Dynamic life calibration method for adjustable kick type temperature controller

Abstract

The invention discloses a dynamic life calibration method for an adjustable kick type temperature controller, which comprises the following steps of: applying adjustable acceleration stress to the temperature controller, driving the temperature controller to circularly kick, and synchronously acquiring a multi-source parameter signal for reflecting a degradation state; performing segment processing on the multi-source parameter signal according to a kick period and a degradation stage, extracting a multi-source degradation characterization quantity and constructing a degradation feature vector sequence; performing multiple times of residual life prediction based on the sequence by a life prediction model to obtain a residual life prediction sequence and generate life trend characteristics; the acceleration stress parameters of the next stage are adaptively updated according to the life trend characteristics until failure judgment conditions are met, failure points are determined, residual life true values at all prediction moments are backfilled, a calibration sample set is established, and a calibration relation for mapping the prediction values into the true values is obtained through fitting; and calibrating the life prediction model based on the calibration relation and outputting a life calibration result. According to the method, the service life information acquisition efficiency and calibration convergence are improved, and the evaluation accuracy and consistency are improved.

Description

A method for dynamic life calibration of an adjustable snap-action temperature controller Technical Field

[0001] This invention relates to the field of thermostat reliability testing, and in particular to a dynamic life calibration method for an adjustable snap-action thermostat. Background Technology

[0002] Snap-on thermostats are commonly used as temperature control or overheat protection components in various applications such as home appliances, automobiles, battery thermal management, and industrial equipment. They achieve periodic switching of contacts through a snap-on mechanism to perform control or protection functions near a set temperature. During long-term cyclic snap-on operation, the thermostat is subjected to the combined effects of electrical load, ambient temperature changes, and repeated movement of the mechanism. This can gradually degrade the electrical performance of the contacts and the mechanical action characteristics, leading to problems such as misalignment, abnormal contact conditions, or shortened lifespan. Therefore, life testing and life assessment of snap-on thermostats are crucial steps in product reliability verification and parameter calibration.

[0003] In existing technologies, thermostat life assessment typically employs fixed or pre-defined accelerated aging test schemes. For example, the thermostat is driven to cycle under preset electrical load and environmental stress conditions, collecting signals such as voltage, current, and displacement, and the lifespan is terminated based on the achievement of preset failure criteria. There are also methods that build life prediction models based on historical test data to output remaining lifespan or lifespan trends. However, these methods mostly rely on fixed stress or offline models, resulting in weak interaction between the testing process and the life prediction process.

[0004] In practical applications, different individual temperature controllers and different degradation stages exhibit varying sensitivities to accelerated stress. Fixed or standardized acceleration schemes are prone to problems such as excessively long test cycles, low information acquisition efficiency, or decreased data validity due to mismatched stress settings. On the other hand, the remaining lifetime output by lifetime prediction models often deviates from the actual remaining lifetime, and this deviation varies with the degradation stage and feature evolution. Without calibration and verification of the mapping relationship between prediction and actual results, the reliability of lifetime assessment and threshold setting will be affected. Furthermore, in lifetime testing, samples may not fail within a preset maximum number of cycles, i.e., censored samples may appear. If the lower bound information of the actual lifetime provided by these censored samples is not utilized, the effectiveness of model calibration and strategy optimization will be reduced. Therefore, there is an urgent need for a lifetime calibration technique that can combine multi-source degradation characterization and remaining lifetime prediction results during testing, dynamically adjust accelerated stress, and calibrate the prediction results to improve the efficiency and consistency of lifetime calibration. Summary of the Invention

[0005] To address the common problems in existing life tests of snap-action temperature controllers, such as poor adaptability of fixed accelerated stress strategies, long test cycles, and low efficiency in acquiring life information, this invention proposes a dynamic life calibration method for adjustable snap-action temperature controllers. This method couples multi-source degradation characterization, multiple predictions of remaining life, and adaptive updates of accelerated stress in a closed loop during life testing. It also constructs a calibration relationship based on the actual remaining life backfilled from the pre-failure sequence, achieving dynamic calibration of life prediction results and improving life calibration efficiency. Simultaneously, a lower bound constraint on the actual remaining life value is introduced for samples that do not fail within a preset maximum number of snap-action cycles to enhance the reliability of calibration and strategy updates.

[0006] To achieve the above objectives, the present invention adopts the following technical solution, providing a dynamic life calibration method for an adjustable snap-action thermostat. The method includes: applying an adjustable acceleration stress to the thermostat and driving it to cyclically snap, simultaneously acquiring multi-source parameter signals reflecting the thermostat's degradation state; segmenting the multi-source parameter signals according to the snap-action cycle and degradation stage, extracting multi-source degradation characterization quantities, and constructing a degradation feature vector sequence, wherein the degradation feature vector sequence is a temporal arrangement sequence of degradation feature vectors obtained by vectorizing and combining the multi-source degradation characterization quantities; and performing multiple lifetime prediction operations based on the degradation feature vector sequence using a life prediction model. The remaining life prediction process generates a remaining life prediction sequence and generates life trend features to characterize the life evolution trend. Based on the life trend features, the accelerated stress parameters for the next stage are adaptively updated until the multi-source degradation characterization quantity meets the failure judgment condition. The failure point is determined and the actual remaining life value at each prediction time is backfilled. A calibration sample set consisting of degradation feature vector, remaining life prediction value, and actual remaining life value is established. A calibration relationship is fitted to map the remaining life prediction value to the actual remaining life value. The life prediction model is calibrated based on the calibration relationship, and the calibrated model outputs the life calibration result of the temperature controller.

[0007] As a further improvement to the above technical solution, the multi-source parameter signal includes at least electrical parameter signals and displacement parameter signals. The electrical parameter signals include voltage signals and current signals, and the displacement parameter signals include the displacement signals of contact jumps. The segmentation processing of the multi-source parameter signals according to the jump cycle and degradation stage includes: using the acquisition window corresponding to a single jump action as the jump cycle segmentation unit; performing time alignment and window truncation on the electrical parameter signals and displacement parameter signals; and dividing multiple jump cycles into at least two degradation stages based on the changing trend of degradation characterization quantities within adjacent jump cycles; extracting electrical characteristic quantities from the electrical parameter signals to characterize the electrical degradation state and extracting displacement characteristic quantities from the displacement parameter signals to characterize the mechanism degradation state within each segmentation unit; and using the electrical characteristic quantities and displacement characteristic quantities to constitute the multi-source degradation characterization quantities.

[0008] As a further improvement to the above technical solution, the remaining lifetime prediction performed by the lifetime prediction model based on the degradation feature vector sequence includes: extracting continuous subsequences from the degradation feature vector sequence based on a preset sliding window as model input, and outputting the remaining lifetime prediction value corresponding to the subsequence at each prediction time by the lifetime prediction model; updating the sliding window in a loop with sudden jumps and repeating the prediction to obtain a remaining lifetime prediction sequence arranged in time sequence.

[0009] As a further improvement to the above technical solution, the lifetime trend characteristics include the declining slope, volatility, moving average difference, inflection point indicator, and number of consecutive monotonically declining values ​​of the remaining lifetime prediction sequence; the volatility is the variance or absolute deviation of the remaining lifetime prediction sequence within the sliding window.

[0010] As a further improvement to the above technical solution, the adaptive update of the accelerated stress parameters for the next stage based on lifetime trend characteristics includes: constructing an accelerated stress strategy set based on historical lifetime test data, wherein the historical lifetime test data includes at least the accelerated stress parameter sequences and their lifetime response data corresponding to different jump cycles or degradation stages; the lifetime response data includes at least: the number of jumps corresponding to the failure point, the predicted remaining lifetime value at each prediction time, and the corresponding actual remaining lifetime value; using the degradation feature vector sequence, the remaining lifetime prediction sequence, and the lifetime trend characteristics as strategy inputs, and the strategy evaluation target as the training target, a strategy generation model for outputting the accelerated stress parameters for the next stage is obtained through training; and the accelerated stress parameters for the next stage are output by calling the strategy generation model based on the lifetime trend characteristics.

