Life prediction method of charging connector by combining accelerated life test with field data

Abstract

This invention relates to the field of reliability assessment technology for electric vehicle charging equipment, specifically disclosing a method for predicting the lifespan of charging connectors by integrating accelerated life testing and field data. The method includes: constructing a laboratory lifespan dataset and a field dataset; preprocessing the two types of data and converting the laboratory degradation time into equivalent field usage time using an acceleration factor model; pre-training a priori lifespan inference model using the preprocessed laboratory data, synthetic data, and the mapped equivalent field time; inputting historical data of the target device into the priori lifespan inference model and directly outputting the lifespan prediction result; and using the lifespan prediction result for operation and maintenance decisions. This invention uses laboratory accelerated life testing as the main data source, combines a small amount of field data, and can infer lifespan status without field training. It can model the lifespan degradation of key components of charging guns under complex operating conditions, predict remaining lifespan, and assess failure risk.

Description

Technical Field

[0001] This application relates to the field of reliability assessment technology for electric vehicle charging equipment, and specifically discloses a method for predicting the lifespan of a charging connector that integrates accelerated life testing and field data. Background Technology

[0002] In actual service, the charging gun is subjected to the combined effects of mechanical stress from insertion and removal, electrothermal stress caused by high current flow, complex and variable natural environment, and multi-physical field coupling stress such as cable bending and insertion / removal holding force. This can easily lead to degradation phenomena such as increased contact resistance, abnormal temperature rise, poor contact, contact point erosion, and mechanical structure wear. In severe cases, it can even cause charging interruption or safety accidents, significantly affecting the reliability and service life of the charging system.

[0003] Currently, the industry mainly relies on the following methods for life assessment and health management of charging guns: the "fixed life / experience curve" assessment method based on laboratory accelerated life tests, the maintenance strategy based on threshold rules, and the machine learning / deep learning model that relies on a large number of field failure samples for training.

[0004] However, laboratory conditions often fail to fully replicate the complex and variable actual working conditions in the field, making it difficult to accurately calibrate the acceleration factor and resulting in a significant deviation between laboratory results and actual field lifespan. Furthermore, setting a threshold for a single parameter cannot fully reflect the degradation process under multiple stress coupling effects, which can easily lead to missed or false alarms. In addition, real failure samples in the field are extremely scarce, and the operating conditions, environmental conditions, and maintenance history of charging guns vary greatly, resulting in severe data distribution drift. Model training faces serious problems of sample imbalance and insufficient generalization ability, making it difficult to promote and apply in actual engineering.

[0005] It is evident that current approaches to predicting the lifespan of charging guns face challenges such as insufficient data to support the training of highly generalizable models, difficulties in mapping accelerated testing conditions in the laboratory to actual field conditions, and the inability of a single model structure to balance physical interpretability with generalization capability under complex conditions. In view of this, the present invention provides a method for predicting the lifespan of charging connectors that integrates accelerated life testing and field data. Summary of the Invention

[0006] The purpose of this invention is to provide a prediction method that uses laboratory accelerated life testing as the main data source, combines a small amount of field data, and can infer life status without field training, so as to solve the related technical problems mentioned in the background art.

[0007] To achieve the above objectives, the basic solution of the present invention provides a method for predicting the lifespan of a charging connector that integrates accelerated life testing and field data, comprising the following steps:

[0008] Step S100: Construct laboratory lifespan dataset and field dataset;

[0009] Step S200: Preprocess the two types of data and convert the laboratory degradation time into equivalent field usage time using the acceleration factor model;

[0010] Step S300: Pre-train the prior lifetime inference model using pre-processed laboratory data, synthetic data, and mapped field equivalent time;

[0011] Step S400: Input the historical data of the target device into the pre-trained prior life inference model with fixed parameters, and directly output the life prediction result.

[0012] Step S500: Use the life prediction results for operation and maintenance decisions.

[0013] Furthermore, in step S100, multidimensional stress accelerated life tests are conducted on charging gun samples of the same model or series, and the temporal characteristics and failure time of each test sample under different stress conditions are recorded to form a laboratory life dataset.

[0014] For the charging guns operating on-site, data on operational behavior, electrical and thermal characteristics, environmental and operating conditions, as well as fault and maintenance are collected according to the preset collection frequency and accuracy requirements to form an on-site operation dataset.

[0015] Furthermore, the preprocessing in step S200 includes performing missing value completion, outlier handling, and normalization on the laboratory lifespan dataset and the field operation dataset, respectively.

[0016] Time period encoding is introduced into the field operation dataset.

[0017] Furthermore, in step S200, the expression for the on-site equivalent time is:

[0018]

[0019] in, As an acceleration factor, The degradation time under accelerated laboratory testing;

[0020]

[0021] or

[0022]

[0023] In the formula, This is the laboratory test temperature. The average operating temperature on site. For equivalent activation energy, Boltzmann's constant, This is the converted on-site equivalent operating time. For laboratory current, The average current at the site, This is the current acceleration index.

