The centroid coordinates calculated by the weighted centroid method are (25.1 mm, 4.9 mm, 2.6 mm), with a neighborhood search radius of 0.3 mm, containing 12 grid nodes, and the local maximum principal stress is... MPa. After time resampling, 400 time-series sampling points were obtained. After applying Savitzky-Golay smoothing, the peak position remained unchanged, the signal-to-noise ratio was improved by about 38%, and the curve correlation coefficient was improved to 0.97 compared with the original sequence, effectively supporting the time-frequency feature extraction of S3.5; S3.5: Perform feature extraction processing on the time-series evolution curve of the local maximum stress point, and use time-frequency analysis to extract key statistical features such as peak value, mean, variance and rate of change, and generate a local stress evolution feature vector with characterization ability.

[0015] Step S4: Fit the elastic modulus degradation curve based on the material aging test data, and combine it with the relationship between the interfacial friction coefficient and time to generate a multi-physics coupling feature vector set. For example... Fig. 3 As shown, it specifically includes: S4.1: Preprocess the data on the change of elastic modulus over time collected in the material aging test to remove outliers and noise interference in order to obtain a time-series degradation sample set of the elastic modulus of the structural material. For the data of elastic modulus changing with time collected in material aging experiment, the input object is a set of original elastic modulus time series records from multiple batches of environmental stress tests, and the data format is a sequence of time stamp and corresponding elastic modulus value pairs; An outlier detection method (parameters: based on the 3σ principle, threshold multiple k is 3.0) is used to remove abnormal points in the elastic modulus sequence caused by measurement errors or sensor instability. Further, a sliding average filtering algorithm (parameters: window width n is 5 sampling points) is used to smooth the elastic modulus sequence after abnormality removal, and the preliminary denoised time series data is obtained. Further, a fast Fourier transform (FFT) spectral analysis method (parameters: sampling frequency 10 Hz, frequency resolution 0.1 Hz) is used to diagnose the frequency domain characteristics of the smoothed sequence, identify and locate the main frequency component of high-frequency noise. Further, according to the frequency domain analysis result, a band-stop filter is constructed (parameters: stop band range covers the main frequency of high-frequency noise ±0.5 Hz, effective stop band attenuation is greater than 40 dB), and the target frequency band attenuation processing of the elastic modulus time series data is realized, and the purified data sequence without significant high-frequency noise is generated. Further, a linear interpolation algorithm (parameters: interpolation interval 0.1 s) is used to interpolate the filtered elastic modulus sequence at the missing time stamp to ensure the integrity and uniform sampling of the time axis. Through the above algorithm processing method, the original noisy and irregularly sampled experimental data is converted into a time-uniform, abnormal-point-free and high-frequency interference-suppressed elastic modulus time series degradation sample set, realizing the expected technical effect of providing high-quality input for the S4.2 curve fitting step. For example, for a material aging test conducted at a constant temperature of 85 ℃ and a relative humidity of 85 %, the original elastic modulus data sampling frequency is 1 Hz and the duration is 7200 s. A total of 15 abnormal points are detected in the initial sequence, and the 3σ rejection method (mean 3.12 GPa, standard deviation 0.08 GPa) is used to judge the points whose modulus values deviate from the mean by more than 0.24 GPa. All abnormal points are deleted. After applying the sliding average filter (window width 5 s) to the remaining data, the FFT spectrum shows that there is a significant high-frequency peak near 1.8 Hz, with a peak amplitude of 0.012 GPa. After constructing a band-stop filter with a stop band range of [1.3 Hz, 2.3 Hz], the energy in this frequency band is reduced by about 45 dB in the filtered result. Finally, a total of 8 missing sampling points are filled by linear interpolation, and the output elastic modulus time series degradation sample set has a length of 7200 points, with a signal-to-noise ratio improved by about 41% compared to the original sequence. This sample set can be directly input into the least squares fitting module to construct the degradation function. S4.2: Curve fitting on the elastic modulus degradation sample set based on the least square fitting algorithm to generate the degradation function expression of the elastic modulus with time to construct the stiffness degradation characteristics in the material aging process; For the material elastic modulus time series degradation sample set output by S4.1 step, the least square fitting algorithm (parameters: fitting polynomial order n is 2~4 order, preferably 3 order, determined according to the Akaike information criterion AIC and the Bayesian information criterion BIC) is used to realize the functional expression of the elastic modulus attenuation trend; Further, by constructing the time variable vector And the corresponding elastic modulus observation value vector , the polynomial function model , the coefficient vector is the fitting parameter; Further, the normal equation is constructed , wherein is the time power matrix, is the observation value column vector, and the coefficient estimate value in the least square sense is obtained by solving the above equation; Further, the residual error vector is