[0011] As a further improvement to the above technical solution, the strategy evaluation objective is determined by any one or a weighted combination of the following: a lifetime calibration efficiency index, used to characterize the efficiency of acquiring lifetime calibration information, and including at least one of the following: the number of jumps or time required to reach the failure determination condition, the number of jumps or time required to reach the preset remaining lifetime interval, and the convergence rate of the remaining lifetime prediction error or calibration error; a mechanism consistency constraint index, used to characterize the stability of the degradation process, and including at least one of the following: the constraint on the variation amplitude of multi-source degradation characterization quantities between adjacent jump cycles, the constraint on the number of abnormal event triggers, and the constraint on the acceleration stress update amplitude; a prediction consistency index, used to characterize the stability of multiple remaining lifetime predictions, and including at least one of the following: the volatility of the remaining lifetime prediction sequence within a preset window, the degree of monotonicity retention of the prediction results, and the variation amplitude of the prediction deviation before and after calibration; adaptively updating the acceleration stress parameters of the next stage also includes: calculating the consistency risk quantity based on the mechanism consistency constraint index or the prediction consistency index, and when the consistency risk quantity exceeds the risk threshold, performing at least one of the following operations: load reduction rollback operation, switching to a lower stress strategy, and pausing the jump cycle.

[0012] As a further improvement to the above technical solution, the acceleration stress parameter includes load current parameter, and at least one of temperature cycle parameter and jump cycle beat parameter; the temperature cycle parameter includes at least one of temperature swing, temperature rise and fall slope, and holding time; the jump cycle beat parameter includes at least one of cycle frequency or duty cycle.

[0013] As a further improvement to the above technical solution, the process of fitting the calibration relationship includes: based on the calibration sample set, using the predicted remaining lifetime as the independent variable and the actual remaining lifetime as the dependent variable, a mapping function is obtained by regression fitting to map the predicted remaining lifetime to the actual remaining lifetime; the mapping function is any one of a linear function, a piecewise linear function, or a nonlinear function, and a minimum constraint is applied to the mapping error during the fitting process.

[0014] As a further improvement to the above technical solution, the calibration relationship includes: dividing the lifetime degradation process into at least two degradation stages based on the change points of the degradation feature vector sequence or the remaining lifetime prediction sequence, and fitting corresponding calibration sub-relationships for different degradation stages to form a segmented lifetime calibration relationship.

[0015] As a further improvement to the above technical solution, the calibration of the lifetime prediction model based on the calibration relationship includes one of the following operations: performing output calibration processing on the remaining lifetime prediction value to obtain the calibrated remaining lifetime prediction value; determining the model calibration parameters based on the calibration relationship; updating the model parameters of the lifetime prediction model using the model calibration parameters; or updating the weight parameters of the multi-source degradation characterization quantity in the lifetime prediction model.

[0016] As a further improvement to the above technical solution, the method further includes: when the temperature controller reaches a preset maximum number of jumps... If the failure criteria are not met, the corresponding sample of the temperature controller is identified as a censored sample, and the lower bound of the true remaining lifetime value at each prediction time is determined based on the preset maximum number of jumps. When fitting the calibration relationship or calibrating the lifetime prediction model, a truncation constraint not lower than the lower bound is applied to the predicted remaining lifetime value that is lower than the true remaining lifetime value, or a penalty loss related to the degree of difference between the predicted remaining lifetime value and the lower bound is applied. For any prediction time... Corresponding number of completed jumps Lower bound of the true remaining lifetime .

[0017] Compared with existing technologies, this invention has at least the following beneficial effects: 1. By using a closed-loop coupling of multi-source degradation characterization, multiple remaining lifetime predictions, and lifetime trend-driven stress adaptive updates, the accelerated stress is dynamically matched with the degradation stage, improving the efficiency of lifetime information acquisition and shortening the lifetime calibration cycle; 2. By backfilling the true remaining lifetime in the near-failure sequence and fitting the calibration relationship, dynamic calibration of the lifetime prediction results is achieved, making the predicted remaining lifetime more consistent with the actual lifetime evolution law, and improving the accuracy and consistency of lifetime calibration; 3. By introducing consistency constraints and risk control mechanisms, the risk of abnormal fluctuations in the degradation process caused by stress updates is reduced, improving the stability of the experimental process and the reliability of the calibration results; 4. By utilizing the remaining lifetime lower bound information provided by censored samples and imposing constraints or penalties on predictions below the lower bound, the effectiveness and generalization ability of model calibration and strategy updates in scenarios with incomplete sample failure are enhanced. Attached Figure Description

[0018] Figure 1 is a flowchart of the dynamic lifetime calibration method provided in one embodiment of the present invention. Detailed Implementation

[0019] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be understood that the following embodiments are only for explaining the present invention and not for limiting the scope of protection of the present invention. Those skilled in the art can make various equivalent substitutions or modifications to the technical solutions of the present invention without departing from the spirit and substance of the present invention, and all such substitutions or modifications should fall within the scope of protection of the present invention. The order of the steps described in the present invention does not constitute a limitation on the order of implementation; the steps can be adjusted, combined, or split for execution while satisfying the purpose of the present invention. The terms "including" and "comprising" used herein are non-exclusive descriptions, indicating that other elements may be included in addition to the listed elements.

[0020] This invention provides a dynamic life calibration method for an adjustable snap-action temperature controller. This method can be implemented by a computing device, such as an industrial computer, controller, or host computer software, in conjunction with a test execution mechanism. The computing device processes the acquired multi-source parameter signals, constructs a degradation feature vector sequence, and performs life prediction and calibration. The test execution mechanism applies adjustable acceleration stress to the temperature controller and drives its cyclic snap-action. The steps, key parameters, and optional implementation methods of this invention will be described in detail below with several specific embodiments.

[0021] As shown in Figure 1, the dynamic lifetime calibration method provided by this invention specifically includes: applying an adjustable acceleration stress to a thermostat and driving it to cyclically jump, simultaneously acquiring multi-source parameter signals reflecting the degradation state of the thermostat; segmenting the multi-source parameter signals according to the jump cycle and degradation stage, extracting multi-source degradation characterization quantities, and constructing a degradation feature vector sequence, wherein the degradation feature vector sequence is a temporal arrangement sequence of degradation feature vectors obtained by vectorizing and combining the multi-source degradation characterization quantities; and performing multiple remaining lifetime predictions based on the degradation feature vector sequence using a lifetime prediction model to obtain the remaining lifetime. The lifespan prediction sequence is generated, and lifespan trend features are used to characterize the lifespan evolution trend. Based on the lifespan trend features, the accelerated stress parameters of the next stage are adaptively updated until the multi-source degradation characterization quantity meets the failure judgment condition. The failure point is determined and the actual remaining lifespan value at each prediction time is backfilled. A calibration sample set consisting of degradation feature vector, predicted remaining lifespan value and actual remaining lifespan value is established. A calibration relationship is fitted to map the predicted remaining lifespan value to the actual remaining lifespan value. The lifespan prediction model is calibrated based on the calibration relationship. The calibrated model outputs the lifespan calibration result of the temperature controller.

[0022] During the cyclical jump of the temperature controller driven by the application of adjustable accelerated stress, this invention acquires multi-source parameter signals in real time and constructs a degradation feature vector sequence. The lifetime prediction model outputs multiple remaining lifetime prediction sequences based on this sequence and further obtains lifetime trend characteristics. The system adaptively adjusts the accelerated stress parameters of the next stage according to the lifetime trend characteristics, so that the test process can obtain effective lifetime information more quickly while ensuring feasibility. When the failure judgment condition is met, the failure point is determined, the true value of the remaining lifetime at each prediction time is backfilled, the calibration relationship between the predicted value and the true value is obtained by fitting, and the lifetime prediction model is calibrated using this calibration relationship, thereby outputting the lifetime calibration result.

[0023] Adjustable accelerated stress refers to external stress conditions applied to shorten life test time and accelerate degradation evolution. It can manifest as electrical stress, thermal stress, or a combination thereof. This invention does not limit the specific implementation form, only requiring that the stress can be adjusted during the test process and can drive the temperature controller to perform cyclic jump actions. During the test, the aforementioned adjustable accelerated stress is applied to the temperature controller and it is made to repeatedly jump according to a set rhythm, thereby forming a cyclic process of life test.