[0024] Furthermore, the prior lifetime inference model in step S300 includes:

[0025] The causal constraint feature modeling submodule is used to filter out the core feature set that has a stable causal relationship with lifespan degradation;

[0026] The degradation characterization submodule is used to encode the original time series data at multiple scales and generate a vector representation that characterizes the degradation trajectory.

[0027] The prior lifetime inference submodule is used to obtain lifetime mapping priors by pre-training with laboratory and synthetic data.

[0028] Furthermore, the causal constraint feature modeling submodule includes constructing a candidate feature set, evaluating the causal relationship between candidate features and lifetime indicators, and determining key degradation factors according to preset rules;

[0029] The degradation characterization submodule includes constructing local degradation segments from time-series data using a sliding window approach, encoding the data within the window, and generating a degradation characterization vector of fixed dimensions.

[0030] The training sample pairs of the prior lifetime inference submodule include a condition part and a response part. The condition part consists of multi-source features, fixed parameters, and degradation characteristics for N consecutive days or N usage cycles. The response part consists of remaining lifetime, failure risk probability, or degradation degree indicators.

[0031] Furthermore, the prior lifetime inference model described in step S300 adopts a Transformer-based architecture, including:

[0032] Conditional sample encoding unit is used to map input features into high-dimensional embedding vectors;

[0033] The functional relationship modeling unit, composed of multiple stacked Transformer encoders, is used to capture the complex nonlinear relationship between the input feature sequence and the lifetime label;

[0034] The inference calculation unit is used to map the hidden state vector output by the functional relationship modeling unit to the final lifetime prediction result and output the confidence interval.

[0035] Furthermore, the life prediction result in step S400 is at least one of the following: remaining life estimate, failure risk probability sequence within the next T time period, degradation level, or maintenance recommendation level.

[0036] Based on the same inventive concept, the present invention also provides a charging connector life prediction system that integrates accelerated life testing and field data, comprising:

[0037] At least one processor;

[0038] At least one memory for storing at least one program;

[0039] When the at least one program is executed by the at least one processor, the at least one processor implements the above-described method for predicting the lifespan of a charging connector by integrating accelerated life testing and field data.

[0040] Furthermore, the system is also equipped with a data acquisition module for performing the function described in step S100 of any one of claims 1-7, including a laboratory accelerated life test data acquisition unit and a field operation data acquisition unit;

[0041] The memory stores the following modules that can be executed by the at least one program:

[0042] The data preprocessing and mapping module is used to perform the functions described in step S200, including a missing value completion unit, an outlier handling unit, a normalization unit, a periodic encoding unit, and an acceleration factor mapping unit.

[0043] The model building module is used to perform the functions described in step S300, including a causal constraint feature modeling unit, a degradation characterization unit, and a prior lifetime inference model training unit.

[0044] The lifetime inference module is used for the function described in step S400. It includes a conditional sample construction unit and a model inference unit. The model inference unit has a pre-stored prior lifetime inference model with fixed parameters trained by the model construction module.

[0045] The output and application module is used to perform the function described in step S500, output the life prediction result and generate corresponding operation and maintenance decision instructions.

[0046] Based on the same inventive concept, the present invention also provides an electronic device that stores computer instructions, which, when executed, run the above-mentioned charging connector life prediction system that integrates accelerated life test and field data, and implement the above-mentioned charging connector life prediction method that integrates accelerated life test and field data.

[0047] The principle and effect of this solution are as follows:

[0048] This solution constructs a priori inference model using laboratory data as a knowledge source. When predicting the lifespan or analyzing the failure probability of industrial products, it leverages the priori inference paradigm to mine the knowledge value of laboratory data, overcoming data bottlenecks and completely eliminating reliance on massive amounts of high-quality field failure samples. This precisely addresses the core industry pain point of scarce field failure samples in industrial scenarios. The solution eliminates spurious correlations between data through causal constraints, allowing the model to focus on the driving factors of physical essence. Simultaneously, it combines an acceleration factor model to achieve accurate physical mapping, enhancing generalization ability and significantly improving the model's generalization and robustness under unknown operating conditions and different batches of products. In practical applications, no additional training or fine-tuning of the model is required. Analysis results can be output through a single forward inference, resulting in low computational overhead. It can be lightweightly deployed on edge computing nodes or in the cloud, reducing deployment costs and significantly reducing on-site deployment and subsequent maintenance investment. At the same time, the solution can output the remaining useful life (RUL) or failure probability curve with confidence intervals, improving decision-making credibility and providing maintenance personnel with richer and more reliable decision-making basis. This helps industrial maintenance upgrade from traditional "planned maintenance" to intelligent "predictive maintenance," effectively reducing the risk of unplanned downtime in industrial production and further optimizing the product's total life cycle cost. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 A flowchart of a charging connector lifetime prediction method integrating accelerated life testing and field data, as proposed in an embodiment of this application, is shown.

