A motor adaptive failure prediction method based on digital twin fusion intelligent optimization

Abstract

This invention relates to the field of intelligent reliability engineering technology for electromechanical products, specifically an adaptive failure prediction method for motors based on digital twin-based intelligent optimization. Through physically-driven intelligent modeling, a physical information neural network is used to automatically learn time-varying coupling coefficients, achieving an organic fusion of data-driven approaches and physical constraints. Failure paths are dynamically reconstructed by using a particle swarm optimization algorithm to identify dominant failure modes in real time and dynamically adjust the weights of the failure propagation chain. A prediction-verification-correction closed loop is implemented, leveraging digital twins and Bayesian inference to achieve continuous self-evolution of the model and constantly improve prediction accuracy. Intelligent design of test schemes is employed, using a multi-objective genetic algorithm to optimize and accelerate test conditions, improving test efficiency and failure mode coverage. Probabilistic lifetime prediction outputs a remaining lifetime distribution with confidence intervals, quantifying prediction uncertainty.

Description

Technical Field

[0001] This invention relates to the field of intelligent reliability engineering technology for electromechanical products, specifically a motor adaptive failure prediction method based on digital twin fusion intelligent optimization. Background Technology

[0002] As the core power unit, the reliability of the motor directly affects the safe operation of the entire system. Existing technologies suffer from the following bottlenecks: multiphysics coupling models use fixed coupling coefficients, which cannot adapt to dynamic changes in motor operating conditions, requiring extensive offline experimental calibration and exhibiting poor model generalization ability; traditional methods pre-set fixed failure propagation paths, failing to consider the switching of dominant failure modes under different operating conditions and lacking a mechanism to dynamically adjust failure chain weights based on real-time operating data; existing methods output a single lifetime prediction value, failing to quantify prediction uncertainty and ignoring the cumulative effects of sensor noise, model errors, and environmental randomness; accelerated life test conditions are often set based on engineering experience, potentially leading to over-excitation of certain failure modes or omission of key failure mechanisms; once established, the prediction model is fixed and cannot be self-corrected using actual operating data, causing prediction errors to accumulate and amplify over time.

[0003] There is an urgent need in this field for an intelligent failure prediction method with self-learning, self-adaptation, and self-optimization capabilities. This method should be able to quickly extract physical laws from a small amount of experimental data, dynamically adjust model parameters and failure evolution paths based on real-time monitoring data, and then provide probabilistic lifetime predictions with confidence intervals. Furthermore, it should be able to automatically design efficient accelerated testing schemes to support the continuous evolution of the prediction model throughout its entire service life and continuously improve its prediction accuracy and reliability. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention proposes a motor adaptive failure prediction method based on digital twin fusion intelligent optimization. This method deeply integrates deep learning, intelligent optimization algorithms, and multiphysics coupling theory to construct a fully intelligent prediction framework that can continuously improve itself.

[0005] This invention is achieved using the following technical solution:

[0006] An adaptive failure prediction method for motors based on digital twin fusion intelligent optimization includes the following steps:

[0007] Step S1: Deploy a multimodal sensor array to collect multi-source data and perform preprocessing and feature extraction;

[0008] Step S2: Identify the failure mechanisms of the four weak interfaces inside the motor: magnet-housing, potting compound-winding / core, carbon brush-commutator, and sealing material, and establish the corresponding physical damage evolution equations and failure propagation networks.

[0009] Step S3: Optimize the accelerated life test scheme using a multi-objective genetic algorithm;

[0010] Step S4: Construct a physical information neural network and automatically learn the time-varying coupling coefficient between state variables and multiphysics fields by embedding a hybrid loss function of physical control equations, boundary conditions and causal constraints;

[0011] Step S5: Based on the particle swarm optimization algorithm, dynamically reconstruct the failure chain weights using real-time monitoring data to identify the dominant failure mode;

[0012] Step S6: Establish a three-level mapping model of material-component-system to map the interface damage state into observable performance indicators at the system level;

[0013] Step S7: Construct a digital twin model integrating PINN and PSO weights, and use Bayesian inference to update the model parameters online to achieve quantitative propagation of uncertainty and probabilistic lifetime prediction;

[0014] Step S8: By comparing the prediction results with the measured data in a closed loop, the model parameters are adaptively updated and the failure chain is reconstructed, forming a continuously self-evolving prediction-verification-correction mechanism.

[0015] Step S2: Physical modeling of coupled failure based on the identification of four weak links includes the following steps:

[0016] Step S2.1: Identify systemic weaknesses; weaknesses include:

[0017] The bonding interface between the magnet and the housing is prone to fatigue cracking of the bonding layer under thermal-vibration coupling stress, which can help monitor torque fluctuations. With vibration acceleration ;

[0018] The potting compound at the winding / core interface is prone to crack propagation due to thermal cycling and hydrolysis, which can be addressed by measuring insulation resistance. With efficiency Conduct monitoring;

[0019] The carbon brush-commutator contact interface suffers from a complex failure resulting from arc erosion, mechanical wear, and electrochemical corrosion, involving multi-field coupling across electro-thermal-mechanical-chemical fields. Spark voltage can be monitored. Contact resistance With electromagnetic interference intensity;

[0020] The sealing performance of the bonding interface of the sealing material deteriorates due to thermo-oxidative aging and stress relaxation; its condition can be monitored by internal humidity. and penetration depth based on impedance spectrum inversion To characterize;

[0021] Step S2.2: Based on the four identified weak links and their dominant failure mechanisms, establish a quantitative physical model for each interface to describe its damage evolution or performance degradation process;

[0022] Step S2.3: Establish the failure propagation network among the four weak links of the motor and reveal their interaction and cascade effect.