calculated, wherein is the corresponding elastic modulus estimate value on the fitting curve, and the residual sum of squares is calculated to evaluate the goodness of fit; Further, the determination coefficient ( is the total sum of squares) is used to calculate the model explanation rate, and the optimal polynomial order is selected according to the value and AIC to balance the model accuracy and complexity; Further, the optimal fitting function expression is output as the elastic modulus degradation function in the material aging process, denoted as , wherein is the time variable, which is used for subsequent S4.4 multi-physical field coupling characteristic construction; Through the least square fitting algorithm and the model optimization strategy, the experimental sample data in S4.1 stage is converted into a continuous differentiable elastic modulus degradation function with clear physical meaning, realizing the quantification and predictable expression of the stiffness degradation characteristics; For example, under the condition of temperature 85℃, humidity 85%RH, using the least square fitting with the fitting polynomial order candidate set {2,3,4}, the time sampling interval is 1s, the total sampling point number is 7200, and the time variable matrix has a dimension of 7200×(n+1). When n=3, the coefficient vector is obtained, the residual sum of squares , and the determination coefficient is 0.9948, and the AIC value is the lowest. Then the degenerate function is (unit: GPa). The function can stably reflect the degradation trend of the elastic modulus throughout the test period, with a relative error of not more than 2.1%, providing high-confidence material aging characteristic input for subsequent characteristic coupling analysis; S4.3: Collect the friction coefficient experimental data of the column-continuous structure interface under the action of different environmental stresses, and generate the characteristic curve of the interface friction coefficient evolution with time through interpolation and smoothing processing to represent the sliding trend of the structure connection interface; The collected object is the original experimental sequence of the friction coefficient of the column-continuous structure connection interface under different temperature, humidity, vibration frequency and acceleration amplitude environmental stress conditions, and the data sources include the synchronous output of the interface friction tester and the environmental stress simulation device; A multi-channel synchronous sampling method (parameters: sampling frequency 1 kHz, channel synchronization error less than 1 μs) is used to realize the synchronous collection of the tangential force and normal force signals of the interface, and the instantaneous friction coefficient of each time sampling point is calculated according to the formula: wherein is the tangential force, is the normal force; Further, the sliding average filtering algorithm (parameters: window width 5 sampling points, overlap rate 80%) is used to perform smoothing processing on the instantaneous friction coefficient sequence, to realize the preliminary suppression of high-frequency random fluctuations and retain the real sliding evolution trend of the interface; Further, the piecewise cubic Hermite interpolation method (parameters: interpolation interval 0.1 s, ensuring the continuity of the first derivative) is used to perform time axis completion and resampling on the smoothed friction coefficient data, to obtain the uniformly sampled friction coefficient time series curve; Further, the exponential smoothing algorithm (parameters: smoothing coefficient a = 0.3) is used to perform trend enhancement processing on the interpolated and resampled curve, to enhance the long-term slow variation characteristics of the interface friction coefficient and suppress short-term disturbances; Through the above interpolation and smoothing joint processing, the original discrete and noisy friction coefficient measurement data is converted into the interface friction coefficient characteristic curve evolving with time, realizing the quantitative characterization of the sliding trend of the column-continuous structure connection interface; For example, under the conditions of temperature 85 ℃, humidity 85 %RH, vibration frequency 10 Hz and acceleration amplitude 5 m / s², the interface friction test is carried out, the tangential force collected by the tester fluctuates between 0.95 N and 1.05 N, and the normal force is stable at 5 N. According to the formula The instantaneous friction coefficient is calculated to be about 0.200±0.005. After the sliding average filtering, the random fluctuation amplitude is reduced by about 42%, the Hermite interpolation of three times to 0.1 s sampling interval forms a friction coefficient sequence of 72000 time points, and the long-term downward trend of the curve is obvious after the exponential smoothing processing, the 24-hour cumulative decrease is 0.015, which corresponds to the increase of 7.5% of the slip trend, which can be directly used as the interface property input of the S4.4 multi-physical field coupling feature construction; S4.4: Based on the elastic modulus degradation function and the interface friction coefficient evolution curve in the material aging process, a time sequence feature matrix of multi-physical field coupling is constructed combined with the structural stress distribution data, so as to enhance the physical explanation ability of the feature to the slip failure mechanism; S4.5: The time sequence feature matrix of multi-physical field coupling is normalized and dimensionally reduced, and the principal component analysis algorithm is used to extract the key feature components to generate a low-dimensional multi-physical field coupling feature vector set with physical consistency.

