Dynamic prediction and adaptive optimization closed-loop method and system for fatigue life of main pull spring of molded case circuit breaker

Abstract

The application discloses a kind of main spring fatigue life dynamic prediction and adaptive optimization closed-loop method of moulded case circuit breaker.It is constructed with online bayesian calibration ability of digital twin, real-time data is collected using multi-source sensor and operation mode is identified;Degeneration sensitive features are dynamically screened from data, life prediction is carried out through physical channel and multi-head attention-based data channel, and dynamic weighted fusion is carried out after consistency verification;The fusion life is used as a reward signal to drive the deep reinforcement learning agent to interact with the digital twin, autonomously search for the optimal structural parameters and virtually verify.The application overcomes the risk of physical / data model mismatch, and opens a real-time closed loop between prediction and optimization, achieving high-robustness main spring fatigue life management and design iteration under multiple working conditions.

Description

Technical Field

[0001] This invention belongs to the field of low-voltage electrical appliance reliability assessment and intelligent operation and maintenance technology, specifically involving a dynamic prediction and adaptive optimization closed-loop method and system for the fatigue life of the main tension spring of a molded case circuit breaker. Background Technology

[0002] Molded case circuit breakers (MCCBs) are core protection devices in low-voltage power distribution systems. The main tension spring of its operating mechanism bears alternating loads during opening and closing, making it a critical weak point that determines the circuit breaker's mechanical lifespan. Fatigue fracture of the main tension spring will directly lead to the circuit breaker malfunctioning or failing to operate, causing power outages or even safety accidents. Statistics show that energy storage spring problems account for over 40% of failures in spring-operated mechanisms. Furthermore, due to the dispersion of factors such as material properties, manufacturing tolerances, and operating conditions, the actual fatigue life of the main tension spring exhibits significant individual differences and randomness, making scheduled maintenance at fixed intervals neither economical nor reliable.

[0003] In existing technologies, there are three main approaches to fatigue life management of main tension springs, but none of them can form an effective closed-loop adaptive management capability: (1) A method combining offline simulation and periodic testing The stress at critical points is obtained through finite element analysis, and the theoretical life is calculated by combining the material's SN curve and Miner's cumulative damage criterion. For example, Li Kui et al. established a rigid-flexible coupling dynamic model of the main tension spring of a molded case circuit breaker and analyzed the distribution characteristics of fatigue failure. This type of method belongs to open-loop static evaluation. Once the design is finalized, it does not change and cannot reflect the actual load spectrum drift and material performance degradation caused by wear, lubrication degradation, and changes in ambient temperature during operation. When the actual operating conditions deviate from the design assumptions, the prediction error increases sharply.

[0004] (2) Online monitoring method for a single physical quantity By installing pressure or displacement sensors at the ends of the springs, the spring force or deformation can be monitored, and an alarm can be triggered when it exceeds a set threshold. For example, Zhang Xu et al. designed a pressure monitoring system for the spring operating mechanism of a circuit breaker. However, the slow decay of spring force can be caused by fatigue damage or non-damaging factors such as stress relaxation. A single steady-state physical quantity cannot distinguish the cumulative damage process, and a warning can only be issued when failure is imminent, lacking the ability to quantitatively predict the remaining life.

[0005] (3) Purely data-driven prediction methods Deep learning models are trained using signals such as vibration and acoustic emission, and then directly mapped to the remaining life. This method has achieved some success in fields such as rolling bearings, but the full life test cycle of the main tension spring of molded case circuit breakers is long and extremely costly, and the real fault samples that can be obtained are extremely scarce; moreover, when the operating frequency, ambient temperature and other operating conditions change, the generalization ability of the trained model drops sharply, and false alarms are very likely to occur.

[0006] A deeper technical contradiction lies in the fact that while physical models possess interpretability and extrapolation capabilities, they struggle to adapt to time-varying operating conditions and individual dispersion; data-driven models can automatically capture implicit degradation patterns, but they heavily rely on data and lack physical consistency constraints. Existing technologies simply juxtapose the two, performing only a weighted average, failing to identify model mismatches or utilize real-time data to correct physical models. This results in predictions that are either redundant under stable operating conditions or distorted when operating conditions change.

