A high voltage connector failure prediction method

Abstract

The application discloses a high-voltage connector fault prediction method, relates to the technical field of fault prediction, and solves the technical problems of incomplete fault mechanism coverage caused by the lack of multi-dimensional sensitive features, inaccurate thermoelectric coupling effect capture, unable to compensate signal transmission delay, lack of dynamic compensation mechanism, and low practicability of prediction results; the interference of environmental vibration is eliminated by extracting the resonance frequency offset of the vibration signal through FFT transformation; a temperature change rate correction term is introduced into the temperature resistance coupling coefficient; the linear coupling features output by the bivariate LSTM and the nonlinear coupling features extracted by the KPLS are combined to construct a thermoelectric coupling feature set, and full-stage coverage of the thermoelectric coupling effect is realized. The improved Transformer model quickly adapts to a new scene through small sample transfer learning. Based on the dynamic compensation coefficient output by the fusion features, the signal transmission delay under the conditions of rapid temperature change and electromagnetic interference can be corrected in real time.

Description

Technical Field

[0001] This invention belongs to the field of fault prediction, specifically a fault prediction method for high-voltage connectors. Background Technology

[0002] High-voltage connectors are core electrical connection components specifically designed for connecting high-voltage power systems and transmitting high-voltage (typically ≥600V AC or ≥1000V DC) and high-current electrical energy. They also function as signal transmission, mechanical fixation, and safety protection devices. Their core role is to enable detachable / separable connections between devices and between devices and cables in high-voltage systems. They must ensure the stability and reliability of high-voltage and high-current transmission while employing insulation, shielding, and locking structural designs to prevent safety risks such as leakage, arcing, and electromagnetic interference. High-voltage connectors are crucial components in fields such as new energy vehicles and power systems; their operational reliability directly determines the safety and stability of the system.

[0003] Traditional coupling coefficients are only applicable to steady-state temperatures and cannot capture abrupt changes in coupling relationships under conditions of drastic temperature fluctuations. The lack of multi-dimensional sensitive features leads to incomplete coverage of fault mechanisms, inaccurate capture of thermoelectric coupling effects, inability to compensate for signal transmission delays, lack of dynamic compensation mechanisms, and low practicality of prediction results. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art; to this end, this invention proposes a high-voltage connector fault prediction method to solve the technical problems of incomplete coverage of fault mechanisms due to the lack of multi-dimensional sensitive features, inaccurate capture of thermoelectric coupling effects, inability to compensate for signal transmission delay, lack of dynamic compensation mechanism, and low practicality of prediction results.

[0005] To address the above problems, a first aspect of the present invention provides a high-voltage connector fault prediction method, comprising the following steps: Based on the electrical signal and mechanical and bolt condition data of high-voltage connectors, sensitive features of bolt fastening failure are extracted from three dimensions: torque change, gap state, and vibration response, combined with the degree of corrosion. Based on the electrical and thermal signals at various locations of the high-voltage connector, and combined with a supplementary temperature change rate correction term, the optimized temperature resistance coupling coefficient is calculated. An optimized signal transmission delay prediction model is constructed, employing a bivariate long short-term memory network. Based on the dynamic correlation characteristics between temperature, current, and signal transmission delay, linear coupling characteristics are output. Nonlinear coupling characteristics between temperature fluctuations, current fluctuations, and contact resistance changes are extracted using a kernel partial least squares algorithm. A thermoelectric coupling feature set is constructed by optimizing the temperature resistance coupling coefficient, linear coupling characteristics, and nonlinear coupling characteristics. Based on temperature cycling data and material properties, we analyze the sensitive characteristics of thermal fatigue life. The LightGBM model is used to score the importance of features and to fuse features sensitive to bolt fastening failure, thermoelectric coupling feature set, and thermal fatigue life. An improved Transformer prediction and lifetime assessment model is established. Fault prediction and lifetime prediction are performed through feature fusion. Based on the optimized signal transmission delay prediction model, dynamic compensation coefficients are generated to compensate the original signal in real time.

[0006] Optionally, in one example of the above aspects, sensitive features of bolt fastening failure are extracted from three dimensions: torque variation, clearance state, and vibration response, combined with the degree of corrosion, including the following steps: From the three dimensions of torque change, gap state and vibration response, combined with the degree of corrosion, torque-related features, gap-resistance correlation features and vibration-corrosion features are extracted as sensitive features of bolt fastening failure. Based on the torque change data, the torque attenuation rate is calculated as a torque-related characteristic: εt=(Bt0-Btt) / Bt0; Where εt is the torque decay rate, Bt0 is the initial tightening torque of the bolt, and Btt is the real-time monitored torque at time t; when εt≥8%, it is determined to be a torque decay warning state, otherwise it is not determined to be a torque decay warning state. Based on the gap between the bolt and the copper busbar, and combined with the resistance of the copper busbar circuit, the relationship between the gap and the resistance is calculated. The vibration signal of the bolt area is subjected to FFT transformation to extract the resonant frequency offset Δf. The corrosion depth hco is obtained by combining the corrosion sensor data. Based on the synergistic acceleration effect of vibration and corrosion, two types of coupling features are constructed, including product coupling features and normalized coupling features. The product-type coupling feature F1 = Δf * hco, where Δf is the currently detected bolt resonance frequency offset and hco is the currently detected bolt corrosion depth threshold. The normalized coupling feature F2 = (Δf / Δfmax)*0.6 + (hco / hcomax)*0.4, where Δfmax is the critical resonant frequency offset of bolt failure and hcomax is the safe corrosion depth threshold of bolt. Vibration-corrosion coupling features are constructed using product-type coupling features and normalized coupling features.

[0007] Optionally, in one example of the above aspects, the optimized temperature resistance coupling coefficient is calculated based on the electrical and thermal signals at various locations of the high-voltage connector, combined with a supplementary temperature change rate correction term, including the following steps: Calculate the temperature-resistivity coupling coefficient based on Joule's law and the heat transfer equation. ; Incorporating the temperature change rate correction term, the formula for optimizing the temperature resistance coupling coefficient is established as follows: in, The optimized temperature-resistance coupling coefficient is given by β, where β is the temperature rate correction coefficient. The rate of temperature change.