[0016] The step S5: a mapping function between the slip trend feature and the material micro-damage mechanism is established, and the multi-physical field coupling feature vector is mapped into a slip trend feature vector with physical consistency based on the function. Specifically, it includes: S5.1: The material micro-damage mechanism is classified and modeled, and the damage evolution equation is constructed based on the fracture mechanics and fatigue damage theory to obtain the damage degree evolution function of the material under different environmental stresses; S5.2: Based on the multi-physical field coupling feature vector set generated by S4, the elastic modulus degradation rate and the interface friction coefficient change rate are extracted as input variables input into the damage evolution equation to calculate the equivalent damage degree of the material at the current time; For the multi-physical field coupling feature vector set generated by S4, the input object is a low-dimensional feature vector set with physical consistency, which includes the elastic modulus degradation function and the interface friction coefficient evolution curve; The first-order difference algorithm (parameters: time interval Δt and feature sampling period are consistent) is used to calculate the elastic modulus degradation function The time derivative operation is performed to realize the calculation of the degradation rate The calculation formula is: Among them represents the elastic modulus degradation rate; Further, the same first-order difference algorithm is used to perform time derivative operation on the interface friction coefficient evolution curve The friction coefficient change rate is extracted, and the calculation formula is: Among them This represents the rate of change of the friction coefficient; Furthermore, the extracted and Aligned by timestamps and combined to form a two-dimensional input vector. , as a driving variable in the damage evolution equation; Furthermore, the damage evolution equation based on the S5.1 output Substitute the input vectors sequentially to calculate the equivalent damage at the current time. The damage evolution equation can be expressed as: , , , These are the calibration coefficients obtained from fitting the materials based on experimental data; Through the above chain calculation process, the multi-physics coupling feature vector set is transformed into a numerical index containing physical degradation rate information, realizing the accurate quantification of equivalent damage and providing a physically consistent input for subsequent slippage trend feature mapping. For example, under operating conditions of 85°C and 85%RH, with Δt taken as 1s, the elastic modulus degradation function is: After differential calculation at t=3600s = -1.247 × 10^-4 GPa / s. The interfacial friction coefficient curve, after the same difference, at t = 3600 s... =-1.2×10^-6 s^-1. Combining these two into a vector (-1.247×10^-4, -1.2×10^-6) and substituting it into the equivalent damage equation, α=8200, β=2.5×10^5, γ=0.002, the calculated... =0.002 - 1.02134 + -0.3 ≈ -1.31934 (the negative sign indicates an increasing trend in damage). This output metric is directly used as the physical label input for training the S5.3 nonlinear mapping, ensuring that the model predictions have interpretable damage mechanisms. S5.3: Construct a nonlinear mapping function between the slippage trend characteristics and the equivalent damage degree, and use the support vector regression algorithm to train the parameters of the function so that the slippage trend characteristics can reflect the physical evolution process of material damage; A nonlinear correlation is established between the equivalent damage value sequence output from step S5.2 and the slippage trend characteristics, with the time-aligned equivalent damage value as the input. With the corresponding set of multi-physics field slippage characteristics at the time point; Support Vector Regression (SVR) method was used (parameters: kernel function type is radial basis function RBF, penalty factor). Value range [1, 100], kernel width Range of values ​​[