[0024] Multi-source parametric signals are a set of observed signals used to reflect the degradation state of a thermostat. They can include at least electrical and displacement parameters, such as voltage, current, and contact jump displacement. Acquisition methods can include continuous sampling or event-triggered sampling, and the acquisition frequency can be determined based on the jump frequency and signal bandwidth. To ensure consistency in subsequent feature extraction, the acquired data is usually timestamped or recorded synchronously under the same acquisition clock.

[0025] The purpose of jump period segmentation is to divide a continuous signal into alignable samples based on a single jump. In implementation, the acquisition window corresponding to a single jump action can be used as the basic segmentation unit: for example, using contact closure / opening events, displacement change points, or electrical parameter transition points as window boundaries, and extracting signal segments of a preset duration before and after the jump; thus, each segmentation unit corresponds to one jump action process.

[0026] The purpose of segmenting the degradation stages is to distinguish the differences in evolutionary characteristics of the thermostat degradation process at different stages. In practice, the stages can be divided based on the changing trends of the characteristic quantities extracted from adjacent jump cycles. For example, when certain characteristic quantities show continuous drift, abrupt changes in slope, or significant fluctuations, it can be determined that a new degradation stage has been entered.

[0027] Multi-source degradation characterization quantities refer to the set of feature quantities extracted from the multi-source parametric signals of each segment unit to quantify the degradation state. In practice, feature quantities reflecting the electrical state can be extracted from electrical parameter signals, such as statistics, energy-related quantities, and on / off characteristic-related quantities obtained from voltage / current waveforms. Feature quantities reflecting the mechanism's operational state can be extracted from displacement parameter signals, such as displacement amplitude, response time, and rebound characteristic-related quantities. The selection of characterization quantities can be configured based on the temperature controller structure and test conditions, but all fall under the category of extracting quantifiable degradation information from multi-source signals.

[0028] A degradation feature vector is a feature representation obtained by vectorizing and combining multiple degradation characteristics extracted within the same segmented unit. For example, several characteristics can be arranged into a vector according to a preset dimension order. Correspondingly, a degradation feature vector sequence is a sequence obtained by arranging the degradation feature vectors corresponding to each jump cycle segmented unit in chronological order. This sequence is used to describe the temporal evolution of thermostat degradation with jump cycles.

[0029] The lifespan prediction model is used to output the predicted remaining lifespan of a thermostat based on a degradation feature vector sequence. This invention does not limit the specific implementation type of the model; it can be a statistical model, a machine learning model, or a deep learning model, etc., as long as it can utilize time-series feature inputs and output remaining lifespan prediction results. To facilitate real-time updates and trend analysis, the model can be run multiple times during the experiment.

[0030] Multiple remaining lifetime predictions refer to repeatedly performing lifetime predictions at different prediction times, such as after each abrupt cycle or after each time window, so that the model output forms a set of remaining lifetime prediction values ​​arranged in time sequence. The resulting remaining lifetime prediction sequence reflects the change of remaining lifetime with degradation evolution, rather than just the result of a single prediction.

[0031] After obtaining the remaining lifetime prediction sequence, lifetime trend features are further generated to characterize the lifetime evolution trend. These features can be calculated from the changing patterns of the prediction sequence, such as the rate of decline, the degree of fluctuation, moving average changes, inflection points, or continuous monotonic changes. The role of lifetime trend features is to compress multiple prediction results into a decision-making basis for controlling accelerated stress updates, thus linking the next stage of stress updates with the lifetime evolution trend.

[0032] Adaptive acceleration stress parameter updates refer to adjusting stress conditions for the next stage based on lifetime trend characteristics. In implementation, lifetime trend characteristics can be mapped to stress update instructions. For example, stress can be increased to accelerate degradation when the lifetime decline rate is slow, and stress can be decreased when the lifetime decline rate is too fast or fluctuates abnormally to avoid test runaway or information distortion. The mapping rules here can be rule-based strategies or model-based strategies, but both fall under the scope of updating stress based on lifetime trend characteristics.

[0033] Failure criteria are used to determine the end of a thermostat's lifespan. They can be based on the abnormality of multiple degradation indicators reaching or exceeding thresholds. For example, when certain degradation indicators show that the thermostat can no longer meet normal operation or safe working requirements, failure is determined. When the multiple degradation indicators meet the failure criteria, the failure point is identified. The failure point can be characterized by the number of jumps or the time elapsed when the failure criteria are met, serving as a reference for end of lifespan.

[0034] After determining the failure point, the remaining lifetime is backfilled with the true value for each predicted time point. The true value is a quantified result determined from the failure point, based on the actual remaining lifetime at each previous predicted time point. In practice, the true remaining lifetime value at the predicted time point can be obtained based on the difference in the number of jumps or the time difference between the failure point and the corresponding predicted time point. This establishes a correspondence between predicted and true values ​​before failure. A calibration sample set is then established, consisting of at least three parts: a degradation feature vector, predicted remaining lifetime values, and actual remaining lifetime values. A calibration relationship is fitted based on this calibration sample set. This calibration relationship maps the predicted remaining lifetime values ​​to the actual remaining lifetime values, thereby compensating for the systematic changes in prediction bias under different degradation stages and different characteristic evolution states. The calibration relationship can be understood as being implemented by a type of mapping function or calibration model.

[0035] After obtaining the calibration relationship, the lifetime prediction model is calibrated based on this relationship. Calibration can be applied to the model output, such as mapping and correcting the predicted remaining lifetime value, or it can be applied to the model parameters, such as determining calibration parameters based on the calibration relationship and updating the model to make its subsequent predictions more closely match the actual remaining lifetime. The calibrated model outputs the lifetime calibration results of the temperature controller. The lifetime calibration results can include lifetime assessment values, estimated remaining lifetime values, or lifetime curves of the temperature controller under the current test conditions and degradation path, which can be used for subsequent reliability evaluation, threshold setting, or product consistency analysis.

[0036] In one embodiment of the present invention, the multi-source parametric signals include at least electrical parametric signals and displacement parametric signals. The electrical parametric signals include voltage signals and current signals, and the displacement parametric signals include the displacement signal of a contact jump. The voltage and current signals can be obtained by connecting a voltage / current acquisition module in series with the temperature controller circuit, and the displacement signal can be obtained by setting a displacement sensor on the contact or a reed linked to the contact. During acquisition, the three signals can be timestamped using the same clock source to ensure alignment consistency during subsequent segmented processing.

[0037] The segmentation processing of multi-source parametric signals according to the jump cycle and degradation stage includes: using the acquisition window corresponding to a single jump action as the jump cycle segmentation unit, performing time alignment and window truncation on the electrical parameter signal and displacement parameter signal, and dividing multiple jump cycles into at least two degradation stages based on the changing trend of degradation characteristics within adjacent jump cycles. In this embodiment, the acquisition window can be set as a time window around a single jump event, with a preset duration before the event plus a preset duration after the event. The jump event can be determined by the abrupt change point of the displacement signal (e.g., the moment when the first-order difference of the displacement exceeds a threshold) or by the on / off transition point of the electrical parameter signal (e.g., the moment when the current rise / fall edge exceeds a threshold). Time alignment can be achieved by interpolating and resampling the sampling timestamps of different channels or by zero-phase delay compensation, so that the voltage, current, and displacement waveforms within the same jump cycle correspond on the same time axis. Window truncation involves extracting corresponding segments from the continuous signal according to the jump event time to form the segmented data of the jump cycle. Furthermore, to achieve the division of degradation stages, a set of degradation characteristics can be calculated on the segmented data of each jump cycle, and trend analysis can be performed on the degradation characteristics sequence of adjacent jump cycles. When the degradation characteristics show trend characteristics such as continuous monotonic drift, abrupt slope change, or significant change in volatility, multiple jump cycles can be divided into at least two degradation stages, such as an early stable stage and a near-failure stage.