[0051] Figure 2 The diagram illustrates the architecture of a charging connector life prediction method that integrates accelerated life testing and field data, as proposed in an embodiment of this application. Detailed Implementation

[0052] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description, in conjunction with the accompanying drawings and preferred embodiments, provides a detailed explanation of the specific implementation methods, structures, features, and effects of the present invention. A method for predicting the lifespan of a charging connector that integrates accelerated life testing and field data is implemented, for example... Figure 1 and Figure 2 As shown, it includes the following steps:

[0053] Step S100: Construct laboratory life test dataset and field data acquisition specifications.

[0054] A multi-source, high-quality foundational database will be constructed to provide data support for subsequent model training and lifetime inference. The data sources mainly include two parts: controllable laboratory accelerated life test data and real-world field operation data, specifically:

[0055] Step S110, Construction of accelerated life test data in the laboratory

[0056] For charging gun samples of the same model or series, conduct multidimensional stress-accelerated life tests under laboratory conditions. The test design should cover the main stress types that the charging gun may encounter in actual use and accelerate its degradation process. Record the temporal characteristics and failure time of each test sample under different stress conditions to form a laboratory life dataset.

[0057] Laboratory accelerated life tests should include, but are not limited to, the following types, and corresponding data should be collected simultaneously:

[0058] Insertion and extraction life test: Under rated current or specified current, perform continuous insertion and extraction operations, and record the changes in parameters such as the number of insertions and extractions, nozzle temperature, contact resistance, and insertion and extraction force throughout the process.

[0059] Current-carrying aging test: Apply a current higher than the rated value to simulate the electrothermal aging caused by long-term high current flow, and record the current, voltage, nozzle temperature, temperature rise rate, contact resistance, etc.

[0060] Environmental stress testing: including but not limited to:

[0061] Damp heat test: Insertion, removal or current flow are performed in a high humidity environment to evaluate the impact of the humid environment on contact performance.

[0062] Immersion test: Immerse the charging gun plug in salt water or clean water for a certain period of time, then plug and unplug it while it is powered on, and record whether breakdown, leakage and changes in contact resistance occur.

[0063] Mechanical performance test:

[0064] Bending test: Repeatedly bend the connection between the cable and the gun tail, and record the number of bends, the continuity of the internal wires, and the change in contact resistance.

[0065] Holding force test: The attenuation of the insertion and extraction force between the plug and the socket is tested by means of hanging weights, etc.

[0066] Data collection must follow the standards shown in the table below.

[0067]

[0068] To accurately label the data with lifetime information, clear failure criteria must be defined. A sample is considered to have failed and the failure time is recorded when any of the following conditions are met:

[0069] Contact resistance failure: When the contact resistance R ≥ 20mΩ, or increases by ≥ 200% relative to the initial value, it is judged as a resistance failure;

[0070] Temperature rise failure: When the temperature rise ΔT ≥ 40℃, or the temperature rise rate ≥ 5℃ / min, it is judged as thermal anomaly failure;

[0071] Mechanical performance failure: insertion and extraction force decreases by ≥ 30%, or holding force is lower than 80N; this is considered a mechanical structural failure.

[0072] In this embodiment, the comprehensive health index HI is defined as follows:

[0073]

[0074] in, , , These are the normal values ​​for resistance, temperature, and insertion / extraction force, respectively. , , These are the weighting coefficients. + + = 1, preferably 0.4, 0.3, or 0.3.

[0075] It is considered invalid when HI ≤ 0.2.

[0076] Step S120, On-site Operation Data Acquisition Specifications

[0077] To ensure structural consistency with laboratory data and the feasibility of subsequent collaborative analysis, the collection of field operational data must follow a unified standard.

[0078] The data acquisition carrier is configured as follows:

[0079] The on-site charging gun needs to be equipped with the following sensors and modules:

[0080] Embedded temperature sensor: Installed at the muzzle or terminal, used to measure the temperature of the core heating point.

[0081] Current and voltage acquisition module: integrated into the charging module or charging gun end.

[0082] Insertion / removal counting module: Microswitches or Hall sensors can be used to record insertion / removal events.

[0083] Abnormal alarm recording module: Records alarm information such as overcurrent, overtemperature, and leakage during the charging process.

[0084] Online contact resistance detection module: Optional, which can be used for detection during charging intervals or maintenance cycles.

[0085] Ambient temperature and humidity sensor: installed near the charging station or gun mount.

[0086] The data acquisition frequency and accuracy requirements are as follows:

[0087] Sampling frequency:

[0088] Operating status data (temperature, current, voltage): sampled every 15 to 60 seconds.

[0089] Insertion / removal count: Event-driven, recorded each time an insertion / removal action occurs.

[0090] Alarm data: Recorded in real time and uploaded immediately upon occurrence.

[0091] Contact resistance: Tested according to maintenance cycles (such as weekly or monthly) or when specific conditions are triggered.

[0092] Accuracy requirements:

[0093] Temperature accuracy: not less than ±1℃.

[0094] Current accuracy: not less than ±1%FS.

[0095] Contact resistance measurement error: not exceeding ±5%.

[0096] Data collected on-site should include at least the categories corresponding to laboratory data, as follows:

[0097] Operational behavior data includes: cumulative number of plug-in / plug-out cycles, cumulative power-on duration, cumulative power consumption (Ah / kWh), charging rate (C-rate), duration of a single charge, and daily / weekly / monthly usage intensity characteristics (such as average number of plug-in / plug-out cycles per day), reflecting user habits and charging load.