[0023] Step S3: Optimizing the accelerated life test scheme using a multi-objective genetic algorithm includes the following processes:

[0024] Based on the multiphysics field coupled failure mechanism modeling and failure chain dynamic reconstruction, a multi-objective optimization model for accelerated life testing oriented towards the failure of key interfaces of motors is constructed.

[0025] Using test temperature, mechanical load, electromagnetic excitation intensity, and environmental stress as decision variables, an accelerated stress feasible region is formed under equipment capacity and safety constraints. With minimum test time, maximum failure mode coverage, and minimum accelerated equivalent deviation as optimization objectives, the non-dominated sorting genetic algorithm NSGA-II is used to solve the problem and obtain the Pareto optimal solution set corresponding to multiple test conditions.

[0026] The Pareto solutions are screened and evaluated to predict the interface damage evolution and failure chain activation characteristics. The accelerated life test scheme with the best overall efficiency, coverage and equivalence is selected to provide high-quality test data support for model calibration and life prediction.

[0027] Step S4: Construct a physical information neural network and automatically learn the time-varying coupling coefficient between state variables and multiphysics fields by embedding a hybrid loss function of physical control equations, boundary conditions and causal constraints. This includes the following steps:

[0028] Step S4.1: Construct the input layer variables of the physical information neural network to characterize the multiphysics operating conditions of the system in the time dimension. The input vector is defined as:

[0029] ;

[0030] in, For temperature parameters, For vibration stress parameters, For current parameters, This refers to the ambient relative humidity parameter. For input voltage parameters, It is a time variable;

[0031] Step S4.2: Construct the output layer variables of the neural network to simultaneously characterize the system state evolution process and the coupling relationship between multiple physics fields. The output vector is defined as:

[0032] ;

[0033] in, , , , These represent damage variables, crack length, contact resistance, and penetration depth, respectively. , , The time-varying coupling coefficients to be learned are used to describe the dynamic interaction strength between different physical fields;

[0034] Step S4.3: Construct a hybrid fully connected neural network, the network structure of which is as follows:

[0035] ;

[0036] Specifically, the Sigmoid activation function is used for the output nodes of state variables to constrain their physical value range, and a linear mapping method is used for the output nodes of coupling coefficients, thereby realizing continuous modeling of time-varying parameters.

[0037] Step S4.4: Construct a joint loss function by introducing data constraints, physical constraints, boundary constraints, and causal consistency constraints:

[0038] ;

[0039] in, , , and These are the weight coefficients for each loss term, used to adjust the relative importance of different constraints during the training process;

[0040] Step S4.5: Introduce supervised fitting loss to the state variables with measured data, which is defined as:

[0041] ;

[0042] in, The data constraint loss term, also known as the supervised fitting loss function, improves the model's ability to fit the real system behavior by minimizing the squared error between the predicted and measured values. This represents the total number of samples for the measured data. The predicted values ​​are for the damage variables at the magnet-casing bonding interface; The measured values ​​represent the damage variables at the magnet-casing bonding interface. The predicted value for the crack length at the potting compound-winding / core interface; This is the measured value of the crack length at the potting compound-winding / core interface;

[0043] Step S4.6: Embed the physical control equations, including damage evolution, crack propagation, wear, and diffusion, into the loss function in residual form to construct the physical constraint loss:

[0044] ;

[0045] The collocation points are generated in the spatiotemporal domain using the Latin hypercube sampling method to ensure that the physical constraints are effective throughout the entire domain. Loss due to physical constraints; This represents the total number of points; Let D be the derivative of the damage variable with respect to time. The physical model function for damage evolution depends on: Current damage variable, Surface temperature of the magnet Equivalent cyclic stress amplitude; Let be the derivative of the crack length 'a' with respect to the number of load cycles 'N'; For the physical model of crack propagation, Current crack length, Effective stress intensity factor range; The time derivative of the wear amount W; This is a wear-corrosion physical model, corresponding to the extended Arcard's law in step S2.2, where W is the cumulative wear volume and I is the contact current. For dynamic contact resistance; The time partial derivative of the concentration C of the permeating medium; The effective diffusion coefficient; The spatial second partial derivative of concentration C;

[0046] Step S4.7: Introduce initial state and boundary condition constraints of the system, and construct a boundary loss term to ensure that the model prediction results satisfy the known physical conditions at the initial time and boundary position:

[0047] ;

[0048] in, Loss is due to boundary condition constraints; The values ​​of the damage variables at the initial time t=0, as predicted by the neural network; The initial crack length predicted by the neural network; The concentration of the medium on the outer surface of the sealing layer is predicted by the neural network; 0 indicates that the damage to the brand-new motor is zero, and it is in a damage-free state. The initial concentration of the medium within the entire sealing layer as predicted by the neural network;