[0038] Within each segmented unit, electrical characteristic quantities characterizing the electrical degradation state are extracted from electrical parameter signals, and displacement characteristic quantities characterizing the mechanism degradation state are extracted from displacement parameter signals. These electrical and displacement characteristic quantities constitute a multi-source degradation characterization quantity. The electrical characteristic quantities can be statistical or process quantities obtained from voltage / current waveforms, such as the peak value of switching transients, the duration of rising / falling edges, the energy integral of a single jump cycle (e.g., obtained by integrating the product of voltage and current over time), or equivalent conduction characteristic quantities (e.g., the characterization value calculated from the voltage-current relationship). The displacement characteristic quantities can be statistical or dynamic quantities obtained from displacement waveforms, such as the jump displacement amplitude, response delay, rebound amplitude, oscillation decay time, or peak displacement velocity. The electrical and displacement characteristic quantities obtained within the same segmented unit are vectorized and combined according to a preset dimensional order to obtain the multi-source degradation characterization quantity for that segmented unit. This quantity is used to subsequently construct a degradation characteristic vector sequence and support lifetime prediction and calibration.

[0039] In one embodiment of the present invention, the remaining lifetime prediction performed by the lifetime prediction model based on a degradation feature vector sequence includes: extracting continuous subsequences from the degradation feature vector sequence using a preset sliding window as model input; and outputting the remaining lifetime prediction value corresponding to the subsequence at each prediction time. The preset sliding window can be set as a sequence segment containing the degradation feature vectors of the most recent L consecutive jump cycles, where L can be selected as a fixed or adaptive value based on the jump cycle and degradation rate. Each prediction time can be set to be triggered once after a preset number of jump cycles are completed, or once when the degradation feature vector sequence is updated to meet the window length. The lifetime prediction model receives the subsequences and outputs the corresponding remaining lifetime prediction value, which can represent the expected number of remaining jumps or the expected remaining time.

[0040] The sliding window is updated cyclically with each jump, and predictions are repeatedly executed to obtain a time-series sequence of remaining lifetime predictions. The sliding window can be updated using a fixed-step update method, where the window slides forward by the same number of feature vectors after each one or several jump cycles, discarding earlier feature vectors and introducing the latest ones; or using an event-triggered update method, where window updates are triggered when the change in a degraded feature vector exceeds a preset threshold. After each window update, the lifetime prediction model is repeatedly executed to output new remaining lifetime prediction values, thus forming a time-series sequence of remaining lifetime predictions that is cyclically updated with each jump. This sequence is then used to generate lifetime trend features and drive adaptive updates of acceleration stress parameters.

[0041] In this embodiment, the lifetime prediction model can be instantiated as a sequence modeling network for processing temporal degradation feature vector sequences, which includes at least an input layer, a temporal feature extraction layer, and a lifetime regression output layer. Specifically, the input layer is used to receive continuous subsequences truncated by a preset sliding window, wherein the subsequences can be represented as L×D matrices, where L is the number of hop cycles contained in the sliding window, and D is the dimension of a single degradation feature vector; the temporal feature extraction layer is used to extract temporal degradation features that evolve with the hop cycles from the L×D input; the lifetime regression output layer is used to map the temporal degradation features to a remaining lifetime prediction value, wherein the remaining lifetime prediction value can be characterized as the expected number of remaining hops or the expected remaining time.

[0042] In one optional implementation, the temporal feature extraction layer can be implemented using a recurrent neural network structure, such as a temporal encoder composed of a gated recurrent unit (GRU) or a long short-term memory (LSTM) network. The temporal encoder progressively updates the hidden states of the degenerate feature vectors within the sliding window in chronological order, obtaining the final hidden state or a sequence of hidden states. The lifetime regression output layer can perform a fully connected mapping on the final hidden state to output the predicted remaining lifetime at the current prediction time, or perform weighted aggregation on the hidden state sequence to output the predicted remaining lifetime. Through this structure, the model can capture the degradation evolution pattern by utilizing the temporal dependencies in the degenerate feature vector sequence, thereby achieving remaining lifetime prediction based on the degenerate feature vector sequence.

[0043] In another optional implementation, the temporal feature extraction layer can be implemented using a one-dimensional convolutional temporal network. Specifically, a one-dimensional convolution is performed on the L×D input along the time dimension to extract local degradation evolution patterns, and the receptive field is expanded through multiple convolutions or dilated convolutions to characterize degradation trends over longer time scales. Subsequently, a lifetime regression output layer converges and maps the convolutional features to output predicted remaining lifetime values. This structure enables the extraction of trend and fluctuation patterns from a sequence of degradation feature vectors with relatively low computational complexity.

[0044] In a further optional implementation, the temporal feature extraction layer can also be implemented using a self-attention structure. This involves using the degradation feature vector within the sliding window as the input to a query-key-value pair, and adaptively selecting time segments that contribute more to remaining lifetime prediction through attention weights, thereby generating a global temporal representation for lifetime regression. The lifetime regression output layer then outputs the remaining lifetime prediction value based on this global representation. The self-attention structure enhances the model's ability to model the nonlinear changes and stage-specific characteristics of the degradation process.

[0045] To achieve multiple remaining lifetime predictions, at each prediction time, a subsequence truncated by the sliding window is input into the lifetime prediction model to obtain the corresponding remaining lifetime prediction value. The sliding window is updated cyclically with each jump, and the prediction is repeated to obtain a time-series sequence of remaining lifetime predictions. If necessary, numerical constraints can be imposed on the model output to ensure the reasonableness of the remaining lifetime prediction values. For example, the output can be limited to non-negative values, or the output can be converted to a uniform dimension of remaining jump count / remaining time using a preset scale, so as to subsequently generate lifetime trend features and drive the adaptive update of acceleration stress parameters.

[0046] In one embodiment of the present invention, the lifetime trend characteristics include the declining slope, volatility, moving average difference, inflection point indicator, and number of consecutive monotonically declining values ​​of the remaining lifetime prediction sequence. Specifically, the declining slope can be determined by the linear fitting slope of the remaining lifetime prediction sequence within a preset time period or a preset number of jumps, characterizing the rate of decline of remaining lifetime with each jump cycle; the moving average difference can be determined by the difference between the moving average of the remaining lifetime prediction values ​​within the current sliding window and the moving average of the previous window or historical benchmark window, characterizing the stage-wise drift of the lifetime trend; the inflection point indicator can be determined by detecting changes in the slope sign, abrupt changes in slope amplitude, or moments when the second-order difference exceeds a threshold in the remaining lifetime prediction sequence, indicating the risk moment when a significant turning point in the lifetime evolution trend occurs; the number of consecutive monotonically declining values ​​can be determined by statistically analyzing the number of prediction points in the remaining lifetime prediction sequence that consecutively satisfy "the subsequent prediction value is not greater than the previous prediction value," characterizing the persistence and consistency of lifetime decline.

[0047] The volatility refers to the variance or absolute deviation of the remaining lifetime prediction sequence within a sliding window. The sliding window can be consistent with the window used in multiple remaining lifetime predictions or set according to a preset ratio. The variance measures the dispersion of each remaining lifetime prediction value within the window relative to its mean, while the absolute deviation measures the deviation of each remaining lifetime prediction value within the window relative to its mean or median. This variation characterizes the strength of the remaining lifetime prediction sequence within a local time period and serves as one of the bases for subsequent adaptive updates of the acceleration stress parameters.

[0048] In one embodiment of the present invention, adaptively updating the accelerated stress parameters for the next stage based on lifespan trend characteristics includes: constructing an accelerated stress strategy set based on historical lifespan test data, wherein the historical lifespan test data includes at least accelerated stress parameter sequences and their lifespan response data corresponding to different jump cycles or degradation stages; the lifespan response data includes at least: the number of jumps corresponding to the failure point, the predicted remaining lifespan at each prediction time, and the corresponding actual remaining lifespan value. The historical lifespan test data can be derived from lifespan test records of the same model or series of temperature controllers under different accelerated stress conditions. The accelerated stress parameter sequences are used to characterize the setting and adjustment process of stress parameters during the jump cycle, and the lifespan response data is used to characterize the lifespan termination position of the temperature controller under the stress parameter sequence and the correspondence between predicted and actual lifespan. Based on the above data, an accelerated stress strategy set can be formed, which may contain multiple sets of selectable stress parameter sequence samples, or contain strategy rule samples for generating stress parameter sequences, to provide a sample basis for subsequent strategy generation model training.