[0098] Electrical and thermal characteristics: muzzle temperature (real-time / daily peak / temperature rise gradient), temperature rise rate dT / dt, effective value and fluctuation of current / voltage, and trend of contact resistance (if online monitoring is available), reflecting the electrothermal stress state during the charging process.

[0099] Environmental and operating condition data: ambient temperature and humidity, indoor / outdoor markings, and whether it is under high load continuous use (such as charging times / duration exceeding preset thresholds), reflecting the external environment and special operating conditions.

[0100] Fault and maintenance data: abnormal alarm records (type, time), maintenance and replacement records (replaced parts, reasons), and fault labels recorded on site (such as poor contact, overheating, and burning) provide direct evidence of degradation and failure.

[0101] Step S200, Data Preprocessing and Co-mapping

[0102] Since laboratory data and field data have inherent differences in sampling frequency, operating condition stability, and data integrity, this step aims to preprocess the two types of data and establish a mapping relationship from high-stress laboratory conditions to normal field operating conditions, laying the foundation for subsequent unified modeling.

[0103] First, data type classification and general preprocessing are performed, dividing the input data into three categories to accommodate different preprocessing rules:

[0104] Category A: Laboratory accelerated test data: continuously collected, with few missing data points, under a single operating condition, and without periodicity.

[0105] Category B: On-site runtime sequence data: contains intermittent sampling, missing data, and noise, and exhibits obvious daily / weekly periodicity.

[0106] Category C: Fixed parameter data: such as equipment model, sensor installation location, design life, etc., which do not change over time.

[0107] Then, missing values ​​were filled in for the accelerated laboratory test data and the field runtime sequence data, according to the following rules:

[0108] Laboratory accelerated test data: Laboratory test data is usually continuously acquired with a low missing rate. For window lengths with continuous missing time ≤ 5%, linear interpolation is used; for test sample segments with a missing rate > 5%, the test sample segments are discarded.

[0109] The expression for the interpolation formula is as follows:

[0110]

[0111] In the formula, For the missing length, This represents the number of interpolation steps.

[0112] On-site runtime sequence data: On-site data may have intermittent sampling or communication anomalies: For short-term missing data of ≤4 time slices, linear interpolation or forward filling is used; for long-term missing data, it is filled based on the mean of similar working condition samples, or KNN interpolation (K=3~5) is used.

[0113] Outlier handling:

[0114] For accelerated laboratory test data: the Z-score method is used, and sample points with an absolute value greater than 3 are considered outliers. Smoothing or interpolation replacement is then performed based on the preceding and following normal values. The expression is as follows:

[0115]

[0116] in, This is the mean of the feature across all samples or a local window. Let | be the standard deviation. If the value is greater than 3, it is considered an outlier and smoothed or interpolated based on the normal values ​​before and after it.

[0117] For on-site runtime sequence data: a combination of physical constraints and statistical methods is used. For example, temperature values ​​exceeding the physical upper limit (e.g., 120℃) or significantly higher than historical averages (e.g., 3 times the standard deviation) are considered outliers, removed, and treated as missing values, as shown in the following expression:

[0118]

[0119] in, and The mean and standard deviation are for a local window (such as the past 24 hours).

[0120] Normalization process:

[0121] For laboratory accelerated test data, Z-score standardization was used to make the data mean 0 and the standard deviation 1, in order to eliminate the influence of dimensions and highlight the degradation trajectory.

[0122]

[0123] in, The characteristic mean, Standard deviation;

[0124] For the on-site running sequence data, Min-Max scaling is used to map the data to the [0, 1] interval to maintain its physical boundaries (e.g., temperature cannot be negative, and current has a rated maximum value).

[0125]

[0126] Periodically encode the field data

[0127] Field operation data exhibits characteristics such as daily periodic fluctuations, changes in usage intensity, environmental and seasonal influences, and the risk of data discontinuity. Therefore, time period coding is introduced for the field data, where t represents the intraday time slice sequence number and T represents the total number of time slices in a day. The period coding formula is as follows:

[0128]

[0129]

[0130] Taking the periodic cycle into account, the periodic coding formula is revised as follows:

[0131]

[0132]

[0133] Where d is the week number (1–7).

[0134] Laboratory accelerated life test data are characterized by single operating conditions, no obvious daily / weekly cycles, high data continuity, and controllable temperature and stress. Therefore, laboratory data does not need to be seasonally or periodically encoded. The focus should be on smoothing the degradation trajectory and removing outliers.

[0135] Laboratory and field data collaboration: To map the accelerated degradation trajectory from the laboratory to the field operational degradation trajectory, the following collaboration strategy is adopted: Maintaining consistency in key physical variables such as temperature, current, and contact resistance; and normalizing the time scale of sampling frequency differences while ensuring unified units and dimensions, as shown in the following expression:

[0136]

[0137] This allows high-frequency laboratory data to be mapped to field statistical scales.