[0049] Step S4.8: Introduce causal consistency loss to constrain the network's learning results from violating known physical causal relationships. Its definition is:

[0050]

[0051] in, The total loss for causal consistency; The total number of causal constraint rules;

[0052] The penalty for the thermo-mechanical coupling coefficient under low-temperature conditions is:

[0053]

[0054] in, It is the thermo-mechanical coupling coefficient; For temperature, Temperature threshold; Indicates the current stress level;

[0055] The causal relationship between penetration depth and internal state is as follows:

[0056]

[0057] in, It is an internal physical quantity; This refers to the penetration depth of the sealing layer. This indicates the sensitivity of internal quantities to penetration depth;

[0058] The damage evolution constraint under low stress conditions is the penetration depth of the sealing layer:

[0059]

[0060] Where D is the damage variable; This represents the rate of damage accumulation. This represents the equivalent cyclic stress amplitude. The stress threshold; It is a tiny positive number;

[0061] The arc erosion penalty under low current conditions is:

[0062]

[0063] in, The cumulative material loss caused by arc erosion; Arc erosion rate; For current; The current threshold;

[0064] Finally, the monotonicity constraint of the temperature-diffusion coefficient is defined as follows:

[0065]

[0066] in, The effective diffusion coefficient; This is the partial derivative of the diffusion coefficient with respect to temperature;

[0067] The above constraints help avoid generating unreasonable strong coupling relationships under low-excitation conditions;

[0068] Step S4.9: Optimize the model according to the phased training strategy, initially using only... and Perform physical pre-training, and then introduce Make minor adjustments to the data, and finally add... The L-BFGS optimization algorithm is used for fine convergence, and Dropout, gradient pruning and early stopping strategies are combined to improve the stability and generalization performance of model training.

[0069] Step S4.10: After the model training is complete, the multiphysics time-varying coupling coefficient is directly obtained from the neural network output, and its expression is: ;in, is the thermo-mechanical coupling coefficient, which characterizes the coupling strength between the temperature field and the mechanical field; The function mapping relationship learned by the neural network; The vibration stress field specifically refers to the equivalent stress amplitude caused by vibration. The chemical-mechanical coupling coefficient characterizes the synergistic effect between chemical corrosion and mechanical damage. This refers to the concentration of corrosive ions. Plastic strain rate, characterizing the deformation rate of a material; is the electro-thermal coupling coefficient, which characterizes the coupling between current effects and thermal effects; These are contact current, dynamic contact resistance, reflecting contact quality, and temperature gradient, respectively, characterizing the non-uniformity of heat flow distribution.

[0070] Step S5: Based on the particle swarm optimization algorithm, the failure chain weights are dynamically reconstructed using real-time monitoring data to identify the dominant failure mode, including the following process:

[0071] Step S5.1: Model the dynamic reconstruction problem of the failure chain as a multivariate optimization problem. Its optimization objective is to find the failure chain weight allocation scheme that best explains the current degradation state of the system under the given monitoring data.

[0072] Step S5.2: Define the spatial position of each particle in the particle swarm as the failure chain weight vector:

[0073] ;

[0074] in, ) indicates the interface Independent degradation weights, Represents the independent degradation weight of the sealed interface; This represents the independent degradation weight of the potting interface. These represent the independent degradation weights of the carbon brush and magnet interface, respectively. Display Interface To the interface The coupling propagation strength is determined, and a normalization constraint is applied. To ensure the physical interpretability of failure contribution allocation;

[0075] Step S5.3: Construct the fitness function for particle evaluation, which is defined as:

[0076] ;

[0077] The prediction error term is defined as follows: ; Configure based on current weights Predicted remaining lifespan; Lifetime estimated by back-calculation of monitoring data; physical consistency penalty. The model complexity term is defined to quantify the degree of violation of physical constraints. To suppress unnecessary coupling paths;

[0078] Step S5.4: At the initial time Randomly generate a particle swarm, with the number of particles set to [number]. The initial position of the particle satisfies , Let be the initial position of the i-th particle. A uniform distribution between 0 and 1. Represents a 10-dimensional hypercube space; particle velocities are initialized to... , Let be the initial velocity of the i-th particle; Indicates the maximum speed limit; further initializes the individual optimal solution. Global optimal solution Defined as:

[0079] ;

[0080] Step S5.5: Update the particle velocity at the (t+1)th iteration, using the following formula:

[0081] ;

[0082] in, Inertial weights; The current particle velocity; For individual learning factors; for Random numbers between; As a social learning factor; The historical best position of the i-th particle; This represents the current position of the i-th particle; The globally optimal position; The physical guiding coefficient; For physical gradient guiding terms;

[0083] Then speed limit is applied. And perform location updates . This represents the truncation function, which will... Limited to between;

[0084] Step S5.6: Trim the boundaries of the updated particle positions:

[0085] ;

[0086] Further normalization is performed on the independent degradation weights to satisfy the physical constraints:

[0087] ;

[0088] in, The first four components of the particle position vector are: 1-sealing, 2-encapsulation, 3-carbon brush, and 4-magnet.