[0008] Optionally, in one example of the above aspects, constructing an optimized signal transmission delay prediction model includes the following steps: Calculate the delay coupling coefficient of electromagnetic interference signals based on the principle of electromagnetic induction. ; Based on the combined effects of temperature and electromagnetic interference, a coupling correction factor is added to optimize the signal transmission delay prediction model: in, For the optimized signal transmission delay, For signal transmission delay under healthy conditions, The optimized temperature resistance coupling coefficient, This represents the average real-time temperature at various locations inside the high-voltage connector. This is the average real-time temperature of various locations inside the high-voltage connector during normal operation at a standard ambient temperature of T=25℃.

[0009] Optionally, in one example of the above aspects, a bivariate long short-term memory network is employed, based on the dynamic correlation characteristics of temperature, current, and signal transmission delay, including the following steps: The current of the high-voltage connector and the real-time temperature of various internal locations are selected as input variables. The signal transmission delay of the optimized signal transmission delay prediction model is used as the target variable. Time series sample pairs are constructed: the data is divided according to time windows and the input data is standardized. A Bi-LSTM bivariate long short-term memory network was constructed, and the input layer dimension was set according to the standardized temperature and current time series. The input layer is connected to a bidirectional LSTM layer, which has two hidden layers: a forward LSTM and a backward LSTM to capture the forward and backward dependencies of the time series data, respectively. The bidirectional LSTM layer is connected to the fully connected layer, which maps the spliced ​​features output by the bidirectional LSTM to a delayed prediction sequence with the same time step as the input. The fully connected layer connects to the output layer and outputs the predicted delay value for each time step; The output of the last hidden state of the trained Bi-LSTM is used as a dynamic correlation feature vector of temperature-current-delay.

[0010] Optionally, in one example of the above aspects, outputting linear coupling features includes the following steps: Based on the dynamic correlation feature vector of temperature-current-delay analyzed by Bi-LSTM bivariate long short-term memory network, the coupling feature deviation value ΔF is calculated: ; Where Fcurrent is the dynamic correlation feature vector of temperature-current-delay for the current sample, Fhealthy is the dynamic correlation feature vector of temperature-current-delay for the high-voltage connector under healthy conditions, and the output coupling feature deviation value ΔF is used as the linear coupling feature.

[0011] Optionally, in one example of the above aspects, the nonlinear coupling characteristics of temperature fluctuations, current fluctuations, and contact resistance changes are extracted using a kernel partial least squares algorithm, including the following steps: The nonlinear coupling characteristics of temperature fluctuation, current fluctuation and contact resistance change are extracted using the KPLS kernel partial least squares algorithm. By calculating the average real-time temperature at various locations inside the high-voltage connector The average real-time temperature of various locations inside the high-voltage connector under normal operating conditions at a standard ambient temperature of T=25℃. The difference between the real-time current of the high-voltage connector and the average current under healthy conditions is used as the change in temperature fluctuation. The temperature fluctuation coefficient and the current fluctuation coefficient are set as input variables to form a 2D input matrix. The output variable is set to the contact resistance change ΔRc=Rc-Rc0, where Rc0 is the contact resistance in the healthy state and Rc is the contact resistance detected in real time by the high-voltage connector. Collect historical data, construct a sample set of input variables and corresponding output variables, and train the KPLS kernel partial least squares algorithm; Extract the top principal components that are most strongly correlated with the output variable, and use these extracted principal components to construct a nonlinear coupling feature vector, which serves as the nonlinear coupling feature vector. Optionally, in one example of the above aspects, the analysis of thermal fatigue life sensitivity characteristics based on temperature cycling data and material properties includes the following steps: Analyze temperature cycling characteristics and fatigue damage characteristics as sensitive features of thermal fatigue life; Calculate the thermal stress of the copper busbar and insulating components based on the material's coefficient of thermal expansion and elastic modulus. , as a characteristic of thermal stress; Thermal stress characteristics, as well as the temperature cycling amplitude, cycling frequency, and dwell time at the high temperature threshold extracted from the high-voltage connector, are used as temperature cycling characteristics. Using Miner's linear cumulative damage theory, combined with a correction coefficient, the cumulative thermal fatigue damage value of insulation and metal components is quantified and used as a fatigue damage characteristic. When the cumulative thermal fatigue damage value of insulation and metal components exceeds a threshold, it is judged as a high-risk warning for thermal fatigue failure; otherwise, it is not judged as a high-risk warning for thermal fatigue failure.

[0012] Optionally, in one example of the above aspects, feature importance scoring is performed using the LightGBM model, and feature fusion is performed on bolt fastening failure sensitive features, thermoelectric coupling feature sets, and thermal fatigue life sensitive features, including the following steps: Construct feature set matrices and fault label vectors for bolt fastening failure sensitive features, thermo-electric coupling feature sets, and thermal fatigue life sensitive features; calculate the MIC value of each type of feature and fault label vector using the maximum mutual information coefficient, and retain features with MIC values ​​greater than the threshold. The selected bolt fastening failure sensitive features, thermo-electric coupling feature set and thermal fatigue life sensitive features are input into the LightGBM classification model, and the feature contribution is automatically evaluated during the model training process. The LightGBM classification model parameters are set as follows: learning rate = 0.01, number of decision trees = 100, maximum tree depth = 6, number of leaf nodes = 32. After training, the gain value of each feature is extracted as the gain importance, and the core features with gain importance greater than or equal to the threshold are retained. The retained bolt fastening failure sensitive features, thermo-electric coupling feature set and thermal fatigue life sensitive features are divided into different domains according to each feature. Attention weights are calculated separately for feature subsets of each domain. Feature scores are generated by tanh activation function and converted into normalized weights by Softmax function. The features within a domain are weighted and summed according to their weights to generate domain-level fusion features. The fusion features from all domains are then concatenated to form the final fusion feature vector.

[0013] Optionally, in one example of the above aspects, an improved Transformer prediction and lifetime assessment model is established. Fault prediction and lifetime prediction are performed through feature fusion. Based on an optimized signal transmission delay prediction model, dynamic compensation coefficients are generated to compensate the original signal in real time, including the following steps: A Transformer model is established, and the input layer divides the fused feature sequence into time windows to construct temporal input samples; physical temporal enhancement position coding is introduced, and position codes are dynamically generated by combining sampling interval and working condition cycle. The multi-branch output layer is configured with parallel output branches, including: fault prediction branch, thermal fatigue life prediction branch, and signal delay compensation branch; The fault prediction branch outputs the probability of typical bolt-related faults and the bolt life prediction. Thermal fatigue life prediction of the remaining life of the branch output high voltage connector; Filter historical data for a preset time period before the high-voltage connector fails and becomes unusable, and for a preset time period before typical bolt-related failures occur, and generate a fused feature vector; The specific time period before the failure to use the connector is marked as the remaining life of the high-voltage connector, and typical bolt-related failures are marked. The specific time period before the occurrence of typical bolt-related failures is marked as the bolt life. The Transformer model is trained using labeled data to obtain an improved Transformer prediction and life assessment model. Based on the fused feature vector of the current high-voltage connection monitoring data, the prediction results are output through the fault prediction branch and the thermal fatigue life prediction branch. The signal delay compensation branch of the improved Transformer prediction and lifetime assessment model is based on the optimized delay prediction model. It outputs the predicted signal transmission delay value and generates dynamic compensation coefficients through the GRU submodule to perform real-time compensation on the original signal. The compensation formula is as follows: Where Kcomp is the dynamic compensation coefficient, Kcomp= , This is the baseline delay value.