[0138] To map laboratory high-stress accelerated life data to the actual field operating environment, an acceleration factor model is established to convert the degradation time under laboratory stress conditions into the equivalent field service time. The expression is as follows:

[0139]

[0140] The on-site equivalent time is calculated as follows:

[0141]

[0142] in, This is the laboratory test temperature. The average operating temperature on site. For equivalent activation energy, Boltzmann's constant, To accelerate the degradation time in laboratory tests, This is the converted on-site equivalent operating time.

[0143] Equivalent on-site running time after conversion It serves as a unified degradation timescale for constructing lifetime labels or as input features for degradation modeling.

[0144] When considering the influence of current stress, the extended Eyring model can be used, and its expression is as follows:

[0145]

[0146] in, For laboratory current, The average current at the site, This is the current acceleration index.

[0147] Equivalent activation energy Current acceleration index Data can be obtained through regression fitting from accelerated laboratory testing data of the same type of device, for example, for two sets of different temperatures. , Lifetime data below , From the above formula, we can obtain:

[0148]

[0149] The equivalent activation energy can be solved using linear regression. Similarly, by keeping the temperature constant and changing the current, the current acceleration exponent can be calculated. If experimental conditions are unavailable, industry-standard empirical parameters can be used as a reference (e.g., equivalent activation energy for contact oxidation). Typically taken as 0.6-1.0 eV; current acceleration index 1-3 is usually used.

[0150] This mapping relationship is mainly used for:

[0151] Unified degradation timescale: Convert the failure time of laboratory samples into "field equivalent usage time" to provide a unified lifetime label for model training.

[0152] Constructing prior samples with similar operating conditions: For a specific operating condition (T) on site field , I field (This can be achieved by using an acceleration factor to reverse-select or by weighting samples with similar stress in the laboratory to build more targeted prior models.)

[0153] Step S300: Construct a priori lifetime inference model

[0154] The aim is to construct a combined lifetime modeling framework of "causal constraints, prior inference, and degradation representation" using preprocessed laboratory data, synthetic data, and mapped field equivalent time. This model is pre-trained on massive amounts of laboratory data to form a powerful lifetime inference prior, eliminating the need for retraining when applied to field equipment. This includes:

[0155] Causal Constraint Feature Modeling Submodule: To avoid the model learning spurious correlations in the data and improve its generalization ability across operating conditions, this module introduces a causal constraint mechanism to select a core feature set that has a stable causal relationship with lifespan degradation. Specifically, it includes the following steps:

[0156] First, a candidate feature set is constructed, including but not limited to: average contact resistance and growth rate, contact resistance fluctuation amplitude, peak nozzle temperature and temperature rise margin, temperature rise rate dT / dt, cumulative insertion and removal times N, power-on time per unit time, cumulative current flow Ah / kWh, average insertion and removal force and attenuation rate, abnormal alarm frequency, ambient temperature and humidity, and high-load continuous use indicator.

[0157] Then, causal strength assessment evaluates the causal relationship between candidate features and lifetime indicators. Methods may include one or a combination of the following: causal discovery algorithms based on conditional independence tests, causal graph inference based on structural constraints, Granger causality tests, feature importance ranking based on random forests or gradient boosting trees, SHAP value contribution analysis, and causal graph structural constraints constructed based on expert experience.

[0158] Finally, key factors are determined. A candidate factor is determined to be a key degradation factor when it meets one or a combination of the following conditions: the p-value of the causal significance test is less than 0.05; it is in the top 30% of the feature importance ranking; its contribution to the life prediction error is not less than 5%; and it maintains a stable contribution in different operating condition subsets.

[0159] The selected set of key factors is used to constrain model training, making the model focus on driving variables with stable physical or engineering significance, thereby improving the generalization ability across batches and operating conditions.

[0160] Degradation Characterization Submodule: To effectively capture the long-term degradation trend of the charging gun, this module performs multi-scale encoding on the original time-series data to generate a vector representation that can characterize the degradation trajectory, specifically including:

[0161] A sliding window is used to construct local degradation segments for time-series data such as temperature rise gradient, contact resistance change trend, and alarm sequence: the window length L ranges from 16 to 128, preferably L = 32 or 64, and the sliding step size is 1 to 4.

[0162] Degenerate encoding encodes the data within a window, generating a fixed-dimensional degenerate representation vector. One or a combination of the following methods can be used:

[0163] Statistical feature encoding: Calculates the mean, variance, slope (characterizing the rate of degradation), maximum value, minimum value, etc. of the data within the window. For example, the slope can be obtained by fitting a straight line using the least squares method.

[0164] One-dimensional convolutional coding (Conv1D): Automatically extracts local temporal patterns using convolutional kernels. For an input sequence X ∈ R^{L×d} (where L is the window length and d is the feature dimension), the convolution operation is represented as:

[0165]

[0166] in, For convolution kernel weights, For bias, The activation function is ReLU, the kernel size K ranges from 3 to 7, preferably K=5; the number of channels ranges from 32 to 128.

[0167] Finally, the degenerate representation vector, the latest values ​​of key causal features, and fixed parameters (such as model number) are concatenated to form a unified model input vector.