[0049] Using degradation feature vector sequences, remaining lifetime prediction sequences, and lifetime trend features as policy inputs, and a policy evaluation objective as the training objective, a policy generation model is obtained through training to output the acceleration stress parameters for the next stage. The policy generation model can be instantiated as either a supervised learning model or a reinforcement learning model: In supervised learning, the degradation feature vector sequences, remaining lifetime prediction sequences, and lifetime trend features can be used as input features, the actual acceleration stress parameters used in the next stage from historical experiments can be used as supervision labels, and the policy evaluation objective can be used as a component of the loss function or objective function for training; In reinforcement learning, the above inputs can be used as state representations, the acceleration stress parameters for the next stage can be used as action outputs, and a reward function can be constructed using the policy evaluation objective, enabling the model to learn to output better stress parameters under different degradation states and lifetime trends. After training, the policy generation model can output the acceleration stress parameters for the next stage given the current degradation state and lifetime trend. The acceleration stress parameters can be a single parameter value or a parameter vector, used to represent the acceleration stress setting for the next stage.

[0050] The strategy generates the next-stage accelerated stress parameters based on the lifetime trend characteristics. This generation can be either periodically or by events. In periodically triggered mode, the generation can occur after a preset number of jump cycles or after each lifetime prediction update. In event triggered mode, the generation can occur when the lifetime trend characteristics meet preset conditions, such as when the descent slope or volatility reaches a preset threshold, when an inflection point indicator is triggered, or when the number of consecutive monotonic decreases reaches a preset number. The accelerated stress parameters output by the strategy generation model are used to update the accelerated stress applied to the temperature controller by the test actuator, thus achieving adaptive stress updates as the degradation state and lifetime trend change during the lifetime test.

[0051] In one embodiment of the present invention, the strategy evaluation objective is determined by any one of the following or by a weighted combination of multiple of them, including lifetime calibration efficiency index, mechanism consistency constraint index, and prediction consistency index.

[0052] Lifetime calibration efficiency indicators are used to characterize the efficiency of acquiring lifetime calibration information, and include at least one of the following: the number of jumps or time required to reach the failure determination condition, the number of jumps or time required to reach the preset remaining lifetime interval, and the convergence rate of the remaining lifetime prediction error or calibration error. Specifically, the number of jumps or time required to reach the failure determination condition can be directly obtained by statistically analyzing the cumulative number of jumps or the cumulative duration from the start of the experiment to the fulfillment of the failure determination condition; the number of jumps or time required to reach the preset remaining lifetime interval can be determined by the cumulative number of jumps or the cumulative duration corresponding to the prediction time when the remaining lifetime prediction sequence first falls into the preset interval; the convergence rate can be determined by the decrease magnitude, decrease slope, or number of steps to reach the preset error threshold of the prediction error or calibration error at adjacent prediction times within a preset window, thereby measuring the speed of acquiring effective lifetime information during the experiment.

[0053] Mechanism consistency constraints are used to characterize the stability of the degradation process and include at least one of the following: constraints on the variation amplitude of multi-source degradation characteristics between adjacent abrupt change cycles, constraints on the number of abnormal event triggers, and constraints on the magnitude of accelerated stress updates. Specifically, the variation amplitude of multi-source degradation characteristics between adjacent abrupt change cycles can be measured by the norm, relative rate of change, or statistical fluctuation range of the difference between degradation characteristics in adjacent cycles, and compared with a preset upper limit to reflect whether unexpected abrupt changes occur in the degradation evolution; the number of abnormal event triggers can be counted against preset abnormal event conditions, which may include any observable abnormality such as abrupt changes in the acquired signal, missing displacement action, on / off anomalies, or abnormal jumps in model output; the constraint on the magnitude of accelerated stress updates sets an upper limit on the change in accelerated stress parameters between adjacent stages to avoid deviation of the degradation path or experimental instability due to excessive stress updates.

[0054] A prediction consistency index is used to characterize the stability of multiple remaining lifetime predictions, and includes at least one of the following: the volatility of the remaining lifetime prediction sequence within a preset window, the degree of monotonicity retention of the prediction results, and the magnitude of change in prediction deviation before and after calibration. Volatility can be determined by the variance, absolute deviation, or range of the remaining lifetime prediction values ​​within the preset window; the degree of monotonicity retention can be determined by statistically analyzing the proportion of prediction points within the preset window that violate the rule that "remaining lifetime should not increase with degradation," the number of consecutive monotonic decreases, or the monotonicity score; the magnitude of change in prediction deviation before and after calibration can be determined by the difference, ratio, or decrease in the prediction error before and after calibration, reflecting the degree of stable improvement in the prediction output with calibration updates.

[0055] In this embodiment, the above-mentioned strategy evaluation objective can be determined by a combination of multiple weighted factors. Specifically, the lifetime calibration efficiency index, mechanism consistency constraint index and prediction consistency index can be normalized and then weighted and summed or compared according to preset weights. The weights can be set to fixed or adjustable values ​​according to the experimental stage, risk preference or calibration objective, so that the strategy generation model can take into account both experimental efficiency and process stability during training and invocation.

[0056] In addition, adaptively updating the accelerated stress parameters for the next stage may also include: calculating the consistency risk based on the mechanistic consistency constraint index or the predicted consistency index; when the consistency risk exceeds the risk threshold, performing at least one of the following operations: load reduction rollback, switching to a lower stress strategy, or pausing the jump cycle. The consistency risk can be determined by the degree to which the mechanistic consistency constraint index or the predicted consistency index exceeds its constraint upper limit; the risk threshold can be set based on historical experimental statistical distribution, a preset safety margin, or an empirical threshold; the load reduction rollback operation can manifest as reducing the magnitude of the accelerated stress parameters for the next stage or restoring them to the stress level of the previous stage; switching to a lower stress strategy can manifest as selecting a stress update path that is more tolerant of the degradation process; pausing the jump cycle can be used to terminate or delay the execution of the next stage update and record the abnormal event information when an abnormal event occurs, for subsequent analysis and model correction.

[0057] In one embodiment of the present invention, the accelerated stress parameters include load current parameters, and at least one of temperature cycle parameters and jump cycle timing parameters. The load current parameter characterizes the electrical load intensity borne by the thermostat contacts during the on / off process, and can be set and adjusted by an adjustable power supply, electronic load, or equivalent load circuit. It can be set to different current levels at different test stages to form adjustable accelerated stress. The temperature cycle parameter characterizes the cyclical change of the ambient temperature around the thermostat over time, and can be controlled in a closed loop by a heating / cooling unit to achieve heating, cooling, and holding according to a set curve. The jump cycle timing parameter characterizes the rhythm of the cyclic jumps that drive the thermostat, and can be controlled by adjusting the applied accelerated stress rhythm or a trigger signal to make the thermostat operate periodically at a set frequency. It can also be dynamically adjusted during the test based on lifespan trend characteristics.

[0058] The temperature cycling parameters include at least one of temperature swing amplitude, temperature rise / fall rate, and hold time; the jump cycle timing parameters include at least one of cycle frequency or duty cycle. Specifically, temperature swing amplitude can be expressed as the temperature difference range between the high-temperature and low-temperature plateaus of the temperature cycle; temperature rise / fall rate can be expressed as the rate of temperature change from the low-temperature plateau to the high-temperature plateau or vice versa; hold time can be expressed as the duration for which the temperature remains stable on the high-temperature plateau and / or the low-temperature plateau; cycle frequency can be expressed as the number of jump cycles completed per unit time; and duty cycle can be expressed as the proportion of time occupied by the on-state and off-state within one cycle of the jump cycle, or the high / low level ratio of the trigger pulse. By combining and setting the above load current parameters, temperature cycling parameters, and jump cycle timing parameters, accelerated stress conditions of different intensities and forms can be formed, and the accelerated stress parameters for the next stage can be updated as needed during the life calibration process to drive the degradation process.