[0014] Compared with the prior art, the beneficial effects of the present invention are: This invention extracts the resonant frequency offset from the vibration signal via FFT transformation, eliminating interference from environmental vibrations. The corrosion depth is quantitatively calculated based on Faraday's law, avoiding subjective judgment errors and exhibiting high feature stability. It accurately quantifies the thermo-electric-magnetic multi-field coupling effect, improving adaptability to complex operating conditions. A temperature change rate correction term is introduced into the temperature-resistance coupling coefficient, solving the problem that traditional coupling coefficients are only applicable to steady-state temperatures and cannot capture abrupt changes in coupling relationships under conditions of drastic temperature fluctuations. The linear coupling features output by the bivariate LSTM accurately capture the temporal dynamic correlation between temperature, current, and signal transmission delay, suitable for characterizing slowly varying coupling effects. The nonlinear coupling features extracted by KPLS effectively uncover the strong nonlinear mapping relationship between temperature fluctuations, current fluctuations, and contact resistance changes, suitable for characterizing abrupt coupling effects. The combined thermo-electric coupling feature set achieves full-stage coverage of the thermo-electric coupling effect.

[0015] This invention employs multi-task parallel prediction, covering the entire lifecycle requirements. It improves the Transformer design by calculating fault probability, remaining lifetime, and latency compensation in three parallel output branches, all based on fused features. This integrated approach combines fault warning, lifetime assessment, and performance optimization, meeting the full lifecycle management needs of high-voltage connectors. Through transfer learning, it achieves strong adaptability and reduces engineering implementation costs. Addressing the characteristic differences of different high-voltage connector models, the improved Transformer can quickly adapt to new scenarios using core features fused with LightGBM and small-sample transfer learning. Based on the dynamic compensation coefficients output from the fused features, it can correct signal transmission delays under drastic temperature changes and electromagnetic interference in real time, effectively improving the signal transmission reliability of high-voltage connectors. This is particularly suitable for autonomous driving and smart charging scenarios where high signal stability is required. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0018] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1 The first aspect of this invention provides a method for predicting high-voltage connector faults, comprising the following steps: In this embodiment, a five-dimensional data acquisition system is constructed to achieve comprehensive and real-time data acquisition of the connector's operating status and the status of bolts and copper busbars; Electrical signal acquisition: A high frame rate current / voltage sensor is used to acquire operating current, voltage fluctuation and signal transmission delay data; a high-precision resistance tester is used to acquire contact resistance, insulation resistance and copper busbar circuit resistance in real time, and to capture resistance changes caused by poor contact, coupling effect and loose copper busbar; Thermal signal acquisition: Distributed fiber optic temperature measurement and infrared thermal imaging are used to simultaneously acquire data, accurately capturing the real-time temperature and temperature change rate of high-voltage connector contact points, shell, copper busbars and insulating components; Mechanical and Bolt Status Acquisition: A new miniature torque sensor is added to monitor bolt torque in real time; an ultrasonic sensor is used to detect the gap between the bolt and the copper busbar; an accelerometer is used to collect vibration signals in the bolt area; and a corrosion sensor is deployed to monitor the degree of bolt corrosion, comprehensively capturing early signs of bolt fastening failure. Electromagnetic signal acquisition: High-frequency electromagnetic sensors acquire the intensity and frequency distribution of electromagnetic interference to quantify the impact of electromagnetic interference on signal transmission; Environmental signal acquisition: Simultaneously acquire ambient temperature to provide data support for adaptive correction of operating conditions.

[0020] Based on the electrical signal and mechanical and bolt condition data of high-voltage connectors, torque-related features, gap and fit features, and vibration and corrosion features are extracted from three dimensions: torque change, gap condition, and vibration response, combined with the degree of corrosion, as sensitive features for bolt fastening failure. Based on the electrical and thermal signals at various locations of the high-voltage connector, and combined with a supplementary temperature change rate correction term, the optimized temperature resistance coupling coefficient is calculated. An optimized signal transmission delay prediction model is constructed, employing a bivariate long short-term memory network. Based on the dynamic correlation characteristics between temperature, current, and signal transmission delay, linear coupling characteristics are output. Nonlinear coupling characteristics between temperature fluctuations, current fluctuations, and contact resistance changes are extracted using a kernel partial least squares algorithm. A thermoelectric coupling feature set is constructed by optimizing the temperature resistance coupling coefficient, linear coupling characteristics, and nonlinear coupling characteristics. Based on temperature cycling data and material properties, we analyze the sensitive characteristics of thermal fatigue life. The LightGBM model is used to score the importance of features and to fuse features sensitive to bolt fastening failure, thermoelectric coupling feature set, and thermal fatigue life. An improved Transformer prediction and lifetime assessment model is established. Fault prediction and lifetime prediction are performed through feature fusion. Based on the optimized signal transmission delay prediction model, dynamic compensation coefficients are generated to compensate the original signal in real time.

[0021] Specifically, in this embodiment, features are constructed from four dimensions: torque change, gap state, vibration response, and corrosion degree. This comprehensively covers the chain evolution process of bolt fastening failure, including torque attenuation, gap increase, vibration frequency shift, and accelerated corrosion, avoiding the omission of the synergistic effect of loosening and corrosion by a single feature. Torque attenuation rate, gap-resistance correlation feature, and vibration-corrosion coupling feature are all sensitive indicators of precursors to bolt failure, which can issue early warnings before serious accidents such as sudden temperature rise of copper busbar and insulation carbonization occur, solving the pain point of traditional methods that only discover failures after they occur. The vibration signal is extracted for resonant frequency shift through FFT transformation, eliminating the interference of environmental vibration; the corrosion depth is quantitatively calculated based on Faraday's law, avoiding subjective judgment errors and ensuring high feature stability.