[0089] Step S5.7: Introduce the physical gradient guiding term The penalty function is as follows: Includes: when sealing weight Larger but monitoring penetration depth Smaller penalty items, when filling weight Larger crack propagation rate Penalty term for slow changes, and when the coupled path weights An infinite penalty term is imposed when the physical causal order is violated, in order to guide the particle to converge toward the physically plausible region;

[0090] Step S5.8: When the cumulative running time reaches a preset threshold, the prediction deviation exceeds 20%, or the operating conditions change significantly, the dynamic reconstruction process is triggered. The particle swarm optimization algorithm is run by extracting the most recent time window data, and the updated failure chain weight matrix is ​​output. The evolution history of the weights is recorded for subsequent analysis.

[0091] Step S6: Establish a three-level mapping model of material-component-system, mapping the interface damage state to system-level observable performance indicators, including the following steps:

[0092] Step S6.1: Definition of material layer state parameters; Based on the modeling results of steps S2 and S3, the damage state quantities of each weak interface are defined as the set of material layer input parameters, including: damage variables of the magnet-casing bonding interface. Encapsulant – Crack length at the winding / core interface Contact resistance of the carbon brush-commutator contact interface With wear Penetration depth of the sealant bonding interface The above parameters characterize the degradation state of the material layer under the action of multiple physics fields, and serve as the basic input for subsequent mapping models.

[0093] Step S6.2: Component layer functional degradation modeling; mapping material layer damage parameters to component layer functional indicators, establishing the influence relationship of interface damage on the performance of key components, including: through magnet bonding damage A mapping model between magnetic flux density non-uniformity and torque pulsation amplitude was established based on the relationship with magnetic circuit offset; the crack length at the potting interface was used as a reference. The influence of winding thermal resistance and insulation performance was investigated, and a mapping relationship between winding loss and efficiency decay was established; this was achieved through carbon brush contact resistance. By varying the spark frequency, a mapping relationship between brush commutation quality and electromagnetic interference intensity was established; the penetration depth at the sealing interface was also investigated. A model is established to assess the impact of environmental media intrusion on component corrosion and electrical performance degradation in relation to changes in internal humidity. The above mapping relationship can be achieved using a semi-empirical model, lookup table function, or a calibrated simplified physical model to ensure computational efficiency and engineering feasibility.

[0094] Step S6.3: System-level performance index mapping; Based on the component-level model, further construct the system-level performance output model to map the component functional degradation to system-level observable performance indices, including: motor output torque fluctuation amplitude. System efficiency The downward trend; electromagnetic interference intensity or equivalent noise index This three-level mapping model enables the transformation from "interface damage state that cannot be directly observed" to "system performance indicators that can be monitored in real time," providing a unified benchmark for subsequent prediction verification and model correction.

[0095] Step S7: Construct a digital twin model integrating PINN and PSO weights, and use Bayesian inference to update the model parameters online to achieve quantified propagation of uncertainty and probabilistic lifetime prediction, including the following steps:

[0096] Step S7.1: Construct a digital twin system consisting of a physical entity layer, a twin model layer, a communication layer, and a decision layer. The twin model layer integrates the PINN multi-field model and the PSO failure chain weights to reflect the system state evolution in real time.

[0097] Step S7.2: Model the uncertainty of the model parameters as a Bayesian inference problem. Based on existing historical knowledge, update the uncertainty of the model parameters using newly acquired data. The posterior distribution is defined as:

[0098] ;

[0099] in, Represents the set of parameters to be estimated; This is a historical observation dataset; This refers to newly acquired online monitoring data; It is the prior distribution; To observe the likelihood function; This is evidence.

[0100] Step S7.3: Set the prior distribution of the parameters as follows:

[0101] ;

[0102] in, The prior mean; The prior variance reflects the uncertainty of the parameters;

[0103] The observation likelihood function is defined as:

[0104] ;

[0105] in, For the first One observation response; For the prediction output of the data-driven model, For the corresponding input variables; To observe the noise variance;

[0106] Step S7.4: Use the Markov chain Monte Carlo method to sample the posterior distribution of the parameters to obtain the expected value of the parameters and the uncertainty measure, which will be used for subsequent prediction and propagation analysis;

[0107] Step S7.5: Calculate the lifetime prediction uncertainty based on the posterior distribution of the parameters:

[0108]

[0109] in, Indicates the remaining lifespan of the motor. For monitoring data / performance observation datasets; The set of parameters for the model to be estimated; It is the posterior distribution of parameters obtained by the Markov chain Monte Carlo method; For a given The lifetime condition distribution under parametric conditions is obtained by joint simulation of the failure mechanism model, failure propagation network, and PINN coupling term;

[0110] Step S7.6: Use the Monte Carlo method to perform uncertainty propagation and statistics to verify the numerical estimation of the integral in step S7.5;

[0111] First, posterior sampling is performed, drawing samples from the posterior distribution of the parameters. Group samples Then, lifetime calculation is performed for each set of parameter samples. The digital twin model is driven to extrapolate the failure evolution into the future, using the first arrival time of "reaching the failure threshold" as the lifetime output; finally, the expected lifetime and interval estimation are performed, the expected lifetime is calculated from the sample set, and the interval estimation is given by quantiles.