[0007] Furthermore, when the predicted lifetime reveals design deficiencies, existing optimization methods (such as genetic algorithms and response surface methodology) are all executed offline in an open-loop manner, requiring manual remodeling and analysis. This creates significant time delays and information gaps between prediction and optimization. Although the concept of digital twins has been introduced into many fields, it has mostly remained at the stage of 3D visualization and data dashboards, and has not yet formed an intelligent closed loop of "perception-cognition-decision-evolution". Summary of the Invention

[0008] This invention aims to solve the following core technical problems in the fatigue life management of main tension springs in molded case circuit breakers in the prior art: When physical models and data-driven models lack dynamic verification and bidirectional calibration, they cannot maintain predictive robustness when operating conditions change. Digital twin models are static and fixed, and cannot be updated synchronously with the operation of equipment; individual differences and degradation evolution are ignored. Fatigue life prediction and structural optimization design are disconnected from each other, and the optimization feedback has a long cycle delay, making it impossible to form an autonomous iterative closed loop. Lacking the ability to adapt to various operating conditions such as frequent operation, long-term stagnation, and high temperature, a single model is difficult to cover all operating scenarios.

[0009] To achieve the above objectives, the present invention is implemented through the following technical solution: A dynamic prediction and adaptive optimization closed-loop method for fatigue life of main tension springs in molded case circuit breakers, characterized by comprising: A digital twin with online Bayesian calibration capability is constructed, whose material properties and boundary condition parameters are described by probability distribution. The main tension spring operation data is collected synchronously through multi-source sensors, and the current operation mode is identified. Degradation-sensitive features are dynamically filtered from the data, and the remaining lifetime and prediction confidence are obtained through the physical mechanism channel and the data-driven channel, respectively. Dynamic weighted fusion is performed based on the consistency test of the two channel predictions. When the deviation exceeds the threshold, model anomaly diagnosis is initiated to output the remaining lifetime of the fusion. The remaining lifetime of the fusion is input into a structure optimization module based on deep reinforcement learning. This module uses the lifetime improvement as a reward, learns and generates an optimization parameter scheme through interaction with the digital twin, and updates the model after virtual verification in the digital twin, forming a closed loop.

[0010] The present invention further provides that the online Bayesian calibration of the digital twin is specifically as follows: when the deviation between the measured spring mechanical behavior and the digital twin simulation exceeds a preset threshold, parameter inversion based on Markov chain Monte Carlo sampling is automatically triggered to perform posterior estimation of material fatigue performance parameters, contact stiffness, damping parameters, etc., and update the digital twin accordingly to keep it synchronized with the physical individual state.

[0011] The present invention further provides that the physical mechanism channel uses measured strain as finite element boundary conditions to calculate the stress spectrum of the critical point in real time, obtains the predicted physical life value through rainflow counting and cumulative damage criteria, and obtains the predicted distribution and confidence level through Monte Carlo propagation using the parameter probability distribution.

[0012] The present invention further specifies that the data-driven channel adopts a multi-head attention temporal neural network, which takes the selected sensitive feature subset sequence as input and outputs lifetime prediction value and confidence level; the network is pre-trained by generating degradation data through digital twin simulation, and performs incremental online learning in conjunction with the elastic weight consolidation method in actual operation.

[0013] The present invention further specifies that the dynamic weighted fusion specifically involves: calculating the relative deviation between the predicted lifetime values ​​of the physical channel and the data channel; if the deviation is less than a dynamic threshold, weighting is performed according to the confidence ratio; if the deviation exceeds the threshold, the channel with lower confidence is frozen, and an anomaly diagnosis process based on sensor effectiveness verification and data distribution drift analysis is initiated to determine whether to recalibrate the digital twin or retrain the data channel.

[0014] The present invention further specifies that the deep reinforcement learning structure optimization module models the structural parameter optimization as a Markov decision process, the state space includes the current structural parameters and corresponding lifetimes, the action space is the parameter adjustment amount, and the reward function is a linear combination of (predicted lifetime after optimization - predicted lifetime before optimization) / predicted lifetime before optimization and constraint penalty term; the agent is trained through a proximal policy optimization algorithm to output the optimal parameter combination that satisfies stiffness, geometry and manufacturing constraints.