[10] ), to realize the definition and parameter initialization of the nonlinear mapping structure from the slippage trend characteristics to the equivalent damage degree; Furthermore, by training the feature vector of slippage trend in the sample set... With equivalent damage Construct the optimization problem for SVR, with the objective function being: in, The weight vector in the high-dimensional feature space. and These are slack variables; Furthermore, a radial basis kernel function is employed. Mapping the slippage trend characteristics to a high-dimensional space enhances the fitting ability of complex nonlinear relationships; Furthermore, the folded cross-validation algorithm (parameter: number of folds) is used. =5), for and The combination of [variables] is used for grid search optimization, and the evaluation metric is the mean squared error. In order to select the hyperparameter combination with the best generalization ability; Furthermore, the SVR model trained with optimal parameters is used as a nonlinear mapping function. This makes the mapped output satisfy This ensures that the slippage trend characteristics can reflect the physical evolution of material damage. By using the SVR nonlinear regression algorithm and parameter optimization strategy, the multi-physics slippage feature vector is transformed into a predicted value that is highly correlated with the equivalent damage degree, thereby enhancing the physical consistency of the features and quantifying the damage mechanism mapping. For example, under environmental conditions of 85°C, 85%RH, and 10Hz vibration frequency, the input feature vector... The model has a dimension of 20, 500 training samples, and uses an RBF kernel-based SVR model. The parameter search space is set to... ∈{1,10,50,100}, ∈{ , , , In 5-fold cross-validation, 1,10}. =50、 =0.01 combination The lowest value was 0.0023, with a Pearson correlation coefficient of 0.962 for the test set and a residual standard deviation of 0.047. The relative error between the output after SVR mapping and the measured equivalent damage degree did not exceed 3.5% over the entire time domain, ensuring the reliability of the slippage trend characteristic in characterizing the material damage process. S5.4: Combine the temporal evolution features of the local maximum stress points extracted in S3 with the mapping function trained in S5.3 to generate a slippage trend feature vector with a time dimension. This vector contains coupled information on structural stress evolution and material damage accumulation. For the time-series evolution features of the local maximum stress point output by the S3 sub-step and the nonlinear mapping function trained by the S5.3 sub-step, the input objects are the time-aligned local stress evolution feature vector and the parameter structure of the nonlinear mapping model. A time synchronization registration algorithm is used (parameter: allowable synchronization error range). The sampling period is adjusted to ensure consistency between the input time base of the local maximum stress point time vector and the equivalent damage degree mapping function, thus guaranteeing the temporal correspondence of the mapping calculation. Furthermore, by using a data feature concatenation method (parameters: concatenation order is peak stress, mean stress, variance, and rate of change), the local stress evolution features are constructed into a fixed-dimensional feature input matrix. , which serves as the set of external driving variables for the nonlinear mapping function; Furthermore, the feature matrix The time-step input is fed into the support vector regression mapping function trained in substep S5.3. Perform mapping operations and output the predicted equivalent damage response. The mathematical form is: in, For optimal parameters and The calibrated radial basis kernel SVR model, It is a set of local stress temporal characteristics; Furthermore, for the output of continuous time steps The sequence is smoothed using a sliding window (parameter: window length). (with an overlap rate of 80%), enhancing the stability of trend features and suppressing the impact of measurement noise on the instantaneous disturbance of the mapping results; Furthermore, the smoothed Damage trend data and the original local stress evolution feature sequence are concatenated and fused at the multi-dimensional vector level to form a composite time series vector. This allows the output to simultaneously include structural mechanical response information and material damage accumulation trend characteristics; The structural stress evolution index output by the chain fusion method is physically consistent with the S5.3 damage mechanism mapping output, generating a time-dimension slip trend feature vector that fully reflects the coupling relationship between structural stress changes and damage accumulation processes; For example, under the conditions of a temperature of 85°C, a humidity of 85% RH, and a vibration frequency of 10 Hz, the local maximum stress point time sequence characteristics include a peak value of 2.85 MPa, a mean value of 2.12 MPa, a variance of 0.15 (MPa) , a rate of change of -1.2 x 10 -3 MPa / s. After time synchronization and registration, these characteristics form a matrix in the order specified, which is input into the S5.3 nonlinear mapping function. = 50, = 0.01 RBF kernel SVR model, the output predicted equivalent damage degree sequence is 0.215 at 3600 s, and after sliding smoothing processing with a window length of 5, the fluctuation amplitude decreases from ±0.012 to ±0.004. The smoothed is spliced with the original characteristics according to time to generate a vector, which, when finally used as input to the prediction model, improves the Pearson correlation coefficient of the prediction to 0.967, verifying the effectiveness of the fusion strategy in enhancing the explanatory ability of the slip trend feature; S5.5: Perform normalization processing on the generated slip trend feature vector to eliminate the dimensional differences between different characteristic dimensions, generating a standardized slip trend feature vector set for use in the prediction model in S6.