[0022] Precise quantification of the multi-field coupling effect of thermo-electricity-magnetism enhances adaptability to complex operating conditions. By introducing a temperature change rate correction term into the temperature resistance coupling coefficient, the problem that traditional coupling coefficients are only applicable to steady-state temperatures and cannot capture sudden changes in coupling relationships under conditions of drastic temperature fluctuations is solved. This is especially suitable for dynamic temperature change scenarios during the driving / charging of new energy vehicles.

[0023] The linear coupling features of the bivariate LSTM output accurately capture the temporal dynamic correlation between temperature, current, and signal transmission delay, making it suitable for characterizing slowly varying coupling effects. The nonlinear coupling features extracted by KPLS effectively uncover the strong nonlinear mapping relationship between temperature fluctuations, current fluctuations, and contact resistance changes, making it suitable for characterizing abrupt coupling effects. The thermoelectric coupling feature set constructed by combining the two achieves "full-scenario, full-stage" coverage of thermoelectric coupling effects.

[0024] By employing an optimized signal transmission delay prediction model, accurate delay prediction and compensation under temperature-electromagnetic interference superposition conditions are achieved based on coupling characteristics. This not only helps predict faults but also directly improves signal transmission reliability, offering the dual value of fault early warning and performance optimization. It also enables full lifecycle assessment, supporting predictive maintenance. Thermal fatigue characteristics, constructed based on temperature cycling amplitude, thermal stress, and Miner cumulative damage, are directly related to the material fatigue mechanism of connector copper busbars and insulating components, rather than relying on empirical statistics. Thermal fatigue characteristics can quantitatively assess the degree of damage accumulation of connectors under long-term temperature cycling, providing data support for the development of on-demand maintenance strategies and avoiding the problems of over-maintenance or under-maintenance in traditional periodic maintenance.

[0025] By employing MIC filtering and LightGBM importance screening, redundant features are eliminated while core features are retained, reducing the computational complexity of subsequent Transformer models and improving training efficiency. Key fault features such as torque attenuation rate and thermal stress are assigned 20%-50% weight gain, guiding the model to focus on core failure indicators, avoiding interference from irrelevant features, and significantly improving early fault identification rates.

[0026] Bolt mechanical features, thermoelectric coupling features, and thermal fatigue features belong to the three major physical domains of mechanics, electricity, and heat, respectively, and suffer from strong heterogeneity and redundancy in some features. LightGBM quantifies the contribution of each feature to the fault prediction task through gain importance scoring, selects core features with a weight ≥ 0.05, and simplifies the high-dimensional feature space to a low-dimensional effective space, significantly reducing the computational complexity of the subsequent Transformer model and improving training and inference efficiency.

[0027] Traditional feature fusion methods often employ simple concatenation to achieve cross-domain feature collaboration, which can easily lead to an imbalance in the weights of mechanical and thermoelectric features. During training, LightGBM can automatically learn the correlation between features from different domains, enabling the fused features to not only contain failure information from a single domain but also reflect the failure evolution patterns of multiple coupled factors, providing highly discriminative input for the Transformer model.

[0028] With strong anti-interference capabilities and adaptability to complex operating conditions, LightGBM is based on a gradient boosting decision tree framework. It has natural robustness to noisy data and outliers, and can effectively filter feature noise caused by electromagnetic interference, environmental vibration, etc., to ensure the stability of fused features. It is especially suitable for the dynamic operating conditions of new energy vehicles during driving / charging.

[0029] By improving the Transformer model's prediction capabilities, the accuracy and timeliness of fault and lifespan predictions are enhanced. Long-term dependencies are captured, enabling early fault warnings. High-voltage connector faults are a gradual evolutionary process, and pre-fault warning signals exhibit long-term temporal correlations. The improved Transformer attention mechanism accurately captures the long-term temporal dependencies of feature sequences. Combined with physical prior constraints, this avoids the model focusing on irrelevant temporal segments, thereby improving the early fault identification rate.

[0030] Multi-task parallel prediction covers the entire lifecycle requirements. The Transformer design improves the three parallel output branches of fault probability, remaining lifespan, and delay compensation, which are completed simultaneously based on fused features. This integrates fault warning, lifespan assessment, and performance optimization to meet the full lifecycle management needs of high-voltage connectors.

[0031] Transfer learning has strong adaptability and reduces engineering implementation costs. To address the differences in features of different high-voltage connector models, the improved Transformer can quickly adapt to new scenarios based on the core features fused by LightGBM through small-sample transfer learning.

[0032] Traditional fault prediction methods only focus on "whether a fault occurs", while this embodiment, based on the dynamic compensation coefficient of the fused feature output, can correct the signal transmission delay under drastic temperature changes and electromagnetic interference in real time, effectively improving the signal transmission reliability of high voltage connectors, and is especially suitable for autonomous driving and smart charging scenarios with high requirements for signal stability.

[0033] In one embodiment of the present invention, sensitive features of bolt fastening failure are extracted from three dimensions: torque change, gap state, and vibration response, combined with the degree of corrosion, including the following steps: From the three dimensions of torque change, gap state and vibration response, combined with the degree of corrosion, torque-related features, gap-resistance correlation features and vibration-corrosion features are extracted as sensitive features of bolt fastening failure. Based on the torque change data, the torque attenuation rate is calculated as a torque-related characteristic: εt=(Bt0-Btt) / Bt0; Where εt is the torque decay rate, Bt0 is the initial tightening torque of the bolt (factory calibration value), and Btt is the real-time monitored torque at time t. When εt ≥ 8%, it is determined to be a torque decay warning state, at which point the risk of copper busbar loosening is significantly increased; otherwise, it is not determined to be a torque decay warning state. Based on the gap between the bolt and the copper busbar, and considering the resistance of the copper busbar circuit, calculate the relationship between the gap and the resistance: in, These are the characteristic coefficients relating gap and resistance. The resistance of the copper busbar circuit under healthy conditions, in mΩ. The resistance of the copper busbar circuit being tested is currently being measured. The gap between the bolt and the copper busbar under healthy conditions, in mm. This is the current gap between the bolt and the copper busbar. This reflects the increase in resistance caused by the increase in unit gap. When the resistance is ≥6 mΩ / mm, it is determined that the copper busbar is loose and an alarm is triggered; otherwise, no alarm is triggered. The vibration signal of the bolt area is subjected to FFT transformation to extract the resonant frequency offset Δf. The corrosion depth hco is obtained by combining the corrosion sensor data. Based on the synergistic acceleration effect of vibration and corrosion, two types of coupling features are constructed, including product coupling features and normalized coupling features. The product-type coupling feature F1 = Δf * hco, where Δf is the currently detected bolt resonance frequency shift and hco is the currently detected bolt corrosion depth threshold, directly characterizing the superimposed effect of the two. The normalized coupling feature F2 = (Δf / Δfmax)*0.6 + (hco / hcomax)*0.4, where Δfmax is the critical resonance frequency offset of bolt failure (experimentally calibrated), hcomax is the bolt safe corrosion depth threshold, set according to the design life, the weighting coefficient is determined by mutual information coefficient optimization, and F2 comprehensively reflects the bolt failure risk level.