[0015] The present invention further provides that the operating modes include at least a frequent operating mode, a long-term static mode, and a high-temperature mode; based on the identified mode, the cumulative damage calculation model of the physical channel is adaptively switched, and the dual-channel fusion weight is adjusted.

[0016] The present invention further provides that, in the long-term static mode, the physical channel is switched to a life prediction model based on stress relaxation and creep damage accumulation, and damage is assessed by monitoring the spring tension decay curve over time; in the high-temperature mode, the material fatigue performance parameters in the physical channel are automatically introduced with a temperature correction factor based on the Arrhenius equation.

[0017] The present invention further provides that the dynamic screening of the degradation-sensitive features is based on the weighted comprehensive score of monotonicity index and trend index, and the screening threshold and feature combination are dynamically adjusted according to the current operation mode.

[0018] The present invention also provides a fatigue life management system for the main tension spring of a molded case circuit breaker, including a multi-source sensor acquisition unit, an edge computing unit and a cloud platform. The system is characterized in that it executes the method described above to realize intelligent operation and maintenance of the entire process from state perception, life prediction, anomaly diagnosis to autonomous optimization closed loop.

[0019] The beneficial effects of this invention are: The physics-data dual-driven verification fusion mechanism proactively identifies single model mismatches, avoiding catastrophic misjudgments during operating condition drift. The mean absolute percentage error of remaining life prediction is significantly lower than that of pure physics and pure data models in individualized verification. Online Bayesian calibration based on MCMC allows model parameters to degrade along with the entity, solving the fundamental problem of model distortion after long-term operation and ensuring the effectiveness of long-term predictions. A real-time closed loop between prediction and optimization is established: fatigue life prediction is directly used as a reward for reinforcement learning, enabling autonomous optimization design iteration without manual intervention. The cycle from triggering optimization to generating a feasible solution is shortened from several weeks to hours. A multi-mode adaptive mechanism allows a single system to cover complex operating scenarios such as frequent operation, long-term standby, and high temperatures, greatly expanding its applicability. Significant economic and safety benefits are achieved: it drives a fundamental shift in molded case circuit breakers from "planned maintenance" to "predictive maintenance," effectively avoiding unplanned downtime and catastrophic failures. Simultaneously, structural optimization extends the lifespan of key components, reducing the total lifespan cost. Attached Figure Description

[0020] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0021] The following will describe in detail the implementation of this application with reference to the accompanying drawings and embodiments, so that the implementation process of how this application uses technical means to solve technical problems and achieve technical effects can be fully understood and implemented accordingly.

[0022] Example 1: As Figure 1 As shown in the figure, this embodiment provides a dynamic prediction and adaptive optimization closed-loop method for the fatigue life of the main tension spring of a molded case circuit breaker. The specific steps are as follows: Step S1: Construct a digital twin with online Bayesian calibration capabilities A high-fidelity digital twin was established, comprising a three-dimensional parametric geometric model of the main tension spring, a nonlinear finite element model, a probabilistic material model, and a fatigue cumulative damage model. The key is that the material fatigue performance parameters (such as fatigue strength coefficient σ'f, fatigue ductility coefficient ε'f, etc.) and boundary condition parameters (such as equivalent damping of the operating mechanism, contact friction coefficient, etc.) in the digital twin are described in probabilistic distribution form, and a Bayesian update engine based on Markov chain Monte Carlo (MCMC) is embedded. When the difference between the measured spring force-displacement curve and the simulated curve in the digital twin exceeds a preset threshold, the system automatically triggers parameter inversion, using MCMC sampling to obtain the posterior distribution of the parameters, updating the digital twin, ensuring that the virtual model is always a high-fidelity mirror of the current physical state.

[0023] Step S2: Synchronous data acquisition and online identification of operating modes from multi-source heterogeneous sensors Miniature strain gauges, triaxial MEMS accelerometers, acoustic emission sensors, laser displacement sensors, and temperature sensors are arranged in the operating mechanism to synchronously acquire the complete waveforms of each opening and closing action using a unified clock. Simultaneously, a lightweight operating mode classifier is constructed, using sliding time windows to statistically analyze operating frequency, average temperature, and temperature gradient as inputs to identify online whether the current mode is a frequent operating mode, a long-term static mode, or a high-temperature mode.