[0017] The step S6: input the slip trend feature vector into a prediction model with online parameter updating capability, and dynamically adjust the model weights through the back propagation algorithm to adapt to the changes in the current environmental conditions. Specifically, it includes: S6.1: input the slip trend feature vector into a prediction model constructed based on a long short-term memory network (LSTM), wherein the slip trend feature vector includes structural stress distribution, elastic modulus degradation curve, and interface friction coefficient change, and other multi-physical field coupling characteristics, to construct an input space representation with physical consistency; S6.2: initialize the hidden layer state of the LSTM model, set the initial weight matrix based on the He normal distribution initialization strategy to ensure the gradient stability at the initial stage of model training, and generate an initial model parameter set; For the technical implementation of S6.2 sub-step, the hidden layer state of the LSTM model is initialized to provide stable and excellent convergence performance initial conditions for dynamic prediction of the slip trend feature vector; The initial hidden state of all time steps is initialized by a hidden state zero vector initialization method The cell state is set as a matrix with all elements being zero, ensuring that the initial state does not introduce additional bias and making the model rely only on input features for iterative updates.

[0018] Further, based on the He normal distribution initialization algorithm (parameters: mean , standard deviation , where is the number of input neurons of the previous layer), initial values are generated for the weight matrices of the input gate, forget gate, output gate and candidate cell in LSTM, to maintain the variance stability in the signal propagation process and reduce the risk of gradient disappearance; Further, He normal random sampling is performed on the weight matrix (input to each gate), so that each input feature has a uniform energy scale in the initial propagation; Further, the recurrent weight matrix (hidden state to each gate) is also initialized with He normal distribution to ensure that the gradient remains numerically stable after multiple steps of propagation in the time recursion structure; Further, the bias vectors of each gate are initialized with a constant, with the forget gate bias set to a positive value to retain more historical information in the initial training stage, and the bias of the remaining gates set to zero to avoid excessive activation in the early stage; Through the above initialization processing, a stable association between the initial parameter state of the LSTM network and the multi-physical field coupling slip trend feature mapping space input in S6.1 is established, achieving balanced propagation and convergence acceleration of the gradient flow in the early training stage; For example, under the condition of input feature dimension , hidden layer unit number , the He normal standard deviation ≈ is calculated , and the normal distribution sampling with mean zero and standard deviation 0.3162 is performed on , the standard deviation ≈ is calculated and the initialized matrix is sampled, and the forget gate bias vector is assigned , and other biases are set to zero. After performing the above initialization, the convergence speed of LSTM in 100 rounds of training is improved by about 18%, and the gradient norm fluctuation amplitude is reduced to within 35% of the initial training, ensuring the prediction stability and high-precision output under the condition of complex multi-working condition data input; S6.3: Perform forward propagation calculation based on the filamentation trend feature vector, generate the predicted output sequence of the filamentation failure trend, and calculate the error with the corresponding true failure trend label to obtain the prediction error vector; S6.4: Perform back propagation algorithm on the prediction error vector, calculate the gradient change of the model parameters based on the gradient descent method, and dynamically update the model weight matrix using the Adam optimizer to reduce the prediction error and enhance the model's fitting ability to the current working condition; S6.5: Store the updated model weight matrix to the online parameter cache area, and synchronously update the sliding average loss value of the model to generate a dynamic prediction model version with online learning ability for subsequent filamentation trend evaluation in the prediction period.

[0019] The step S7: judge whether the model prediction error exceeds the preset threshold, if it exceeds, trigger the model parameter online optimization mechanism, adjust the learning rate and update the historical feature database. Specifically, it includes: S7.1: Time alignment processing is performed on the filamentation failure trend prediction results output by the prediction model and the actual observed filamentation state data to ensure the time sequence consistency of the error calculation; S7.2: Calculate the mean square error (MSE) index based on the time-aligned prediction results and observation data as the quantitative evaluation basis for model prediction error; S7.3: Compare the calculated mean square error with the preset error threshold to determine whether the model is in an abnormal prediction state, if the error exceeds the threshold, generate a model optimization trigger signal; S7.4: In response to the model optimization trigger signal, execute the learning rate adaptive adjustment strategy, dynamically reduce or increase the learning rate based on the current error gradient to enhance the model's convergence stability and adaptation speed; S7.5: Update the current environmental condition label and the corresponding filamentation trend feature vector to the historical feature database for subsequent model retraining and feature distribution update to improve the long-term prediction performance of the model.