[0168] Prior lifetime inference submodule: Using processed laboratory and synthetic data, a mapping function is pre-trained to infer lifetime based on historical conditions.

[0169] Laboratory data can cover hundreds to thousands of test samples (devices / batch), while field data can cover dozens to hundreds of operational segments of equipment; and it allows for cross-model expansion to a "collection of devices with similar structures".

[0170] Each training sample pair includes:

[0171] Conditional Part C: Multi-source table / time series features for N consecutive days (or N usage cycles) + fixed parameters + degradation characterization;

[0172] Response section Y: Lifetime label / degradation index, such as: Remaining Lifetime RUL (expressed as number of cycles / time), or failure risk probability (probability of failure within the next T), or degradation level index (such as contact resistance margin, temperature rise margin).

[0173] Where N is an adjustable parameter, preferably a usage cycle of 7–60 days or 50–500 times.

[0174] The training / validation set is divided into two parts (e.g., 70–90% / 10–30%) for model training, either by sample or batch. One of Adam, AdamW, or SGD is used as the optimizer. Taking Adam as an example, the parameter update rule for the Adam optimizer is as follows:

[0175]

[0176]

[0177]

[0178]

[0179]

[0180] In the formula, For the first moment estimate at step t, The decay rate of the first moment (hyperparameter) is typically taken as 0.9. This is the momentum value from the previous step. Let t be the gradient at step t, which is the derivative of the loss function with respect to the current model parameters; For the second moment estimate at step t, This represents the decay rate of the second moment (a hyperparameter). It is typically taken as 0.999. For the second moment estimation at step t-1, The square of the gradient at step t; This is the first-order moment estimate after bias correction. for The power of t, where t is the current iteration step; This is the second-order moment estimate after bias correction; These are the model parameters updated in step t. These are the model parameters updated at step t-1. For learning rate, It is a constant used to prevent the denominator from being zero and to ensure numerical stability.

[0181] Learning rate: 1e-3 to 1e-5, using piecewise decay or cosine annealing strategies. For the Adam optimizer, an initial learning rate of 3e-4 is preferred.

[0182] Loss function: can be selected according to the output type.

[0183] For RUL regression, a weighted combination of MAE (mean absolute error) and Huber loss can be used. Huber loss is defined as:

[0184]

[0185] in, This represents Huber's loss value. For the true value, For predicted values, The threshold parameter is used when the absolute value of the prediction error is less than or equal to... When the error is greater than 1, use mean squared error (MSE); when the error is greater than 1, use mean squared error (MSE). When using linear functions (similar to MAE), it is necessary to avoid being dominated by excessively large errors.

[0186] For risk probabilities (binary classification), binary cross-entropy loss can be used:

[0187] in, This represents the cross-entropy loss value for a single sample.

[0188] The training rounds can be several to hundreds of rounds, or the convergence / early stopping of the validation set error can be used as the criterion; the convergence criterion is that the validation set index does not improve significantly for several consecutive rounds or reaches a preset threshold.

[0189] In one specific embodiment, the prior lifetime inference model employs a Transformer-based architecture. Its specific structure is as follows:

[0190] Conditional sample encoding unit: Composed of an embedding layer and a positional encoding layer. The embedding layer maps fixed input parameters (such as model number), time features (such as day of the week, hour), and other categorical or numerical features into a high-dimensional vector. The positional encoding layer adds sequential information based on the temporal order of the samples (such as the sampling point of each day); the output dimension can be 64–512.

[0191] Functional Relationship Modeling Unit: Composed of multiple stacked Transformer encoders. Each layer contains a multi-head self-attention mechanism and a feedforward neural network with hidden layers ranging from 64 to 1024 dimensions, used to capture the complex nonlinear relationships and long-range dependencies within the input feature sequence and between it and the lifetime label. This unit outputs a hidden state vector containing contextual information;

[0192] The inference computation unit, consisting of a fully connected layer, maps the hidden state vector output by the function relationship modeling unit to the final lifetime prediction result (e.g., a scalar value representing the remaining lifetime, or a vector representing the probability of failure at various future time points). This unit can also output the confidence interval of the prediction result using techniques such as Monte Carlo Dropout.

[0193] To balance model expressiveness and generalization performance, the preferred parameter settings are as follows: Number of model layers: 2–12 layers, preferably 4–6 layers; Hidden dimension: 64–1024, preferably 256 or 512; Output embedding dimension: 128–256; Batch size: 32–256, preferably 128; Learning rate: 1e-3 to 1e-5, preferably 3e-4 (Adam optimizer); Loss function: a weighted combination of MAE and Huber loss; Early stopping criterion: validation set error shows no significant improvement for 10 consecutive rounds. The above parameter ranges can be adjusted according to sample size and hardware resources.

[0194] Step S400: Input conditional samples and perform inference.

[0195] This step applies the pre-trained model to various target devices in the field. First, for a target charging gun in the field, historical operating data for N consecutive days (e.g., *N=30* days) prior to its prediction time is obtained. Preprocessing and periodic encoding are performed according to step S200, and key causal features and degradation representations are extracted according to step S300, concatenating them into an input vector.