[0059] In one embodiment of the present invention, the process of fitting the calibration relationship includes: based on a calibration sample set, using the predicted remaining lifetime as the independent variable and the actual remaining lifetime as the dependent variable, a mapping function is obtained through regression fitting to map the predicted remaining lifetime to the actual remaining lifetime. The calibration sample set consists of multiple sets of samples, each corresponding to a prediction time. Each sample contains the degradation feature vector at that prediction time, the predicted remaining lifetime output by the lifetime prediction model, and the actual remaining lifetime determined by failure point backfilling. To facilitate regression fitting, the predicted remaining lifetime at each prediction time in the sample set can be used to form a sequence of independent variables, and the corresponding actual remaining lifetime can be used to form a sequence of dependent variables. After preprocessing outliers or missing values, regression fitting is performed to obtain the parameters or structure of the mapping function.

[0060] In this embodiment, the mapping function is any one of a linear function, a piecewise linear function, or a nonlinear function, and a minimization constraint is applied to the mapping error during the fitting process. The linear function can be used to characterize the overall proportional deviation and bias term between the predicted and actual values. The piecewise linear function can be used to characterize the differences in the mapping relationship within different predicted value intervals, for example, using different slopes / intercepts in the early lifespan and near-failure intervals. The nonlinear function can be used to characterize the nonlinear law of the prediction deviation changing with the lifespan interval; the nonlinear function can be any one of a polynomial function, an exponential / logarithmic function, or a nonlinear regression function implemented by a neural network. The mapping error can be characterized by the difference between the remaining lifespan estimate output by the mapping function and the actual remaining lifespan value in the sample, and a minimization constraint is applied through least squares, absolute deviation minimization, or weighted error minimization, so that the obtained mapping function can reduce the overall mapping deviation between the predicted and actual remaining lifespan values, thereby forming a calibration relationship for subsequent lifespan prediction model calibration.

[0061] In one embodiment of the present invention, the structured calibration relationship further includes: dividing the lifetime degradation process into at least two degradation stages based on the change points of the degradation feature vector sequence or the remaining lifetime prediction sequence, and fitting corresponding calibration sub-relationships for different degradation stages to form a segmented lifetime calibration relationship. The change points can be determined by the locations where the statistical characteristics of the degradation feature vector sequence or the remaining lifetime prediction sequence undergo significant changes. For example, when the change amplitude of certain dimensions of the degradation feature vector increases sharply between adjacent jump cycles, or the trend slope changes significantly, or when the remaining lifetime prediction sequence shows inflection point indication triggering, a sudden change in the decreasing slope, or a significant increase in volatility, this location can be determined as a change point. Based on the change points, the lifetime degradation process can be divided into at least two degradation stages, such as an early degradation stage and a late degradation stage, or further divided into at least two stages among a stable stage, an accelerated degradation stage, and a near-failure stage.

[0062] After completing the degradation stage division, calibration samples falling within each degradation stage can be selected to form a stage sub-sample set. Regression fitting is then performed on each stage sub-sample set using the predicted remaining lifetime as the independent variable and the actual remaining lifetime as the dependent variable, thereby obtaining the corresponding calibration sub-relationship for that stage. The calibration sub-relationship can be any of a linear, piecewise linear, or nonlinear mapping function, and different function forms or parameter values ​​can be used for the calibration sub-relationships of different degradation stages. Finally, the calibration sub-relationships corresponding to each degradation stage are combined according to the chronological order of the degradation stages, with the point of change serving as the segment boundary, forming a piecewise lifetime calibration relationship. This relationship is used for stage-matching mapping calibration of the predicted remaining lifetime at different degradation stages.

[0063] In one embodiment of the present invention, calibrating the lifetime prediction model based on the calibration relationship includes one of the following operations: performing output calibration processing on the remaining lifetime prediction value to obtain the calibrated remaining lifetime prediction value; determining the model calibration parameters based on the calibration relationship; updating the model parameters of the lifetime prediction model using the model calibration parameters; or updating the weight parameters of the multi-source degradation characterization quantity in the lifetime prediction model.

[0064] For the previous calibration operation, the output calibration processing can be achieved by inputting the remaining lifetime prediction value output by the lifetime prediction model at the current prediction time into the mapping function corresponding to the calibration relationship to obtain the calibrated remaining lifetime prediction value. When the calibration relationship is a segmented calibration relationship, the corresponding calibration sub-relationship can be selected according to the degradation stage at the current prediction time, and then the mapping calibration can be performed on the remaining lifetime prediction value to obtain a calibrated output that is more consistent with the actual remaining lifetime. This calibrated output can be used as a component of the lifetime calibration result or as one of the inputs for subsequent lifetime trend feature calculation.

[0065] For the subsequent calibration operation, the model calibration parameters can be determined by the function parameters, piecewise boundary parameters, or nonlinear mapping structure parameters of the calibration relationship. For example, when the calibration relationship is a linear mapping function, its slope and intercept can be used as model calibration parameters; when the calibration relationship is a piecewise mapping function, the mapping parameters of each segment and the segment boundaries can be used as model calibration parameters. Updating the model parameters of the lifetime prediction model using the model calibration parameters can be achieved by applying calibration adjustments to the parameters of the model output layer or by retraining / fine-tuning the overall model parameters, so that the remaining lifetime prediction value output by the model under the same degradation feature input is more in line with the calibrated mapping law. Updating the weight parameters of the multi-source degradation characterization in the lifetime prediction model can be achieved by adjusting the input weights, attention weights, or feature fusion weights of each dimension of degradation characterization, so as to enhance the contribution of key degradation characterization related to lifetime deviation, thereby improving the consistency between the subsequent prediction output and the actual lifetime.

[0066] In yet another embodiment provided by the present invention, the dynamic life calibration method further includes: when the temperature controller reaches a preset maximum number of jumps... If the failure determination condition is not met, the sample corresponding to the temperature controller is identified as a censored sample, and the lower bound of the true value of the remaining lifetime at each prediction time is determined according to the preset maximum number of jumps. When fitting the calibration relationship or calibrating the lifetime prediction model, a truncation constraint not lower than the lower bound is applied to the remaining lifetime prediction value that is lower than the lower bound of the true value of the remaining lifetime, or a penalty loss related to the degree of difference between the remaining lifetime prediction value and the lower bound is applied.

[0067] In this embodiment, a maximum number of jumps is preset. The maximum test duration for a single sample can be preset by the test plan or equipment capabilities; when the sample is completed If the failure condition is not triggered after the second jump cycle, it can be considered that the true failure point of the sample has not been observed, but its true lifetime can be determined to be at least greater than or equal to 1. The corresponding lifetime length can be used to form the lower bound of the true remaining lifetime value at each prediction time, based on the information that the device has not yet failed, for subsequent calibration relationship fitting or model calibration processes.

[0068] In this embodiment, the truncation constraint can be executed at the model output or the calibration relation mapping end. That is, when the predicted remaining lifetime value at a certain prediction time is less than the corresponding lower bound, the predicted value is adjusted to a value not less than the lower bound to avoid violating the known lower lifetime limit of the censored samples. The penalty loss can be introduced in the objective function of regression fitting or model parameter update. When the predicted remaining lifetime value is lower than the corresponding lower bound, the loss weight or penalty term is increased according to the degree of difference from the lower bound, thereby driving the fitting or calibration process to tend to output a remaining lifetime prediction result not lower than the lower bound. When the predicted value is not lower than the lower bound, the penalty effect is not applied or is weakened.

[0069] For any prediction time Corresponding number of completed jumps The lower bound of the true remaining lifetime can be expressed as: ;in, It can be obtained from the count of the jump cycles that have been completed during the experiment. Indicates the time of prediction Based on "at most it can continue to execute until The lower bound of the remaining lifetime that can be determined by the "secondary jump". For example, when Less than hour, for and The difference is used to constrain the predicted remaining lifetime at that prediction time to not be lower than this difference; when Greater than or equal to hour, A value of 0 is used to ensure that the lower bound is non-negative. By introducing the above lower bound constraint or penalty treatment to the censored samples, the stability and reliability of the calibration relationship fitting and the calibration of the lifetime prediction model can be enhanced by utilizing the lifetime lower bound information provided by the censored samples, even when the actual failure point has not been observed.