[0020] The step S8: output the filamentation failure trend curve based on the optimized prediction model, and generate the prediction results containing the failure path explanation, which is used to guide the preventive maintenance decision of high-reliability electronic equipment. Specifically, it includes: S8.1: Post-processing filtering is performed on the original filamentation trend sequence output by the optimized prediction model, and a sliding window average algorithm is used to remove the prediction fluctuation noise to obtain a smooth filamentation displacement evolution trend curve; S8.2: Perform cross-correlation analysis based on the filamentation displacement evolution trend curve and the material elastic modulus degradation curve to identify the time sequence coupling relationship between the filamentation trend and the material aging rate to quantify the contribution of filamentation evolution to structural failure; S8.3: According to the structural mechanics analysis model and the multi-physical field coupling feature set, a physical mapping function of the wire pull failure path is constructed, the prediction result is mapped to the three-dimensional structural coordinate space, and an explainable failure path atlas with spatial positioning ability is generated; The input conditions are the wire pull displacement evolution trend curve, the material elastic modulus degradation curve obtained in step S8.2, and the three-dimensional geometric structure parameter set in the structural mechanics analysis model, and the feature vector set including the structural stress distribution characteristics and the interface friction coefficient change characteristics in the multi-physical field coupling feature set is called at the same time; A three-dimensional finite element result back-projection algorithm (parameters: node coordinate matrix , element connection matrix ) is used to realize the spatial mapping between the wire pull amplitude at each time in the prediction result and the structural geometric grid nodes, and an initial spatial field distribution matrix is obtained; Further, a structural stress weight superposition method (parameters: stress component weight coefficient set obtained through physical calibration) is used to realize the matching enhancement of the wire pull amplitude and the local stress concentration area, and a physically weighted node wire pull response value matrix is obtained; Further, a three-dimensional deformation field reconstruction algorithm (based on trilinear interpolation and time sequence shape function) is used to realize the continuous spatial interpolation expression of the weighted node wire pull response at multiple times, and a time sequence three-dimensional deformation tensor field is generated; Further, a failure path search algorithm (Dijkstra path optimization, parameters: node failure probability vector ) is used to extract the maximum cumulative wire pull damage path in the three-dimensional deformation tensor field, and the node sequence set is output as the failure path data; Through a three-dimensional visualization mapping module, the failure path node sequence and the three-dimensional geometric model of the columnar structure are fused and rendered to form an explainable failure path atlas with spatial positioning ability, realizing the mapping from the prediction result to the explainable path in the structural physical coordinate system; For example, in a columnar solder joint monitoring task of a high-reliability communication device, the input three-dimensional geometric parameters are a node coordinate matrix with a size of 1500x3, an element connection matrix with a size of 2800x4, and the wire pull displacement prediction amplitude range is ~ mm, and the node stress weight coefficient set is ~ ​​Interval. The back-projection algorithm assigns each time step predicted displacement value to the corresponding spatial node, resulting in an initial distribution matrix size 1500 × 1 at time t0. After stress weight superposition processing, the node value at the local stress concentration site is improved by about 12%, and the value in the low stress area is attenuated by about 8%. The tensor field generated by the reconstruction of the three-dimensional deformation field Significant continuous deformation aggregation is formed at the location of the key welding point. The path length extracted by the failure path search algorithm on this tensor field is 12.5 mm, corresponding to 18 continuous nodes, and the cumulative wire-slip damage probability is The failure path explanation atlas rendered by the above is marked with a red path in the visualization platform, clearly marking the first failure location and the expansion direction, providing precise spatial location guidance for maintenance personnel. Through this method, the on-site maintenance time is shortened by about 41% compared to traditional manual troubleshooting, and the prediction accuracy is improved by about 16% compared to the no-mapping method; S8.4: Perform pattern matching algorithm based on the failure path explanation atlas and the historical wire-slip failure case database to identify the typical failure mode category to which the current prediction result belongs, in order to improve the explainability and engineering practicability of the prediction result; S8.5: Integrate the wire-slip failure trend curve, failure path explanation atlas, and failure mode recognition result into a structured prediction report, and output it to the electronic device lifecycle management system through the visualization interface, to assist in formulating preventive maintenance strategies.