[0024] Step S3: Multi-domain degradation feature extraction and dynamic selection of sensitive feature subsets. The steps are as follows: After wavelet denoising and normalization of the multi-source signals, time-domain, frequency-domain, and time-frequency-domain features are extracted, and the monotonicity and trend indices of each feature are calculated. The feature subset most sensitive to the current degradation mode is automatically selected. This subset will be dynamically adjusted according to changes in the operating mode.

[0025] Step S4: The fusion steps of physical-data dual-driven remaining lifetime prediction and dynamic verification are as follows: Physical channel: The measured strain history is used as the displacement boundary condition of the finite element model. The stress spectrum of the critical point is calculated in real time. By rainflow counting and Miner's linear / nonlinear cumulative damage criterion considering the mean stress correction, the remaining physical channel lifetime N-phy and the confidence level C-phy obtained from the parameter uncertainty are obtained.

[0026] Data channel: The selected sensitive feature sequences are input into a pre-trained multi-head attention temporal neural network. The network simultaneously outputs the remaining lifetime N-data and the standard deviation of its predicted distribution, thus obtaining the confidence score C-data.

[0027] Dynamic verification fusion: The relative deviation between the lifetime predictions of the two channels is calculated. If the deviation is less than a dynamic threshold, an adaptive weighted fusion based on confidence is used to obtain the final lifetime N-fusion; if the deviation exceeds the threshold, the model anomaly diagnosis module is activated—checking sensor effectiveness, assessing the distribution shift of the current load spectrum compared to the historical training set, and selecting to freeze a channel or re-trigger the parameter calibration of the digital twin based on the diagnosis results. This mechanism makes the fusion itself a "stethoscope" for model health.

[0028] Step S5: Dynamic classification and early warning of fatigue degradation state levels driven by lifespan Based on the design life of the machine, five health levels are defined according to the proportion of N-fusion, and different warning actions are triggered for each level.

[0029] Step S6: Deep reinforcement learning closed loop of prediction-optimization-verification When N-fusion drops below a preset percentage of the design life (e.g., 80%), the structural optimization process is automatically activated: Using a digital twin as a high-fidelity environment, a Kriging proxy model is constructed to represent the key structural parameters of the main tension spring (spring wire diameter, mean diameter, number of coils, hook transition radius, etc.) and fatigue life.

[0030] The optimization problem is formalized as a Markov decision process: the state is the current parameter vector and the corresponding fatigue life, the action is the parameter adjustment amount, and the reward function is designed as (optimized life - unoptimized life) / unoptimized life - constraint penalty term.

[0031] The agent is trained using the Proximal Policy Optimization (PPO) algorithm, which directly searches for the optimal parameter combination based on the lifetime improvement. Since the reward signal comes directly from the lifetime prediction of the digital twin, and the digital twin has undergone Bayesian calibration, the optimization direction has physical realism.

[0032] The optimal solution generated by the intelligent agent is first verified through high-fidelity simulation in the digital twin. After successful verification, the design parameters of the digital twin are updated, and the solution is stored in the knowledge base to guide future optimization, thus forming a closed loop.

[0033] This embodiment uses the main tension spring of the operating mechanism of a certain type of molded case circuit breaker (rated current 250A) as the object to construct and run the complete closed-loop system of the present invention.

[0034] First, the digital twin was initially constructed and probabilistically modeled. The initial design parameters for the main tension spring were: material SWP-B piano wire, wire diameter d = 2.5 mm, mean diameter D = 15 mm, effective number of coils 12, free length 60 mm, and hook inner corner radius R = 1.5 mm. A parametric three-dimensional nonlinear finite element model was established in Abaqus, using C3D10 quadratic tetrahedral elements, with the mesh at the hook root refined to 0.15 mm. The material constitutive model adopted a kinematic hardening model, with an elastic modulus E = 206 GPa and Poisson's ratio 0.3.