[0112] Step S8: By comparing the predicted results with the measured data in a closed loop, adaptively triggering model parameter updates and failure chain reconstruction, a continuously self-evolving prediction-verification-correction mechanism is formed, including the following steps:

[0113] Step S8.1: In each prediction period Within the digital twin model based on step S7, using the monitoring data at the current moment and the updated model parameters, the system-level performance indicators and remaining lifetime prediction results are generated, including: system performance prediction values, probability distribution of remaining lifetime and confidence intervals.

[0114] Step S8.2: Compare the predicted system-level performance indicators with the actual monitored values, and calculate the prediction error indicators, including mean square error, relative error or probability coverage, to quantify the model prediction accuracy.

[0115] Step S8.3: Based on the preset error threshold or operating condition change criterion, determine whether to trigger the model self-correction process; the self-evolution mechanism is triggered when any of the following conditions are met:

[0116] The prediction error exceeds the set threshold for multiple consecutive periods;

[0117] The confidence interval for predicted remaining lifespan deviates significantly from the measured degradation trend.

[0118] The operating conditions have changed significantly, causing the original failure mode assumptions to no longer hold.

[0119] Step S8.4: When the triggering condition is met, automatically perform the following operations: call the Bayesian inference method in step S7 to update the posterior distribution of the key model parameters online; call the particle swarm optimization algorithm in step S5 to dynamically reconstruct the failure chain weights and adjust the dominant failure path and its propagation intensity.

[0120] Step S8.5: Synchronize the updated model parameters and failure chain structure to the digital twin model and proceed to the next prediction cycle. This forms a closed-loop self-evolutionary operation mode of "prediction-verification-correction-re-prediction", enabling the motor failure prediction model to continuously adapt and improve its performance throughout the entire service life.

[0121] This invention achieves an organic integration of data-driven and physical constraints through intelligent modeling driven by physical knowledge, utilizing a physical information neural network to automatically learn time-varying coupling coefficients; dynamic reconstruction of failure paths, using a particle swarm optimization algorithm to identify dominant failure modes in real time and dynamically adjust the weights of the failure propagation chain; a prediction-verification-correction closed loop, leveraging digital twins and Bayesian inference to achieve continuous self-evolution of the model and continuously improve prediction accuracy; intelligent design of test schemes, using a multi-objective genetic algorithm to optimize and accelerate test conditions, improving test efficiency and failure mode coverage; and probabilistic lifetime prediction, outputting the remaining lifetime distribution with confidence intervals to quantify prediction uncertainty. This invention deeply integrates physical mechanisms and data-driven approaches, achieving high-precision, adaptive, and interpretable prediction of multi-field coupling failures in motors. Attached Figure Description

[0122] Figure 1 This is a flowchart illustrating the present invention. Detailed Implementation

[0123] The present invention will now be described in further detail with reference to the accompanying drawings and specific preferred embodiments.

[0124] This invention achieves an organic integration of data-driven and physical constraints through intelligent modeling driven by physical knowledge, utilizing a physical information neural network to automatically learn time-varying coupling coefficients; dynamic reconstruction of failure paths, using a particle swarm optimization algorithm to identify dominant failure modes in real time and dynamically adjust the weights of the failure propagation chain; a prediction-verification-correction closed loop, leveraging digital twins and Bayesian inference to achieve continuous self-evolution of the model and constantly improve prediction accuracy; intelligent design of experimental schemes, using a multi-objective genetic algorithm to optimize and accelerate experimental conditions, improving experimental efficiency and failure mode coverage; and probabilistic lifetime prediction, outputting the remaining lifetime distribution with confidence intervals to quantify prediction uncertainty.

[0125] An adaptive failure prediction method for motors based on digital twin fusion intelligent optimization includes the following steps:

[0126] Step S1: Deploy a multimodal sensor array to collect multi-source data and perform preprocessing and feature extraction;

[0127] Step S1.1: Deploy a multimodal sensor array at key locations on the motor, specifically including:

[0128] Temperature sensing network used to measure the surface temperature of magnets. Winding end temperature Bearing temperature Casing surface temperature: Sampling frequency 1Hz, accuracy ±0.5℃.

[0129] Vibration sensing network, monitoring signals are triaxial accelerations , , Sampling frequency ≥ 5kHz, measuring range ± 50g; installation positions are distributed at the front and rear end covers of the motor and the middle of the housing.

[0130] Electrical parameter monitoring signals include three-phase current. , , Bus voltage commutator spark voltage Sampling frequency 10kHz.

[0131] Environmental parameter monitoring includes relative humidity (RH(t)) and salt spray concentration. (Measured indirectly via a conductivity sensor) Atmospheric pressure .

[0132] Step S1.2: Preprocess and extract features from the original monitoring signal at the edge computing layer to obtain key features that can directly characterize the motor's operating status and degradation trend.