[0070] In this embodiment, combined with practical engineering applications, a dynamic life calibration process for an adjustable snap-action thermostat is provided. This process completes online life assessment and calibration without disassembling the thermostat, and supports cloud aggregation of data of the same model and OTA model refeedback, achieving the effect of becoming more accurate with each calibration.

[0071] S1 Online Accelerated Life Test Setup and Multi-Source Synchronous Acquisition: The jump-type temperature controller under test is installed in the test fixture and placed in a stepper servo temperature control box. The temperature controller is driven to cyclically jump at a preset rate through temperature cycling. At the same time, an electrical load circuit is set up to apply an adjustable load current, and a data acquisition module is configured to synchronously acquire contact voltage signals, contact current signals, and contact jump displacement signals.

[0072] To ensure consistency in subsequent feature extraction, voltage / current / displacement signals are timestamped using the same clock source, or time alignment is performed on multi-channel signals. During acquisition, a single jump action is used as the basic event unit. Acquisition windows are set before and after each jump event, such as a preset duration before the event plus a preset duration after the event, continuously recording to form a time-series data stream of multi-source parameter signals. The platform or host computer displays the current number of jumps, current load current and temperature cycle parameters, as well as voltage / current / displacement waveforms in real time for monitoring and tracing the experimental process.

[0073] S2 Single Jump Feature Instantiation: For each jump corresponding to the acquisition window, a three-dimensional degradation feature vector is extracted from the electrical and displacement parameters to characterize the degradation state. Specifically, this may include: arcing time, during the transient period when the jump causes the contact to open / close, the arc duration is determined based on the voltage / current waveform, such as the time period when the current is continuously non-zero and the voltage shows arc voltage characteristics, and the arcing duration is calculated accordingly; fused energy, the instantaneous power within the arc duration is integrated to obtain the energy accumulation characterization of a single jump, such as integrating the product of voltage and current over time to obtain the energy value, used to characterize the cumulative thermal-electric shock of the arc on the contact material; and reed rebound speed, the rebound speed characterization is calculated based on the displacement signal in the rebound segment after the jump, such as differentiating the displacement signal and taking the peak / average speed of the rebound segment, used to characterize the degradation changes in the elasticity and motion stability of the mechanism.

[0074] The three representations are combined in a predetermined dimensional order to form a three-dimensional degenerate feature vector. The feature vectors obtained from each jump are then arranged in order of the number of jumps to form a feature vector sequence. If necessary, normalization / standardization processing can be performed on the feature vectors to reduce the impact of dimensional differences on the model input.

[0075] S3 is based on lightweight LSTM online inference: It uses a lightweight LSTM network as the lifetime prediction model to perform remaining lifetime prediction on the above feature vector sequence. The lightweight LSTM may include an input layer, at least one LSTM temporal coding layer, and a lifetime regression output layer; the input is a continuous subsequence of length L, which can be extracted from the feature vector sequence by a sliding window, and the output is the predicted number of remaining hops at the prediction time.

[0076] To achieve online, multi-stage prediction, the sliding window is continuously updated as the jump cycle progresses. After each or several jump cycles, the latest feature vector is introduced and the oldest feature vector is discarded. The lightweight LSTM is repeatedly called to output new predicted values ​​for the remaining number of jumps, thus obtaining a time-series sequence of remaining lifetime predictions. The platform can further calculate characteristics of lifetime evolution trends from the prediction sequence, such as the rate of decline, the degree of fluctuation, and inflection point indicators, which serve as the basis for subsequent stress adaptive updates.

[0077] S4 threshold-triggered adaptive acceleration: The predicted lifetime threshold is used as the control trigger condition: When the predicted number of remaining jumps or its trend characterization meets the control threshold, such as when the predicted remaining lifetime drops to a certain proportion of the nominal lifetime, or when the predicted sequence shows a continuous monotonous decrease and approaches the preset range, the platform automatically issues a stress update command to execute accelerated aging.

[0078] The main method for accelerating aging is to automatically increase the load current: for example, by increasing the load current by a preset step size ΔI, or by amplifying it to the next current level by a preset factor; at the same time, the temperature cycling and jump timing can be kept constant, or parameters such as temperature swing amplitude / holding time / cycle frequency can be adjusted in conjunction with the test conditions to form adjustable accelerated stress. Each stress update is recorded as part of the stress parameter sequence, including the update time, the current values ​​before and after the update, the predicted lifetime and degradation characteristics at that time, etc., for subsequent calibration backtracking and model training.

[0079] As the experiment continues, the platform continuously monitors whether the multi-source degradation characterization quantities meet the failure judgment conditions, such as abnormal action of electrical and displacement parameters reaching the threshold, abnormal increase and persistence of arcing time or energy accumulation, and instability of displacement action amplitude / rebound characteristics. When the failure judgment conditions are met, the failure point is determined, and the key threshold parameters that lead to failure are recorded, such as the number of jumps when the failure occurs, the load current level at that time, the temperature cycling stage, and the threshold of key degradation characterization quantities, thereby completing the dynamic calibration during the experiment.

[0080] S5 Failure Point Backfilling and Cloud Aggregation: To address the challenge of uniformly calibrating the discrepancy between predicted and actual lifetimes as degradation stages change, calibration backfilling and model iteration are performed after failure points are determined: 1) Failure point backfilling of actual remaining lifetime: Using the number of jumps corresponding to the failure point as a reference, the actual remaining lifetime value is backfilled for each prediction time. The actual remaining jump count can be determined by subtracting the number of jumps completed at the prediction time from the number of failure jumps, forming a pair of predicted and actual values; 2) Constructing a calibration sample set and fitting calibration relationships: A calibration sample set is formed by combining feature vectors, predicted remaining lifetime values, and actual remaining lifetime values. The predicted remaining lifetime value is used as the independent variable, and the actual value as the dependent variable. A mapping function is obtained through regression fitting, serving as the calibration relationship. If necessary, the degradation process can be divided into at least two stages based on the changes in the feature vector sequence or prediction sequence, and a segmented calibration relationship can be formed by fitting calibration sub-relationships to improve performance at different degradation stages. 3) Calibrate the lifetime prediction model: You can choose to perform calibration mapping on the model output to obtain the calibrated predicted lifetime, or extract the model calibration parameters from the calibration relationship to update the LSTM output layer parameters / feature weights, so that the subsequent prediction is closer to the real lifetime evolution; 4) Aggregate the same model data in the cloud and update the general model: Upload the feature sequence, stress parameter sequence, failure point information, prediction-real pairing data, calibration relationship parameters, etc. of this test to the cloud; The cloud aggregates multiple sample data by model / batch, and performs incremental training or retraining on the general lifetime model and stress update strategy model to obtain more robust general parameters; 5) OTA backfeed: The updated general lifetime model parameters, calibration parameters or strategy parameters in the cloud are distributed to the test end or production line end in OTA, so that the subsequent lifetime calibration and online evaluation can obtain more stable and accurate prediction results under the same test conditions, achieving the effect of becoming more accurate with calibration.

[0081] Finally, the platform outputs lifetime calibration results, which may include: calibrated lifetime assessment values ​​(expressed as remaining number of jumps or equivalent lifetime), calibration relationship parameters / stage boundaries corresponding to the assessment, and failure point and threshold parameter records associated with the test task, which serve as a quantitative decision-making basis for production line selection, reliability assessment, and predictive maintenance.

[0082] The above description is merely an explanation of preferred embodiments of this application and the technical principles upon which they are based, and is not intended to limit this application. Those skilled in the art should understand that the scope of protection of this application is not limited to technical solutions formed by specific combinations of the above-mentioned technical features, but should also cover other technical solutions formed by arbitrary combinations, substitutions, or modifications of the above-mentioned technical features and their equivalent features without departing from the inventive concept of this application. For example, solutions obtained by substituting the above-mentioned features with technical features disclosed in this application that have the same or similar functions.