[0035] The probabilistic material model is constructed based on standard fatigue test data of the same batch of steel wires, and the posterior distribution of the Manson-Coffin formula parameters is inferred using the Bayesian method. It is assumed that the fatigue strength coefficient σ'f follows a normal distribution, and its mean μ-σ'f = 2015 MPa and standard deviation 60 MPa are obtained through MCMC sampling; the fatigue ductility coefficient ε'f follows a log-normal distribution, with a geometric mean of 0.58 and a geometric standard deviation of 0.12. The digital twin encapsulates these distributions as parameterized objects.

[0036] Two miniature strain gauges (sensitivity coefficient 2.08) are attached to the base of the hooks at both ends of the main tension spring. A triaxial accelerometer (ADXL356B) is mounted on the side plate of the operating mechanism. An acoustic emission sensor (Nano30) is mounted on the frame using a magnetic clamp. A laser displacement sensor (KEYENCE IL-030) is used to monitor the axial displacement at one end of the spring, and a thermocouple is attached to the spring surface. All sensors are connected to a synchronous acquisition unit, which samples acoustic emission at 25.6kHz and strain and acceleration at 2.56kHz. Each complete opening and closing of the circuit breaker triggers a complete waveform recording and timestamp.

[0037] Define a window length W = 720 hours (30 days). Count the number of operations (Nop) within the window: Nop > 1500 is defined as a frequent operation mode; 150 ≥ Nop is defined as a long-term inactive mode; the rest are normal modes. Additionally, if the average temperature within the window exceeds 55℃, a high-temperature flag is added. Mode classification is implemented using a shallow fully connected network, and the current mode code is output.

[0038] After each action, the peak strain ε-max, strain range Δε, and root mean square value are extracted from the denoised strain signal; the centroid frequency and energy proportion of the 1-3kHz frequency band are extracted from the vibration signal; and the ringing count rate and cumulative energy are calculated from the acoustic emission signal. These constitute a high-dimensional feature vector. For each feature sequence, the monotonicity index Mon (mean of sign difference quotient) and Spearman trend coefficient ρ are calculated. With α=0.4 and β=0.6, a comprehensive score is calculated, and a threshold of 0.7 is set to filter features with scores >0.7. In frequent operation mode, the vibration frequency band energy and acoustic emission count rate scores are prominent; in long-term static mode, the monotonicity score of the slowly drifting peak strain re-enters the sensitive set, ensuring adaptive updating of the feature set.

[0039] The physical channel employs a reduced-order beam element model for accelerated computation. Measured strain data is converted into displacement boundary conditions, and a stress analysis is run once every 100 operations, using rainflow counting to obtain the stress amplitude-mean matrix. Goodman correction and the Miner cumulative damage criterion are used to calculate the cumulative damage D in real time. Considering material parameter uncertainties, 100 sets of parameters are sampled from the posterior distribution for each calculation, resulting in 100 remaining lifetime samples. The mean is taken as N-phy, and the ratio of the standard deviation to the mean is used as the uncertainty, with the confidence level C-phy = 1 - (σ-N / μ-N). As the digital twin is calibrated online, C-phy gradually approaches 1.

[0040] Network Structure: The input dimension is the current number of sensitive features (dynamically changing, handled by padding masks), which is then positionally encoded and fed into a 4-layer Transformer encoder with 8 attention heads per layer. The feedforward network has a dimension of 1024. After global average pooling, two output heads are generated: a lifetime prediction head (linear) and a confidence head (Sigmoid). Pre-training data consists of 1.2 million degradation simulation sequences generated by a digital twin under different material and load combinations. A cosine annealing learning rate is used, and the loss function is a weighted combination of MSE and confidence-calibrated binary cross-entropy. The model achieves a MAPE of 6.8% on the simulation data.