[0034] Vibration-corrosion coupling features are constructed using product-type coupling features and normalized coupling features.

[0035] In one embodiment of the present invention, based on the electrical and thermal signals at various locations of the high-voltage connector, and in conjunction with a supplementary temperature change rate correction term, the optimized temperature resistance coupling coefficient is calculated, including the following steps: Calculate the temperature-resistivity coupling coefficient based on Joule's law and the heat transfer equation. The formula is: in, The contact resistance of the high-voltage connector is expressed in mΩ. This represents the average real-time temperature at various locations inside the high-voltage connector, in degrees Celsius (°C). The temperature coefficient of resistance of copper is set to 3.9 × 10⁻⁶ / ℃ in this embodiment. The contact resistance of a high-voltage connector under normal operating conditions at a standard ambient temperature of T=25℃. This is the average real-time temperature of various locations inside the high-voltage connector during normal operation at a standard ambient temperature of T=25℃. Incorporating the temperature change rate correction term, the formula for optimizing the temperature resistance coupling coefficient is established as follows: in, The optimized temperature resistance coupling coefficient is given by β, which is the temperature rate correction coefficient. In this embodiment, the value is experimentally calibrated to a range of 0.02-0.05 s / ℃. The rate of temperature change is expressed in °C / s.

[0036] In one embodiment of the present invention, constructing an optimized signal transmission delay prediction model includes the following steps: Based on the principle of electromagnetic induction, calculate the delay coupling coefficient of electromagnetic interference signals: in, The electromagnetic interference signal delay coupling coefficient is given by Δτ, where Δτ is the change in signal transmission delay in nanoseconds. The change in electromagnetic interference intensity is expressed in V / m. Based on the combined effects of temperature and electromagnetic interference, a coupling correction factor is added to optimize the signal transmission delay prediction model: in, For the optimized signal transmission delay, For signal transmission delay under healthy conditions, The optimized temperature resistance coupling coefficient, This represents the average real-time temperature at various locations inside the high-voltage connector. This is the average real-time temperature of various locations inside a normally operating high-voltage connector at a standard ambient temperature of T=25℃; enabling accurate prediction of delay under conditions of drastic temperature changes and electromagnetic interference.

[0037] In one embodiment of the present invention, a bivariate long short-term memory network is employed, based on the dynamic correlation characteristics of temperature, current, and signal transmission delay, including the following steps: The current of the high-voltage connector and the real-time temperature of various internal locations are selected as input variables. The signal transmission delay of the optimized signal transmission delay prediction model is used as the target variable. Time series sample pairs are constructed: the data is divided according to the time window and the input data is standardized. In this embodiment, the window size is 20 sampling periods, the step size is 5, and each sample contains 20 sets of input sequences and corresponding 20 target sequences (i=1,2,...,20).

[0038] A Bi-LSTM bivariate long short-term memory network is constructed, and the input layer dimension is set according to the standardized temperature and current time series. In this embodiment, the input layer dimension is set to a time step of 20 and a feature number of 2.

[0039] The input layer is connected to a bidirectional LSTM layer, which has two hidden layers. The forward LSTM and backward LSTM capture the forward and backward dependencies of the time series data, respectively. The number of neurons in each hidden layer is 64, the activation function is tanh, and the dropout rate is 0.2 to prevent overfitting.

[0040] The bidirectional LSTM layer is connected to the fully connected layer, which maps the spliced ​​features output by the bidirectional LSTM to a delayed prediction sequence with the same time step as the input. The fully connected layer connects to the output layer and outputs the predicted delay value for each time step; The output of the last hidden state of the trained Bi-LSTM is used as a dynamic correlation feature vector of temperature-current-delay.

[0041] In this embodiment, the mean squared error of MSE is used as the loss function, and the learning rate of the Adam optimizer is set to 0.001.

[0042] In one embodiment of the present invention, outputting linear coupling features includes the following steps: Based on the dynamic correlation feature vector of temperature-current-delay analyzed by Bi-LSTM bivariate long short-term memory network, the coupling feature deviation value ΔF is calculated: ; Wherein, Fcurrent is the dynamic correlation feature vector of temperature-current-delay for the current sample, and Fhealthy is the dynamic correlation feature vector of temperature-current-delay for the high-voltage connector under healthy conditions; the output coupling feature deviation value ΔF is used as the linear coupling feature. In this embodiment, Fhealthy is a baseline vector obtained by training with healthy data, and ΔF represents the degree of deviation between the current coupling state and the healthy state.

[0043] In one embodiment of the present invention, the nonlinear coupling characteristics of temperature fluctuation, current fluctuation and contact resistance change are extracted using a kernel partial least squares algorithm, including the following steps: The nonlinear coupling characteristics of temperature fluctuation, current fluctuation and contact resistance change are extracted using the KPLS kernel partial least squares algorithm. By calculating the average real-time temperature at various locations inside the high-voltage connector The average real-time temperature of various locations inside the high-voltage connector under normal operating conditions at a standard ambient temperature of T=25℃. The difference between the real-time current of the high-voltage connector and the average current under healthy conditions is used as the change in temperature fluctuation. The temperature fluctuation coefficient and the current fluctuation coefficient are set as input variables to form a 2D input matrix. The output variable is set to the contact resistance change ΔRc=Rc-Rc0, where Rc0 is the contact resistance in the healthy state and Rc is the contact resistance detected in real time by the high-voltage connector. Collect historical data, construct a sample set of input variables and corresponding output variables, and train the KPLS kernel partial least squares algorithm; Extract the top principal components that are most strongly correlated with the output variable, and use these extracted principal components to construct a nonlinear coupling feature vector, which serves as the nonlinear coupling feature vector. In this embodiment, the radial basis function (RBF) is selected as the kernel function. The input X is mapped to a high-dimensional feature space through the kernel function, and then partial least squares regression is performed in the high-dimensional space to extract the top three principal components with the strongest correlation to the output variable, or the principal components with a cumulative contribution rate ≥ 95%. The extracted principal components are used to form a nonlinear coupled feature vector, which serves as the input feature for subsequent cross-domain fusion.