[0133] Specifically, the vibration signal undergoes time-domain, frequency-domain, and time-frequency-domain analysis to extract time-domain features (including root mean square value, peak value, kurtosis, and skewness) and frequency-domain features (including 1×, 2×, and 3× harmonic characteristic amplitudes related to rotational speed). Energy distribution is then obtained through wavelet packet transform to effectively capture fault information in non-stationary vibrations. The current signal is analyzed through armature current to calculate the total harmonic distortion (THD), reflecting electrical performance degradation, and the pulse count rate, used to characterize commutation status and count the number of spike pulses per unit time. After processing, the temperature signal is used to extract features related to the thermal state, including the maximum temperature rise relative to the ambient temperature. And the temperature gradient between key components used to characterize internal heat conduction properties (such as magnets and housing). .

[0134] S1.3: To ensure the reliability and consistency of the input model data, a rigorous real-time quality control process is implemented. Based on the statistical 3σ criterion, significant outlier data points (outliers) caused by sensor or transmission noise are identified and removed in real time. For data gaps caused by brief signal interruptions, a Kalman filter algorithm is used for real-time, optimal estimation and interpolation to ensure the continuity of the data stream. To fuse information from multiple sensor sources, high-precision timestamp alignment is performed on data streams from different physical channels to ensure the accuracy of subsequent correlation analysis.

[0135] S1.4: Perform data standardization. To ensure that multi-source heterogeneous feature data can be learned efficiently and stably by subsequent deep learning models, after edge feature extraction, the data stream is uniformly standardized using the Z-score standardization method based on the mean and standard deviation of the training set, in order to accelerate model convergence and improve prediction accuracy.

[0136] Step S2: Identify the failure mechanisms of the four weak interfaces inside the motor: magnet-housing, potting compound-winding / core, carbon brush-commutator, and sealing material, and establish the corresponding physical damage evolution equations and failure propagation networks.

[0137] Step S2.1: Identify systemic weaknesses; using a comprehensive method combining failure mode and effect analysis (FMEA) and finite element simulation, systematically identify four core weak interfaces within the motor and their dominant failure mechanisms. Weak points include:

[0138] The bonding interface between the magnet and the housing is prone to fatigue cracking of the bonding layer under thermal-vibration coupling stress, which can help monitor torque fluctuations. With vibration acceleration ;

[0139] The potting compound at the winding / core interface is prone to crack propagation due to thermal cycling and hydrolysis, which can be addressed by measuring insulation resistance. With efficiency Conduct monitoring;

[0140] The carbon brush-commutator contact interface suffers from a complex failure resulting from arc erosion, mechanical wear, and electrochemical corrosion, involving multi-field coupling across electro-thermal-mechanical-chemical fields. Spark voltage can be monitored. Contact resistance With electromagnetic interference intensity;

[0141] The sealing performance of the bonding interface of the sealing material deteriorates due to thermo-oxidative aging and stress relaxation; its condition can be monitored by internal humidity. and penetration depth based on impedance spectrum inversion To characterize;

[0142] Step S2.2: Establish a set of basic physical equations. Based on the four identified weak links and their dominant failure mechanisms, establish a quantitative physical model for each interface to describe its damage evolution or performance degradation process.

[0143] Damage evolution equation for the magnet-casing bonding interface; the fatigue damage accumulation process of the bonding layer at this interface under alternating thermal-vibration coupled stress is described by an evolution equation based on damage mechanics:

[0144]

[0145] in, For the interface damage variable, the range is: to These correspond to the lossless state and complete failure, respectively. From temperature field and vibration acceleration field The equivalent cyclic stress amplitude obtained from coupled calculations; A constant reflecting the fatigue performance of a material; Activation energy associated with material damage evolution, in units of ; This is the universal gas constant. .

[0146] Crack propagation equation at the potting compound-winding / core interface; the crack propagation behavior at this interface under the combined effects of thermal stress and hydrolysis is described by a modified Paris equation:

[0147] ;

[0148] Effective stress intensity factor range It combines the contributions of mechanical, thermal loads, and chemical effects: .

[0149] in, Crack length, in units of ; The number of thermomechanical load cycles; For fatigue crack propagation parameters of the material; This represents the range of stress intensity factor variation caused by mechanical stress. The range of stress intensity factor variation caused by thermal stress; The rate constant for the hydrolysis reaction is denoted by temperature. and media intrusion rate The function.

[0150] Wear-corrosion equation for the carbon brush-commutator contact interface; the material loss at this interface is the superposition of mechanical wear, arc erosion, electrochemical corrosion, and their synergistic effects, described by an extended Arcard wear law:

[0151]

[0152] The specific expressions for each component are as follows:

[0153]

[0154] in, The total material volume loss rate, in units of ; The mechanical wear coefficient; Contact pressure, unit is ; The relative sliding speed is expressed in units of 1000 m / s. ; Material hardness, unit: ; This is the arc erosion coefficient; Contact current, unit: ; Dynamic contact resistance, unit: ; Spark frequency, unit: ; The electrochemical corrosion coefficient; This represents the concentration of corrosive ions, in units of... ; The activation energy of the corrosion reaction is expressed in units of 1. ; This is the coefficient representing the synergistic effect of wear and corrosion.

[0155] The permeation-diffusion equation at the bonding interface of sealing materials; the process of external medium permeating into the sealing material under humid heat and stress aging is described by Fick's second law, which is influenced by temperature and stress:

[0156]

[0157] Among them, the effective diffusion coefficient It is a function of temperature and stress:

[0158]

[0159] Boundary conditions and initial conditions are set as follows , .