[0196] Sample sequence encoding and position mapping are used to characterize the chronological order and periodic position. The methods include one or a combination of the following: assigning a sequence index to each 15-minute time slice / each event; expanding the time index into tabular features (hour, minute, intraday sequence number, weekday, holiday, etc.); introducing periodic features (sine, cosine, or periodic index) to represent the daily / weekly periodic position; and setting field position identifiers for data from different sources (temperature, contact resistance, alarms, etc.) to achieve tabular column mapping.

[0197] The input vector is fed into a fixed prior lifetime inference model, and a forward propagation calculation is performed to directly obtain the lifetime prediction result.

[0198] The expression for the conditional lifetime inference relation is as follows:

[0199]

[0200] in, The input features for the predicted time include fixed parameters, time features, known historical time-sharing features, and degradation characterization features. This is a set of conditional samples consisting of historical data from the target device over N days (or N periods). As a priori lifetime inference model, the parameters remain fixed during the prediction phase and are not adjusted in any way based on the target device data. The lifetime estimate output can be RUL, failure risk, or degradation level.

[0201] Step S500: Output and Application

[0202] During the prediction phase, the model parameters remain fixed, without any adjustments or fine-tuning based on the target device. The lifespan prediction results are directly output, including one or a combination of the following: Remaining Life Estimate (RUL); Probability of Failure within the Next T Years; Degradation Level / Maintenance Recommendation Level. These can be used for triggering maintenance work orders, spare parts strategies, lifespan expiration warnings, warranty decisions, etc.

[0203] When the lifetime output is a risk / degradation sequence, it can be represented as:

[0204]

[0205] in, This can be the number of time slices within the prediction period. Indicates the first The failure risk / degradation value corresponding to each time slice.

[0206] When the output is a scalar of remaining lifetime, it can be expressed as:

[0207]

[0208] Taking a certain model of charging gun as an example:

[0209] Laboratory dataset: Conduct tests such as insertion and removal life, immersion-after hot insertion and removal, bending, holding force, and current-carrying aging, and collect data such as temperature, dT / dt, contact resistance trend, alarm / abnormal events, etc. to form a test sample set and failure criterion labels;

[0210] On-site data: Collect operational behavior data of equipment at several sites (number of plug-in / unplug-out times, power-on duration, current consumption), nozzle temperature and fluctuations, ambient temperature and humidity, alarm and maintenance records;

[0211] Preprocessing: Perform missing completion, anomaly handling, normalization, and temporal feature expansion according to step S200;

[0212] Sample organization: N days of historical data are used as conditional samples (preferably N=30 days), and RUL / risk / degradation level are used as response labels to construct training samples;

[0213] Pre-training: Complete the pre-training according to the data partitioning, objective function and optimization strategy in step S300 to obtain the prior lifetime inference model;

[0214] Field inference: Input the target equipment with the condition samples of the most recent N days, without performing any field training or parameter adjustment, and output the remaining lifespan or failure risk;

[0215] Verification and Application: The prediction results are used for maintenance strategy formulation; they can be compared and verified with actual alarm / replacement records after the prediction window ends to evaluate the applicability of the project.

[0216] It is evident that using laboratory experiments as the primary data source can significantly reduce reliance on field failure samples; no field training is required during the prediction phase, resulting in low deployment and maintenance costs; causal constraints suppress spurious correlations, improving generalization across operating conditions; and it supports two types of outputs: lifetime scalar and risk sequence, adapting to different operation and maintenance strategies.

[0217] This method can model the life degradation of key components of charging guns under complex operating conditions, predict their remaining life, and assess their failure risk.

[0218] Based on the same inventive concept, another embodiment provides a charging connector life prediction system that integrates accelerated life testing and field data, including a processor, a storage device, and a program stored on the processor, wherein when the program is executed by the processor, the processor implements the above-described charging connector life prediction method that integrates accelerated life testing and field data.

[0219] The system also includes a data acquisition module for performing the function of step S100 in the above method, including a laboratory accelerated life test data acquisition unit and a field operation data acquisition unit; it also includes the following modules stored in the memory and executed by the program:

[0220] The data preprocessing and mapping module is used to perform the functions of step S200 in the above method, including a missing value completion unit, an outlier handling unit, a normalization unit, a periodic encoding unit, and an acceleration factor mapping unit.

[0221] The model building module is used to perform the function of step S300 in the above method, including the causal constraint feature modeling unit, the degradation characterization unit and the prior lifetime inference model training unit.

[0222] The lifetime inference module is used for the function of step S400 in the above method. It includes a conditional sample construction unit and a model inference unit. The model inference unit has a pre-stored prior lifetime inference model with fixed parameters trained by the model construction module.

[0223] The output and application module is used to perform the function of step S500 in the above method, output the lifetime prediction results and generate corresponding operation and maintenance decision instructions.