[0041] Assume the physical channel outputs N-phy = 18500 times with a confidence level of 0.82; the data channel outputs N-data = 21000 times with a confidence level of 0.75. The relative deviation |21000-18500| / 21000 ≈ 11.9%, which is less than the set threshold of 20%. Therefore, the fusion weights w-phy = 0.82 / (0.82+0.75) = 0.522, w-data = 0.478, and the fusion lifetime N-fusion = 0.522×18500+0.478×21000 ≈ 19630 times. If the confidence level of a data channel output suddenly drops to 0.4 and the deviation suddenly increases to 35%, the system will determine that there may be a loose sensor or network mismatch, immediately freeze the data channel, and only rely on the physical channel output, triggering sensor self-check and model recalibration requests.

[0042] The design life is N-design = 40,000 cycles. N-fusion = 19,630 cycles, accounting for 49% of the design life, entering "attention state". The system pushes an attention notification to the maintenance personnel's mobile phones via the IoT module, suggesting "maintenance during the next planned shutdown window". When the life is below 2,000 cycles, the on-site audible and visual alarm is activated, and the control loop can be automatically cut off through the communication interface, forcibly locking the operation.

[0043] Optimization was initiated when N-fusion decayed to below 32,000 cycles (80% of the design life). First, 200 parameter combinations were sampled using Latin hypercube sampling, and the corresponding fatigue life was obtained through digital twin simulation to construct a Kriging model, achieving an R² validation of 0.98. Then, a PPO agent was constructed: state vector = [d, D, n, R_hook, current life], and action space consisting of discrete combinations of parameter increment / decrement steps. The reward R = (Nfnew - Nfold) / Nfold - 10 * I'violate, where I'violate is the constraint violation flag (set to 1 if stiffness exceeds ±15% or spring index C < 4 or C > 12). PPO adopted an Actor-Critic architecture, with both Actor and Critic being fully connected networks of [256, 256], a learning rate of 3e-4, a pruning parameter of 0.2, and a total of 500 episodes of training. Ultimately, the intelligent agent learned to slightly increase the spring wire diameter to 2.65mm and significantly increase the hook radius to 2.8mm while maintaining stiffness constraints, and also fine-tuned the mean diameter and number of turns. In digital twin virtual verification, this solution improved the fatigue life to approximately 61,000 cycles, an increase of 52%, and the opening and closing speed fluctuations of the moving and stationary contacts were within 3%, meeting the functional requirements.

[0044] When the equipment reached 18,000 operations, the measured spring force was 4.7% lower than the digital twin's predicted value, exceeding the 2% threshold and triggering calibration. Using the spring force attenuation as the observed data, assuming that the material's elastic modulus E and the local plastic parameters of the hook transition region drifted, 10,000 samples were taken using MCMC (discarding the first 2,000 preheated samples). The posterior mean of E decreased from 206 GPa to 203.5 GPa, and the local yield strength of the hook decreased by approximately 3%. The calibrated parameters were updated to the digital twin, and simulation was performed again. The force deviation was reduced to 0.9%, completing synchronization.

[0045] Three typical operating condition simulation scenarios were constructed: frequent operation (100 operations per day), long-term static operation (once per month under continuous tension), and high-temperature frequent operation (ambient temperature 90℃). Results showed that in the long-term static scenario, if the data channel model under the frequent operation mode was forcibly adopted, it would fail to capture the force attenuation caused by stress relaxation. The remaining life prediction still showed over 80% at 20,000 hours, but in reality, the tensile force was insufficient due to relaxation. This invention automatically switches to the stress relaxation cumulative damage model, effectively predicting the functional failure threshold of the tensile force. In the high-temperature scenario, due to the automatic introduction of a temperature correction factor, the physical channel prediction error improved from -32% without correction to -7%.

[0046] As used in the specification and claims, certain terms refer to specific components. Those skilled in the art will understand that hardware manufacturers may use different names to refer to the same component. This specification and claims do not distinguish components based on differences in name, but rather on differences in function. The term "comprising" throughout the specification and claims is an open-ended term and should be interpreted as "comprising but not limited to." "Approximately" means that within an acceptable margin of error, those skilled in the art can solve the technical problem and substantially achieve the technical effect within a certain margin of error.