[0160] in, The concentration of the permeation medium is expressed in units of 1000 mg / L. , The coordinates are along the direction of the sealing layer thickness, in units of ; For time, the unit is ; The effective diffusion coefficient is expressed in units of 1000 ppm. , The pre-exponential factor of the diffusion coefficient, in units of ; The activation energy for the diffusion process is expressed in units of 1000 kJ / m². ; This is the stress enhancement factor, in units of... ; The hydrostatic pressure exerted on the sealing layer, in units of ; Concentration of environmental media, unit: ; The thickness of the sealing layer is expressed in units of 1. .

[0161] Step S2.3: Establish the failure propagation network among the four weak links of the motor and reveal their interaction and cascade effect.

[0162] The penetration depth of the sealing interface is reduced due to thermo-oxidative aging and stress relaxation. Increase, thereby causing the rate of intrusion into the environmental media. The reaction rate rises and triggers three parallel propagation paths: firstly, it accelerates the hydrolysis reaction rate at the potting interface. This, in turn, promotes the crack propagation rate. Increased; secondly, it exacerbated the electrochemical corrosion rate at the carbon brush interface. This leads to contact resistance This increases the stress corrosion rate at the magnet interface, ultimately reducing the equivalent stress intensity factor. Increase, thereby accelerating the rate of interface damage accumulation. To quantify the propagation strength of the failure chain, a coupling strength matrix is ​​defined:

[0163] ;

[0164] Among them, matrix elements Display Interface For the interface Failure propagation strength weight ( These correspond to the interfaces of sealing, potting, carbon brush, and magnet, respectively, and their initial values ​​can be set based on expert experience.

[0165] Step S3: Optimize the accelerated life test scheme using a multi-objective genetic algorithm;

[0166] Based on the multiphysics field coupled failure mechanism modeling and failure chain dynamic reconstruction, a multi-objective optimization model for accelerated life testing oriented towards the failure of key interfaces of motors is constructed.

[0167] Using test temperature, mechanical load, electromagnetic excitation intensity, and environmental stress as decision variables, an accelerated stress feasible region is formed under equipment capacity and safety constraints. With minimum test time, maximum failure mode coverage, and minimum accelerated equivalent deviation as optimization objectives, the non-dominated sorting genetic algorithm NSGA-II is used to solve the problem and obtain the Pareto optimal solution set corresponding to multiple test conditions.

[0168] By combining the failure mechanism model in step S2 with the digital twin system in step S7, the Pareto solution is screened and evaluated to predict the interface damage evolution and failure chain activation characteristics. The accelerated life test scheme with the best overall efficiency, coverage and equivalence consistency is selected to provide high-quality test data support for model calibration and life prediction.

[0169] Step S4: Construct a physical information neural network and automatically learn the time-varying coupling coefficient between state variables and multiphysics fields by embedding a hybrid loss function of physical control equations, boundary conditions and causal constraints;

[0170] Step S4.1: Constructing network input variables based on multi-source operating condition information; constructing the input layer variables of the physical information neural network to characterize the multi-physics operating condition state of the system in the time dimension. The input vector is defined as:

[0171] ;

[0172] in, For temperature parameters, For vibration stress parameters, For current parameters, This refers to the ambient relative humidity parameter. For input voltage parameters, The input is a time variable; the above inputs enable unified modeling of multi-physics operating condition information.

[0173] Step S4.2: Joint output modeling of state variables and time-varying coupling parameters; construct the output layer variables of the neural network to simultaneously characterize the system state evolution process and the coupling relationship between multiple physics fields. Its output vector is defined as:

[0174] ;

[0175] in, , , , These represent damage variables, crack length, contact resistance, and penetration depth, respectively. , , The time-varying coupling coefficients to be learned are used to describe the dynamic interaction strength between different physical fields;

[0176] Step S4.3: Construction of a hybrid neural network structure with embedded physical constraints; construct a hybrid fully connected neural network, the network structure of which is as follows:

[0177] ;

[0178] Specifically, the Sigmoid activation function is used for the output nodes of state variables to constrain their physical value range, and a linear mapping method is used for the output nodes of coupling coefficients, thereby realizing continuous modeling of time-varying parameters.

[0179] Step S4.4: Construction of the joint loss function with multiple constraints; By introducing data constraints, physical constraints, boundary constraints, and causal consistency constraints, a joint loss function is constructed:

[0180] ;

[0181] in, , , and These are the weight coefficients for each loss term, used to adjust the relative importance of different constraints during the training process;

[0182] Step S4.5: Introduce supervised fitting loss to the state variables with measured data, which is defined as:

[0183] ;

[0184] in, The data constraint loss term, also known as the supervised fitting loss function, improves the model's ability to fit the real system behavior by minimizing the squared error between the predicted and measured values. This represents the total number of samples for the measured data. The predicted values ​​are for the damage variables at the magnet-casing bonding interface; The measured values ​​represent the damage variables at the magnet-casing bonding interface. The predicted value for the crack length at the potting compound-winding / core interface; This is the measured value of the crack length at the potting compound-winding / core interface;