[0021] Thus, the technical solutions of the present application have been described in combination with the preferred embodiments shown in the drawings, but it is easy for those skilled in the art to understand that the protection scope of the present application is obviously not limited to these specific embodiments. Those skilled in the art can make equivalent changes or replacements to related technical features without departing from the principles of the present application, and the technical solutions after these changes or replacements will fall within the protection scope of the present application.

[0022] The above only describes the preferred embodiments of the present application and is not intended to limit the present application; for those skilled in the art, the present application can have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and rules of the present application shall be included within the protection scope of the present application.

Claims

1. A method for predicting wire pull failure of a corrosion resistant circuit board pillar connection structure, characterized by, The method comprises the following steps: S1: Collecting the running data of the circuit board column connection structure under different environmental stress conditions and recording the corresponding environmental working condition labels; S2: Filtering and normalizing the collected structure stress response signal to generate a standardized structure stress sequence; S3: Based on the fusion of finite element simulation and standardized structure stress sequence, calculating the stress distribution cloud of each key node of the column connection structure, and extracting the time sequence evolution characteristics of the local maximum stress points; S4: Fitting the elastic modulus degradation curve according to the material aging experiment data, and combining the time-varying relationship of the interface friction coefficient to generate a multi-physical field coupling feature vector set; S5: Establishing a mapping function between the sliding trend characteristics and the material micro-damage mechanism, and mapping the multi-physical field coupling feature vector into a sliding trend feature vector with physical consistency; S6: Inputting the sliding trend feature vector into the prediction model with online parameter updating capability, and dynamically adjusting the model weight through the back propagation algorithm; S7: Judging whether the model prediction error exceeds the preset threshold, if yes, triggering the model parameter online optimization mechanism, adjusting the learning rate and updating the historical feature database; S8: Outputting the sliding failure trend curve based on the optimized prediction model, and generating the prediction result including the failure path explanation.

2. The method of claim 1, wherein the method further comprises: The step S1 specifically comprises: Based on the industrial sensor network, obtaining the multi-modal running data of the circuit board column connection structure under the action of multiple typical environmental stresses, the running data including the thermal field distribution signal output by the temperature sensor, the environmental humidity time series data collected by the humidity sensor, the vibration frequency spectrum information recorded by the three-axis accelerometer, and the structure stress response signal collected by the strain gauge, and constructing a multi-dimensional environmental-structure coupling data set; Performing frequency domain filtering processing on the structure stress response signal to obtain a filtered structure stress time series signal; Using the Z-score standardization method to normalize the filtered structure stress time series signal to generate a standardized structure stress sequence; Based on the control parameters of the environmental stress test box and the vibration bench, recording the environmental working condition labels corresponding to each set of collected data, and establishing the mapping relationship between the structure response and the environmental excitation; Timestamp aligning and data packaging the standardized structure stress sequence and the corresponding environmental working condition labels to generate a structured running data sample set.

3. The method of claim 1, wherein the method further comprises: The step S2 specifically comprises: Performing band-pass filtering processing on the collected original structure stress response signal to obtain a denoised structure stress time series signal; Performing multi-scale decomposition on the denoised structure stress time series signal based on the wavelet transform algorithm to extract wavelet coefficients at each scale; Performing mean normalization operation on the structure stress signal reconstructed from the wavelet coefficients to map the signal amplitude to a unified interval to form a normalized stress sequence; Based on the sliding window mechanism, segmenting the normalized stress sequence to calculate the mean, variance and kurtosis characteristics in each window to generate a structure stress feature sequence; Time sequence aligning the structure stress feature sequence and the corresponding environmental working condition labels to generate a joint data set of structure stress response and environmental stress conditions.

4. The method of claim 3, wherein the method further comprises: The filtering processing in the step S2 includes Butterworth low-pass filtering and FIR band-pass filtering, the lower limit frequency of the band-pass filtering is 1-10 Hz, the upper limit frequency is 100-500 Hz, the filter order is 64-256, and the normalized operation is performed on the structural stress signal after filtering and wavelet decomposition.