[0224] Based on the same inventive concept, another embodiment of the present invention provides an electronic device storing computer instructions. When the computer instructions are executed, they run the aforementioned charging connector life prediction system integrating accelerated life testing and field data, and implement the aforementioned charging connector life prediction method integrating accelerated life testing and field data. Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The computer-readable storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.

[0225] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any indirect modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method of predicting the life of a charging connector by fusing accelerated life testing with field data, characterized by, Includes the following steps: Step S100: Construct laboratory lifespan dataset and field dataset; Step S200: Preprocess the two types of data and convert the laboratory degradation time into equivalent field usage time using the acceleration factor model; Step S300: Pre-train the prior lifetime inference model using pre-processed laboratory data, synthetic data, and mapped field equivalent time; Step S400: Input the historical data of the target device into the pre-trained prior life inference model with fixed parameters, and directly output the life prediction result. Step S500: Use the life prediction results for operation and maintenance decisions.

2. A method according to claim 1, characterized in that In step S100, multidimensional stress accelerated life tests are conducted on charging gun samples of the same model or series, and the temporal characteristics and failure time of each test sample under different stress conditions are recorded to form a laboratory life dataset. For the charging guns operating on-site, data on operational behavior, electrical and thermal characteristics, environmental and operating conditions, as well as fault and maintenance are collected according to the preset collection frequency and accuracy requirements to form an on-site operation dataset.

3. The method according to claim 1, characterized in that, The preprocessing in step S200 includes performing missing value completion, outlier handling, and normalization on the laboratory lifespan dataset and the field operation dataset, respectively. Time period encoding is introduced into the field operation dataset.

4. The method according to claim 3, characterized in that, In step S200, the expression for the on-site equivalent time is: in, As an acceleration factor, The degradation time under accelerated laboratory testing; or In the formula, This is the laboratory test temperature. The average operating temperature on site. For equivalent activation energy, Boltzmann's constant, This is the converted on-site equivalent operating time. For laboratory current, The average current at the site, This is the current acceleration index.

5. A method according to claim 1, characterized in that, The prior lifetime inference model in step S300 includes: The causal constraint feature modeling submodule is used to filter out the core feature set that has a stable causal relationship with lifespan degradation; The degradation characterization submodule is used to encode the original time series data at multiple scales and generate a vector representation that characterizes the degradation trajectory. The prior lifetime inference submodule is used to obtain lifetime mapping priors by pre-training with laboratory and synthetic data.

6. A method according to claim 5, characterized in that, The causal constraint feature modeling submodule includes constructing a candidate feature set, evaluating the causal relationship between candidate features and lifetime indicators, and determining key degradation factors according to preset rules. The degradation characterization submodule includes constructing local degradation segments from time-series data using a sliding window approach, encoding the data within the window, and generating a degradation characterization vector of fixed dimensions. The training sample pairs of the prior lifetime inference submodule include a condition part and a response part. The condition part consists of multi-source features, fixed parameters, and degradation characteristics for N consecutive days or N usage cycles. The response part consists of remaining lifetime, failure risk probability, or degradation degree indicators.

7. The method according to claim 1, characterized in that, The prior lifetime inference model described in step S300 adopts a Transformer-based architecture, including: Conditional sample encoding unit is used to map input features into high-dimensional embedding vectors; The functional relationship modeling unit, composed of multiple stacked Transformer encoders, is used to capture the complex nonlinear relationship between the input feature sequence and the lifetime label; The inference calculation unit is used to map the hidden state vector output by the functional relationship modeling unit to the final lifetime prediction result and output the confidence interval. The life prediction result is at least one of the following: remaining life estimate, failure risk probability sequence within the next T time period, degradation level, or maintenance recommendation level.

8. A charging connector life prediction system integrating accelerated life testing and field data, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the charging connector lifetime prediction method that integrates accelerated lifetime testing and field data as described in any one of claims 1 to 7.

9. A system according to claim 8, characterized in that, The system is also equipped with a data acquisition module for performing the function described in step S100 of any one of claims 1-7, including a laboratory accelerated life test data acquisition unit and a field operation data acquisition unit. The memory stores the following modules that can be executed by the at least one program: The data preprocessing and mapping module is used to perform the function described in step S200 of any one of claims 1-7, including a missing value completion unit, an outlier processing unit, a normalization unit, a periodic encoding unit, and an acceleration factor mapping unit. The model building module is used to perform the function described in step S300 of any one of claims 1-7, and includes a causal constraint feature modeling unit, a degradation characterization unit, and a prior lifetime inference model training unit. The lifetime inference module is used to perform the function described in step S400 of any one of claims 1-7, and includes a conditional sample construction unit and a model inference unit, wherein the model inference unit has a pre-stored prior lifetime inference model with fixed parameters trained by the model construction module. The output and application module is used to perform the function described in step S500 of any one of claims 1-7, output the lifetime prediction result and generate corresponding operation and maintenance decision instructions.

10. An electronic device storing computer instructions, characterized in that, When the computer instructions are executed, the charging connector life prediction system integrating accelerated life test and field data as described in claim 8 or 9 is run, and the charging connector life prediction method integrating accelerated life test and field data as described in any one of claims 1 to 7 is implemented.

Images