[0044] In one embodiment of the present invention, the analysis of thermal fatigue life sensitivity characteristics based on temperature cycling data and material properties includes the following steps: Analyze temperature cycling characteristics and fatigue damage characteristics as sensitive features of thermal fatigue life; Calculate the thermal stress of the copper busbar and insulating components based on the material's coefficient of thermal expansion and elastic modulus. As a characteristic of thermal stress, the formula is: Where E is the elastic modulus of the material; in this embodiment, the copper busbar has a modulus of 110 GPa, and the insulating shell has a modulus of 3.5 GPa. The coefficient of thermal expansion of the material is 16.5 × 10⁻⁶. In this embodiment, the copper busbar has a coefficient of thermal expansion of 16.5 × 10⁻⁶. -6 / ℃, Insulating shell: 75×10 -6 / ℃, μ is Poisson's ratio, in this embodiment, copper busbar: 0.34, insulating shell: 0.38, Temperature cycling amplitude; thermal stress of copper busbars and insulating components. When the insulation aging risk exceeds 30% of the material's yield strength, it is determined to be in an insulation aging risk warning state; otherwise, it is not determined to be in an insulation aging risk warning state; the risk of insulation aging and deformation increases dramatically. The thermal stress characteristics, as well as the temperature cycle amplitude ΔTt=Ttmax-Ttmin, cycle frequency fcu=1 / total duration of a single temperature cycle, and dwell time at the high temperature threshold of the high voltage connector are extracted as temperature cycle characteristics; where Ttmax is the highest temperature in a single cycle and Ttmin is the lowest temperature in a single cycle. Using Miner's linear cumulative damage theory, combined with a correction coefficient, the cumulative thermal fatigue damage value of insulation and metal components is quantified and used as a fatigue damage characteristic. When the cumulative thermal fatigue damage value of insulation and metal components exceeds a threshold, it is judged as a high-risk warning for thermal fatigue failure; otherwise, it is not judged as a high-risk warning for thermal fatigue failure.

[0045] In this embodiment, the cumulative thermal fatigue damage of the insulating and metal components is calculated using the following formula: in, This is due to the cumulative damage from thermal fatigue in insulating and metallic components. Let j be the actual number of cycles under temperature cycling conditions. The limit number of cycles at which fatigue failure occurs under the j-th working condition is determined by the SN curve. The condition correction factor is k for the j-th working condition. Under high temperature and high humidity conditions, k = 1.2-1.5. When D ≥ 0.8, it is judged to be at high risk of thermal fatigue failure, and protection strategies need to be activated.

[0046] In one embodiment of the present invention, the LightGBM model is used to score the importance of features, and feature fusion is performed on bolt fastening failure sensitive features, thermoelectric coupling feature sets, and thermal fatigue life sensitive features, including the following steps: Construct feature set matrices and fault label vectors for bolt fastening failure sensitive features, thermo-electric coupling feature sets, and thermal fatigue life sensitive features; calculate the MIC value of each type of feature and fault label vector using the maximum mutual information coefficient, and retain features with MIC values ​​greater than a threshold; in this embodiment, the threshold for MIC values ​​is set to 0.3. The selected bolt fastening failure sensitive features, thermo-electric coupling feature set and thermal fatigue life sensitive features are input into the LightGBM classification model, and the feature contribution is automatically evaluated during the model training process. The LightGBM classification model parameters are set as follows: learning rate = 0.01, number of decision trees = 100, maximum tree depth = 6, number of leaf nodes = 32. After training, the gain value of each feature is extracted as the gain importance, and the core features with gain importance greater than or equal to the threshold are retained. The Gain value represents the total contribution of the feature to the reduction of model error. The gain importance threshold is set to 0.05, and core features with Gain importance ≥ 0.05 are retained, such as torque attenuation rate, thermal stress, vibration-corrosion coupling feature F2, etc. The retained bolt fastening failure sensitive features, thermo-electric coupling feature set and thermal fatigue life sensitive features are divided into different domains according to each feature. Attention weights are calculated separately for feature subsets of each domain. Feature scores are generated by tanh activation function and converted into normalized weights by Softmax function. The features within a domain are weighted and summed according to their weights to generate domain-level fusion features. The fusion features from all domains are then concatenated to form the final fusion feature vector.

[0047] In one embodiment of the present invention, an improved Transformer prediction and lifetime assessment model is established. Fault prediction and lifetime prediction are performed through feature fusion. Based on an optimized signal transmission delay prediction model, dynamic compensation coefficients are generated to compensate the original signal in real time. This includes the following steps: A Transformer model is established. The input layer divides the fused feature sequence into time windows. In this embodiment, the window size is 20 sampling periods and the step size is 5 to construct the temporal input samples. Physical temporal enhancement position coding is introduced, and position coding is dynamically generated by combining the sampling interval and the working condition cycle. The multi-branch output layer is configured with parallel output branches, including: fault prediction branch, thermal fatigue life prediction branch, and signal delay compensation branch; The fault prediction branch outputs the probability of typical bolt-related faults and the bolt life prediction. Thermal fatigue life prediction of the remaining life of the branch output high voltage connector; Filter historical data for a preset time period before the high-voltage connector fails and becomes unusable, and for a preset time period before typical bolt-related failures occur, and generate a fused feature vector; The specific time period before the failure to use the connector is marked as the remaining life of the high-voltage connector, and typical bolt-related failures are marked. The specific time period before the occurrence of typical bolt-related failures is marked as the bolt life. The Transformer model is trained using labeled data to obtain an improved Transformer prediction and life assessment model. Based on the fused feature vector of the current high-voltage connection monitoring data, the prediction results are output through the fault prediction branch and the thermal fatigue life prediction branch. The signal delay compensation branch of the improved Transformer prediction and lifetime assessment model is based on the optimized delay prediction model. It outputs the predicted signal transmission delay value and generates dynamic compensation coefficients through the GRU submodule to perform real-time compensation on the original signal. The compensation formula is as follows: Where Kcomp is the dynamic compensation coefficient, Kcomp= , This is the baseline delay value.