5. The method of claim 1, wherein the method further comprises: The step S3 specifically includes: The normalized structural stress sequence is subjected to spatial mapping processing, the corresponding relationship between the structural stress data and the geometric model of the column-connected structure is established, and the stress boundary condition input is generated; Based on the three-dimensional geometric model and material attribute parameters of the column-connected structure, a finite element analysis model is constructed, and the standardized stress boundary condition is imported, and static and dynamic coupled structural mechanics simulation is performed to obtain the stress response distribution of each key node under different environmental working conditions; The stress distribution data output by the finite element simulation is subjected to visual processing to generate stress nephograms of the column-connected structure at different time steps; Based on the stress nephograms, local maximum stress points in the column-connected structure are identified, and their stress change sequences in the time dimension are extracted to form time sequence evolution curves of the local maximum stress points; The time sequence evolution curves of the local maximum stress points are subjected to feature extraction processing, and key statistical features are extracted by using time-frequency analysis method to generate local stress evolution feature vectors.

6. The method of claim 5, wherein the method further comprises: In the step S3, the finite element analysis model is a three-dimensional CAD model, the mesh element type is a ten-node tetrahedral element, the average element length is 0.2-1 mm, the key area is refined to 0.2 mm, the material parameters include elastic modulus, Poisson's ratio, yield strength and density, the simulation calculation includes static-dynamic coupling analysis, and the three-dimensional visualization of the stress distribution is performed by using the equivalent surface rendering method and the principal stress extraction method.

7. The method of claim 1, wherein the method further comprises: determining a corrosion resistance of the circuit board post connection structure. The step S4 specifically includes: The data of the change of the elastic modulus with time collected in the material aging experiment are preprocessed to obtain a time sequence degradation sample set of the elastic modulus of the structural material; Based on the least square fitting algorithm, the time sequence degradation sample set is subjected to curve fitting to generate a degradation function expression of the change of the elastic modulus with time, and a stiffness degradation feature in the material aging process is constructed; Experimental data of the friction coefficient of the interface of the column-connected structure under different environmental stress actions are collected, and the friction coefficient evolution feature curve of the interface with time is generated through interpolation and smoothing processing; Based on the elastic modulus degradation function in the material aging process and the friction coefficient evolution curve, in combination with the structural stress distribution data, a time sequence feature matrix of multi-physical field coupling is constructed; The time sequence feature matrix is subjected to normalization and dimension reduction processing, key feature components are extracted by using the principal component analysis algorithm, and a multi-physical field coupling feature vector set is generated.

8. The method of claim 7, wherein the method further comprises: In the step S4, the elastic modulus degradation curve is fitted by using the polynomial least square method, the fitting order is 2-4 orders, the friction coefficient evolution curve is obtained by using multi-channel synchronous sampling, sliding mean, cubic Hermite interpolation and exponential smoothing method, and all feature data are subjected to dimension reduction processing by using the principal component analysis.

9. The method of claim 1, wherein the method further comprises: determining a corrosion resistance of the circuit board column connector structure. The step S5 specifically includes: The micro-damage mechanism of the material is classified and modeled, a damage evolution equation is constructed based on fracture mechanics and fatigue damage theory, and a damage degree evolution function of the material under different environmental stresses is obtained; Based on the multi-physics field coupling characteristic vector set, the elastic modulus degradation rate and the interface friction coefficient change rate are extracted as input variables into the damage evolution equation to calculate the equivalent damage degree of the material at the current time; A nonlinear mapping function between the wire slipping trend feature and the equivalent damage degree is constructed, and a support vector regression algorithm is used to train the function parameters; The time sequence evolution features of the local maximum stress points are combined with the mapping function to generate a wire slipping trend feature vector with time dimension; The wire slipping trend feature vector is normalized to generate a standardized wire slipping trend feature vector set.

10. The method of claim 9, wherein the method further comprises: In step S5, the damage evolution equation is a linear or nonlinear function, the input variables include the material elastic modulus degradation rate and the interface friction coefficient change rate, and the mapping equation parameters are calibrated based on actual material test data and trained through a support vector regression.

Images