[0083] It should also be understood that the step numbers in the invention content and embodiments are for illustrative purposes only and do not necessarily limit the execution order. The order of each process should be based on its functional implementation and internal logical relationship. Based on the teachings of this application, those skilled in the art can make various modifications, variations or equivalent substitutions to the implementation methods without departing from the spirit and substance of this application, and all such modifications or substitutions should fall within the protection scope of this application.

Claims

1. A method for dynamic life calibration of an adjustable snap-action temperature controller, characterized in that, The method includes: applying adjustable acceleration stress to the thermostat and driving it to cycle and jump, while simultaneously acquiring multi-source parameter signals reflecting the degradation state of the thermostat; segmenting the multi-source parameter signals according to the jump cycle and degradation stage, extracting multi-source degradation characterization quantities, and constructing a degradation feature vector sequence, wherein the degradation feature vector sequence is a time-series arrangement sequence of degradation feature vectors obtained by vectorizing and combining multi-source degradation characterization quantities; performing multiple remaining lifetime predictions based on the degradation feature vector sequence using a lifetime prediction model to obtain a remaining lifetime prediction sequence, and generating lifetime trend features to characterize the lifetime evolution trend; adaptively updating the acceleration stress parameters of the next stage according to the lifetime trend features until the multi-source degradation characterization quantities meet the failure judgment conditions, determining the failure point and backfilling the true remaining lifetime values ​​at each prediction time, establishing a calibration sample set composed of degradation feature vectors, predicted remaining lifetime values, and true remaining lifetime values, and fitting a calibration relationship to map the predicted remaining lifetime values ​​to the true remaining lifetime values; calibrating the lifetime prediction model based on the calibration relationship, and outputting the lifetime calibration results of the thermostat from the calibrated model.

2. The dynamic life calibration method for the adjustable snap-action temperature controller according to claim 1, characterized in that, The multi-source parameter signals include at least electrical parameter signals and displacement parameter signals. The electrical parameter signals include voltage signals and current signals, and the displacement parameter signals include the displacement signals of contact jumps. Segmenting the multi-source parameter signals according to the jump cycle and degradation stage includes: using the acquisition window corresponding to a single jump action as a jump cycle segmentation unit; performing time alignment and window truncation on the electrical parameter signals and displacement parameter signals; and dividing multiple jump cycles into at least two degradation stages based on the changing trend of degradation characterization quantities within adjacent jump cycles; extracting electrical characteristic quantities from the electrical parameter signals to characterize the electrical degradation state and extracting displacement characteristic quantities from the displacement parameter signals to characterize the mechanism degradation state within each segmentation unit; and using the electrical characteristic quantities and displacement characteristic quantities to constitute multi-source degradation characterization quantities.

3. The dynamic life calibration method for the adjustable snap-action temperature controller according to claim 1, characterized in that, The process of performing multiple remaining lifetime predictions based on the degradation feature vector sequence by the lifetime prediction model includes: extracting continuous subsequences from the degradation feature vector sequence as model input based on a preset sliding window, and outputting the remaining lifetime prediction value corresponding to the subsequence at each prediction time by the lifetime prediction model; updating the sliding window in a loop with sudden jumps and repeating the prediction to obtain a remaining lifetime prediction sequence arranged in time sequence.

4. The dynamic life calibration method for the adjustable snap-action temperature controller according to claim 3, characterized in that, The lifespan trend characteristics include the declining slope, volatility, moving average difference, inflection point indicator, and number of consecutive monotonically declining values ​​of the remaining lifespan prediction sequence; the volatility is the variance or absolute deviation of the remaining lifespan prediction sequence within the sliding window.

5. The dynamic life calibration method for an adjustable snap-action temperature controller according to claim 1, characterized in that, The adaptive update of the accelerated stress parameters for the next stage based on lifetime trend characteristics includes: constructing an accelerated stress strategy set based on historical lifetime test data, wherein the historical lifetime test data includes at least the accelerated stress parameter sequences and their lifetime response data corresponding to different jump cycles or degradation stages; the lifetime response data includes at least: the number of jumps corresponding to the failure point, the predicted remaining lifetime value at each prediction time, and the corresponding actual remaining lifetime value; using the degradation feature vector sequence, the remaining lifetime prediction sequence, and the lifetime trend characteristics as strategy inputs, and the strategy evaluation target as the training target, a strategy generation model for outputting the accelerated stress parameters for the next stage is obtained through training; and the accelerated stress parameters for the next stage are output by calling the strategy generation model based on the lifetime trend characteristics.

6. The dynamic life calibration method for an adjustable snap-action temperature controller according to claim 5, characterized in that, The strategy evaluation objective is determined by any one of the following or a weighted combination of multiple: a lifetime calibration efficiency index, used to characterize the efficiency of acquiring lifetime calibration information, and including at least one of the following: the number of jumps or time required to reach the failure determination condition, the number of jumps or time required to reach the preset remaining lifetime interval, and the convergence rate of the remaining lifetime prediction error or calibration error; a mechanism consistency constraint index, used to characterize the stability of the degradation process, and including at least one of the following: the constraint on the variation amplitude of multi-source degradation characterization quantities between adjacent jump cycles, the constraint on the number of abnormal event triggers, and the constraint on the amplitude of accelerated stress updates; a prediction consistency index, used to characterize the stability of multiple remaining lifetime predictions, and including at least one of the following: the volatility of the remaining lifetime prediction sequence within a preset window, the degree of monotonicity retention of the prediction results, and the variation amplitude of the prediction deviation before and after calibration; The adaptive update of the accelerated stress parameters for the next stage also includes: calculating the consistency risk based on the mechanism consistency constraint index or the predicted consistency index; when the consistency risk exceeds the risk threshold, performing at least one of the following operations: load reduction rollback, switching to a lower stress strategy, or pausing the jump cycle.

7. The dynamic life calibration method for an adjustable snap-action temperature controller according to claim 1, characterized in that, The acceleration stress parameters include load current parameters, and at least one of temperature cycle parameters and jump cycle beat parameters; the temperature cycle parameters include at least one of temperature swing, temperature rise / fall slope, and hold time; the jump cycle beat parameters include at least one of cycle frequency or duty cycle.

8. The dynamic life calibration method for an adjustable snap-action temperature controller according to claim 1, characterized in that, The process of fitting the calibration relationship includes: based on the calibration sample set, using the predicted remaining lifetime as the independent variable and the actual remaining lifetime as the dependent variable, a mapping function is obtained by regression fitting to map the predicted remaining lifetime to the actual remaining lifetime; the mapping function is any one of a linear function, a piecewise linear function, or a nonlinear function, and a minimum constraint is applied to the mapping error during the fitting process.

9. The dynamic life calibration method for an adjustable snap-action temperature controller according to claim 1, characterized in that, The calibration relationship includes: dividing the lifetime degradation process into at least two degradation stages based on the change points of the degradation feature vector sequence or the remaining lifetime prediction sequence, and fitting corresponding calibration sub-relationships for different degradation stages to form a segmented lifetime calibration relationship; calibrating the lifetime prediction model based on the calibration relationship includes one of the following operations: performing output calibration processing on the remaining lifetime prediction value to obtain the calibrated remaining lifetime prediction value; determining the model calibration parameters based on the calibration relationship, and updating the model parameters of the lifetime prediction model using the model calibration parameters, or updating the weight parameters of the multi-source degradation characterization quantity in the lifetime prediction model.

10. The dynamic life calibration method for an adjustable snap-action temperature controller according to claim 1, characterized in that, The method further includes: when the temperature controller reaches a preset maximum number of jumps... If the failure criteria are not met, the corresponding sample of the temperature controller is identified as a censored sample, and the lower bound of the true remaining lifetime value at each prediction time is determined based on the preset maximum number of jumps. When fitting the calibration relationship or calibrating the lifetime prediction model, a truncation constraint not lower than the lower bound is applied to the predicted remaining lifetime value that is lower than the true remaining lifetime value, or a penalty loss related to the degree of difference between the predicted remaining lifetime value and the lower bound is applied. For any prediction time... Corresponding number of completed jumps Lower bound of the true remaining lifetime 。

Images