[0047] The foregoing description illustrates and describes several preferred embodiments of the present invention. However, as previously stated, it should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept described herein through the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

Claims

1. A method for dynamic prediction and adaptive optimization closed-loop method for fatigue life of main tension springs in molded case circuit breakers, characterized in that, include: A digital twin with online Bayesian calibration capability is constructed, whose material properties and boundary condition parameters are described by probability distribution. The main tension spring operation data is collected synchronously through multi-source sensors, and the current operation mode is identified. Degradation-sensitive features are dynamically filtered from the data, and the remaining lifetime and prediction confidence are obtained through the physical mechanism channel and the data-driven channel, respectively. Dynamic weighted fusion is performed based on the consistency test of the two channel predictions. When the deviation exceeds the threshold, model anomaly diagnosis is initiated to output the remaining lifetime of the fusion. The remaining lifetime of the fusion is input into a structure optimization module based on deep reinforcement learning. This module uses the lifetime improvement as a reward, learns and generates an optimization parameter scheme through interaction with the digital twin, and updates the model after virtual verification in the digital twin, forming a closed loop.

2. The method according to claim 1, characterized in that, The online Bayesian calibration of the digital twin specifically involves: when the deviation between the measured spring mechanical behavior and the digital twin simulation exceeds a preset threshold, parameter inversion based on Markov chain Monte Carlo sampling is automatically triggered to perform posterior estimation of material fatigue performance parameters, contact stiffness, damping parameters, etc., and update the digital twin accordingly to keep it synchronized with the physical individual state.

3. The method according to claim 1, characterized in that, The physical mechanism channel uses measured strain as finite element boundary conditions to calculate the stress spectrum of critical points in real time. It obtains the predicted physical lifetime value through rainflow counting and cumulative damage criteria, and obtains the predicted distribution and confidence level through Monte Carlo propagation using the parameter probability distribution.

4. The method according to claim 1, characterized in that, The data-driven channel employs a multi-head attention temporal neural network, taking a selected subset of sensitive features as input and outputting lifetime predictions and confidence levels. The network is pre-trained using degradation data generated through digital twin simulation and undergoes incremental online learning in conjunction with an elastic weight consolidation method during actual operation.

5. The method according to claim 1, characterized in that, The dynamic weighted fusion specifically involves: calculating the relative deviation between the predicted lifetime values ​​of the physical channel and the data channel; if the deviation is less than a dynamic threshold, weights are assigned according to the confidence ratio for weighted fusion. If the deviation exceeds the threshold, channels with lower confidence are frozen, and an anomaly diagnosis process based on sensor validity verification and data distribution drift analysis is initiated to determine whether to recalibrate the digital twin or retrain the data channels.

6. The method according to claim 1, characterized in that, The deep reinforcement learning structure optimization module models the structural parameter optimization as a Markov decision process. The state space includes the current structural parameters and their corresponding lifetimes, the action space is the parameter adjustment amount, and the reward function is a linear combination of (predicted lifetime after optimization - predicted lifetime before optimization) / predicted lifetime before optimization and constraint penalty term. The agent is trained through a proximal policy optimization algorithm to output the optimal parameter combination that satisfies stiffness, geometry, and manufacturing constraints.

7. The method according to claim 1, characterized in that, The operating modes include at least a frequent operation mode, a long-term static mode, and a high-temperature mode; based on the identified mode, the cumulative damage calculation model of the physical channel is adaptively switched, and the dual-channel fusion weight is adjusted.

8. The method according to claim 7, characterized in that, In the long-term static mode, the physical channel switches to a life prediction model based on stress relaxation and creep damage accumulation, and assesses damage by monitoring the spring tension decay curve over time; in the high-temperature mode, the material fatigue performance parameters in the physical channel are automatically adjusted by a temperature correction factor based on the Arrhenius equation.

9. The method according to claim 1, characterized in that, The dynamic screening of degradation-sensitive features is based on a weighted composite score of monotonicity and trend indicators, and the screening threshold and feature combination are dynamically adjusted according to the current operating mode.

10. A fatigue life management system for the main tension spring of a molded case circuit breaker, comprising a multi-source sensor acquisition unit, an edge computing unit, and a cloud platform, characterized in that, The system executes the method of any one of claims 1 to 9 to achieve intelligent operation and maintenance throughout the entire process, from state awareness, lifespan prediction, anomaly diagnosis to autonomous optimization closed loop.

Images