[0048] In this embodiment, the stable delay under normal operating conditions, the historical average delay, or the design reference value are used to standardize the error and achieve dynamic delay compensation under drastic temperature changes and electromagnetic interference.

[0049] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for predicting faults in high-voltage connectors, characterized in that, Includes the following steps: Based on the electrical signal and mechanical and bolt condition data of high-voltage connectors, sensitive features of bolt fastening failure are extracted from three dimensions: torque change, gap state, and vibration response, combined with the degree of corrosion. Based on the electrical and thermal signals at various locations of the high-voltage connector, and combined with a supplementary temperature change rate correction term, the optimized temperature resistance coupling coefficient is calculated. An optimized signal transmission delay prediction model is constructed, employing a bivariate long short-term memory network. Based on the dynamic correlation characteristics between temperature, current, and signal transmission delay, linear coupling characteristics are output. Nonlinear coupling characteristics between temperature fluctuations, current fluctuations, and contact resistance changes are extracted using a kernel partial least squares algorithm. A thermoelectric coupling feature set is constructed by optimizing the temperature resistance coupling coefficient, linear coupling characteristics, and nonlinear coupling characteristics. Based on temperature cycling data and material properties, we analyze the sensitive characteristics of thermal fatigue life. The LightGBM model is used to score the importance of features and to fuse features sensitive to bolt fastening failure, thermoelectric coupling feature set, and thermal fatigue life. An improved Transformer prediction and lifetime assessment model is established. Fault prediction and lifetime prediction are performed through feature fusion. Based on the optimized signal transmission delay prediction model, dynamic compensation coefficients are generated to compensate the original signal in real time.

2. The high-voltage connector fault prediction method according to claim 1, characterized in that, Based on three dimensions—torque variation, gap condition, and vibration response—and combined with the degree of corrosion, sensitive characteristics of bolt fastening failure are extracted, including the following steps: From the three dimensions of torque change, gap state and vibration response, combined with the degree of corrosion, torque-related features, gap-resistance correlation features and vibration-corrosion features are extracted as sensitive features of bolt fastening failure. Based on the torque change data, the torque attenuation rate is calculated as a torque-related characteristic: εt=(Bt0-Btt) / Bt0; Where εt is the torque decay rate, Bt0 is the initial tightening torque of the bolt, and Btt is the real-time monitored torque at time t; when εt≥8%, it is determined to be a torque decay warning state, otherwise it is not determined to be a torque decay warning state. Based on the gap between the bolt and the copper busbar, and considering the resistance of the copper busbar circuit, calculate the relationship between the gap and the resistance: in, These are the characteristic coefficients relating gap and resistance. The resistance of the copper busbar circuit under healthy conditions. The resistance of the copper busbar circuit being tested is currently being measured. This refers to the gap between the bolt and the copper busbar under healthy conditions. This refers to the current gap between the bolt and the copper busbar; when If the value is greater than or equal to the threshold, it is determined that the copper busbar is loose and an alarm is triggered; otherwise, no alarm is triggered. The vibration signal of the bolt area is subjected to FFT transformation to extract the resonant frequency offset Δf. The corrosion depth hco is obtained by combining the corrosion sensor data. Based on the synergistic acceleration effect of vibration and corrosion, two types of coupling features are constructed, including product coupling features and normalized coupling features. The product-type coupling feature F1 = Δf * hco, where Δf is the currently detected bolt resonance frequency offset and hco is the currently detected bolt corrosion depth threshold. The normalized coupling feature F2 = (Δf / Δfmax)*0.6 + (hco / hcomax)*0.4, where Δfmax is the critical resonant frequency offset of bolt failure and hcomax is the safe corrosion depth threshold of bolt. Vibration-corrosion coupling features are constructed using product-type coupling features and normalized coupling features.

3. The high-voltage connector fault prediction method according to claim 1, characterized in that, Based on the electrical and thermal signals at various locations of the high-voltage connector, and combined with a supplementary temperature change rate correction term, the optimized temperature resistance coupling coefficient is calculated, including the following steps: Calculate the temperature-resistivity coupling coefficient based on Joule's law and the heat transfer equation. The formula is: in, For the contact resistance of the high-voltage connector, This represents the average real-time temperature at various locations inside the high-voltage connector. The temperature coefficient of resistance of copper. The contact resistance of a high-voltage connector under normal operating conditions at a standard ambient temperature of T=25℃. This is the average real-time temperature of various locations inside the high-voltage connector during normal operation at a standard ambient temperature of T=25℃. Incorporating the temperature change rate correction term, the formula for optimizing the temperature resistance coupling coefficient is established as follows: in, The optimized temperature-resistance coupling coefficient is given by β, where β is the temperature rate correction coefficient. The rate of temperature change.

4. The high-voltage connector fault prediction method according to claim 1, characterized in that, Constructing an optimized signal transmission delay prediction model includes the following steps: Based on the principle of electromagnetic induction, calculate the delay coupling coefficient of electromagnetic interference signals: in, Here, Δτ is the electromagnetic interference signal delay coupling coefficient, and Δτ is the change in signal transmission delay. This represents the change in electromagnetic interference intensity. Based on the combined effects of temperature and electromagnetic interference, a coupling correction factor is added to optimize the signal transmission delay prediction model: in, For the optimized signal transmission delay, For signal transmission delay under healthy conditions, The optimized temperature resistance coupling coefficient, This represents the average real-time temperature at various locations inside the high-voltage connector. This is the average real-time temperature of various locations inside the high-voltage connector during normal operation at a standard ambient temperature of T=25℃.

5. The high-voltage connector fault prediction method according to claim 1, characterized in that, A bivariate long short-term memory network is used, based on the dynamic correlation characteristics of temperature, current, and signal transmission delay, including the following steps: The current of the high-voltage connector and the real-time temperature of various internal locations are selected as input variables. The signal transmission delay of the optimized signal transmission delay prediction model is used as the target variable. Time series sample pairs are constructed: the data is divided according to time windows and the input data is standardized. A Bi-LSTM bivariate long short-term memory network was constructed, and the input layer dimension was set according to the standardized temperature and current time series. The input layer is connected to a bidirectional LSTM layer, which has two hidden layers: a forward LSTM and a backward LSTM to capture the forward and backward dependencies of the time series data, respectively. The bidirectional LSTM layer is connected to the fully connected layer, which maps the spliced ​​features output by the bidirectional LSTM to a delayed prediction sequence with the same time step as the input. The fully connected layer connects to the output layer and outputs the predicted delay value for each time step; The output of the last hidden state of the trained Bi-LSTM is used as a dynamic correlation feature vector of temperature-current-delay.

6. The high-voltage connector fault prediction method according to claim 5, characterized in that, Outputting linear coupling characteristics includes the following steps: Based on the dynamic correlation feature vector of temperature-current-delay analyzed by Bi-LSTM bivariate long short-term memory network, the coupling feature deviation value ΔF is calculated: ; Where Fcurrent is the dynamic correlation feature vector of temperature-current-delay for the current sample, Fhealthy is the dynamic correlation feature vector of temperature-current-delay for the high-voltage connector under healthy conditions, and the output coupling feature deviation value ΔF is used as the linear coupling feature.

7. The high-voltage connector fault prediction method according to claim 1, characterized in that, The nonlinear coupling characteristics of temperature fluctuations, current fluctuations, and contact resistance changes are extracted using a kernel partial least squares algorithm, including the following steps: The nonlinear coupling characteristics of temperature fluctuation, current fluctuation and contact resistance change are extracted using the KPLS kernel partial least squares algorithm. By calculating the average real-time temperature at various locations inside the high-voltage connector The average real-time temperature of various locations inside the high-voltage connector under normal operating conditions at a standard ambient temperature of T=25℃. The difference between the real-time current of the high-voltage connector and the average current under healthy conditions is used as the change in temperature fluctuation. The temperature fluctuation coefficient and the current fluctuation coefficient are set as input variables to form a 2D input matrix. The output variable is set to the contact resistance change ΔRc=Rc-Rc0, where Rc0 is the contact resistance in the healthy state and Rc is the contact resistance detected in real time by the high-voltage connector. Collect historical data, construct a sample set of input variables and corresponding output variables, and train the KPLS kernel partial least squares algorithm; Extract the top principal components that are most strongly correlated with the output variables, and use these extracted principal components to form a nonlinear coupling feature vector, which serves as the nonlinear coupling feature.

8. The high-voltage connector fault prediction method according to claim 1, characterized in that, Based on temperature cycling data and material properties, the sensitivity characteristics of thermal fatigue life are analyzed, including the following steps: Analyze temperature cycling characteristics and fatigue damage characteristics as sensitive features of thermal fatigue life; Calculate the thermal stress of the copper busbar and insulating components based on the material's coefficient of thermal expansion and elastic modulus. As a characteristic of thermal stress, the formula is: Where E is the elastic modulus of the material. Let μ be the coefficient of thermal expansion of the material, and μ be Poisson's ratio. Temperature cycling amplitude; thermal stress of copper busbars and insulating components. When the insulation aging risk warning state is exceeded by 30% of the material yield strength, it is determined to be an insulation aging risk warning state; otherwise, it is not determined to be an insulation aging risk warning state. Thermal stress characteristics, as well as the temperature cycling amplitude, cycling frequency, and dwell time at the high temperature threshold extracted from the high-voltage connector, are used as temperature cycling characteristics. Using Miner's linear cumulative damage theory, combined with a correction coefficient, the cumulative thermal fatigue damage value of insulation and metal components is quantified and used as a fatigue damage characteristic. When the cumulative thermal fatigue damage value of insulation and metal components exceeds a threshold, it is judged as a high-risk warning for thermal fatigue failure; otherwise, it is not judged as a high-risk warning for thermal fatigue failure.

9. The high-voltage connector fault prediction method according to claim 1, characterized in that, Using the LightGBM model, feature importance scoring is performed, and feature fusion is conducted on bolt fastening failure sensitive features, thermoelectric coupling feature sets, and thermal fatigue life sensitive features, including the following steps: Construct feature set matrices and fault label vectors for bolt fastening failure sensitive features, thermo-electric coupling feature sets, and thermal fatigue life sensitive features; calculate the MIC value of each type of feature and fault label vector using the maximum mutual information coefficient, and retain features with MIC values ​​greater than the threshold. The selected bolt fastening failure sensitive features, thermo-electric coupling feature set and thermal fatigue life sensitive features are input into the LightGBM classification model, and the feature contribution is automatically evaluated during the model training process. The LightGBM classification model parameters are set as follows: learning rate = 0.01, number of decision trees = 100, maximum tree depth = 6, number of leaf nodes = 32. After training, the gain value of each feature is extracted as the gain importance, and the core features with gain importance greater than or equal to the threshold are retained. The retained bolt fastening failure sensitive features, thermo-electric coupling feature set and thermal fatigue life sensitive features are divided into different domains according to each feature. Attention weights are calculated separately for feature subsets of each domain. Feature scores are generated by tanh activation function and converted into normalized weights by Softmax function. The features within a domain are weighted and summed according to their weights to generate domain-level fusion features. The fusion features from all domains are then concatenated to form the final fusion feature vector.

10. A high-voltage connector fault prediction method according to claim 4, characterized in that, An improved Transformer prediction and lifetime assessment model is established. Fault and lifetime predictions are performed through feature fusion. Based on an optimized signal transmission delay prediction model, dynamic compensation coefficients are generated to compensate the original signal in real time. The process includes the following steps: A Transformer model is established, and the input layer divides the fused feature sequence into time windows to construct temporal input samples; physical temporal enhancement position coding is introduced, and position codes are dynamically generated by combining sampling interval and working condition cycle. The multi-branch output layer is configured with parallel output branches, including: fault prediction branch, thermal fatigue life prediction branch, and signal delay compensation branch; The fault prediction branch outputs the probability of typical bolt-related faults and the bolt life prediction. Thermal fatigue life prediction of the remaining life of the branch output high voltage connector; Filter historical data for a preset time period before the high-voltage connector fails and becomes unusable, and for a preset time period before typical bolt-related failures occur, and generate a fused feature vector; The specific time period before the failure to use the connector is marked as the remaining life of the high-voltage connector, and typical bolt-related failures are marked. The specific time period before the occurrence of typical bolt-related failures is marked as the bolt life. The Transformer model is trained using labeled data to obtain an improved Transformer prediction and life assessment model. Based on the fused feature vector of the current high-voltage connection monitoring data, the prediction results are output through the fault prediction branch and the thermal fatigue life prediction branch. The signal delay compensation branch of the improved Transformer prediction and lifetime assessment model is based on the optimized delay prediction model. It outputs the predicted signal transmission delay value and generates dynamic compensation coefficients through the GRU submodule to perform real-time compensation on the original signal. The compensation formula is as follows: Where Kcomp is the dynamic compensation coefficient, Kcomp= , This is the baseline delay value.

Images