[0185] Step S4.6: Constructing the residual constraint loss based on the physical control equations; embedding the physical control equations, including damage evolution, crack propagation, wear, and diffusion, into the loss function in residual form to construct the physical constraint loss:

[0186] ;

[0187] The collocation points are generated in the spatiotemporal domain using the Latin hypercube sampling method to ensure that the physical constraints are effective throughout the entire domain. This represents the physical constraint loss (the residual of the physical governing equations). The total number of collocation points (physical constraint points generated in the spatiotemporal domain through Latin hypercube sampling); The derivative of the damage variable D with respect to time (obtained by automatic differentiation from a neural network); The physical model function for damage evolution depends on: Current damage variable, Surface temperature of the magnet Equivalent cyclic stress amplitude (calculated by coupling the temperature field and the vibration field); The derivative of crack length *a* with respect to the number of load cycles *N* (calculated using a neural network); For the crack propagation physical model (modified Paris formula), Current crack length, Effective stress intensity factor range; The time derivative of the wear amount W; This is a wear-corrosion physical model, corresponding to the extended Arcard's law in step S2.2, where W is the cumulative wear volume and I is the contact current. For dynamic contact resistance; The time partial derivative of the concentration C of the permeating medium; The effective diffusion coefficient (a function of temperature and stress); The spatial second-order partial derivative of concentration C (automatic differentiation by neural network);

[0188] Step S4.7: Constructing Consistency Constraints for Initial and Boundary Conditions; Introducing initial state and boundary condition constraints of the system, constructing boundary loss terms to ensure that the model prediction results satisfy the known physical conditions at the initial time and boundary position:

[0189] ;

[0190] in, Loss is due to boundary condition constraints; The values ​​of the damage variables at the initial time t=0, as predicted by the neural network; The initial crack length predicted by the neural network; The medium concentration on the outer surface of the sealing layer (at position x=0) is predicted by the neural network; 0 indicates that the damage to the brand-new motor is zero, and it is in an undamaged state. The initial concentration of the medium within the entire sealing layer as predicted by the neural network;

[0191] Step S4.8: Constructing consistency constraints based on physical causal logic; introducing causal consistency loss to restrict the network learning results from violating known physical causal relationships, which is defined as:

[0192]

[0193] in, The total loss for causal consistency; This represents the total number of causal constraint rules (M=5 in this example);

[0194] The penalty for the thermo-mechanical coupling coefficient under low-temperature conditions is:

[0195]

[0196] in, This is the thermo-mechanical coupling coefficient (output by the neural network, corresponding to the time-varying coupling parameter in step S4.2); For temperature, Temperature threshold; Indicates the current stress level;

[0197] The causal relationship between penetration depth and internal state is as follows:

[0198]

[0199] in, These are internal physical quantities (such as: effective sealed volume, internal cleanliness, insulation performance, etc.); This refers to the penetration depth of the sealing layer. This indicates the sensitivity of internal quantities to penetration depth;

[0200] The damage evolution constraint under low stress conditions is the penetration depth of the sealing layer:

[0201]

[0202] Where D is the damage variable; The rate of damage accumulation (calculated by the neural network through automatic differentiation); This represents the equivalent cyclic stress amplitude. This is the stress threshold (e.g., 50% of the fatigue limit); For small positive numbers (e.g.: / h indicates acceptable minute damage);

[0203] The arc erosion penalty under low current conditions is:

[0204]

[0205] in, Cumulative material loss (mm³) caused by arc erosion; Arc erosion rate (mm³ / h); For current; The current threshold;

[0206] Finally, the monotonicity constraint of the temperature-diffusion coefficient is defined as follows:

[0207]

[0208] in, The effective diffusion coefficient (m² / s, controlling the permeation rate of the medium in the sealing layer); The partial derivative of the diffusion coefficient with respect to temperature ( (for temperature)

[0209] The above constraints help avoid generating unreasonable strong coupling relationships under low-excitation conditions;

[0210] Step S4.9: Execution of phased training and optimization strategy; optimize the model according to the phased training strategy, initially using only... and Conduct physical pre-training, and then introduce Make minor adjustments to the data, and finally add... The L-BFGS optimization algorithm is used for fine convergence, and Dropout, gradient pruning and early stopping strategies are combined to improve the stability and generalization performance of model training.

[0211] Step S4.10: Extraction and characterization of multiphysics time-varying coupling coefficients; After model training is completed, the multiphysics time-varying coupling coefficients are directly obtained from the neural network output, and their expression is: ;in, is the thermo-mechanical coupling coefficient, which characterizes the coupling strength between the temperature field and the mechanical field; The function mapping relationship learned by the neural network; Vibration stress field (MPa), specifically refers to the equivalent stress amplitude caused by vibration; The chemical-mechanical coupling coefficient characterizes the synergistic effect between chemical corrosion and mechanical damage. This refers to the concentration of corrosive ions (mol / L or ppm), such as chloride ions and sulfate ions. σ is the plastic strain rate (s⁻¹), which characterizes the deformation rate of the material; is the electro-thermal coupling coefficient, which characterizes the coupling between current effects and thermal effects; These are contact current (A), dynamic contact resistance (Ω), reflecting contact quality, and temperature gradient (℃ / mm or K / m), characterizing the non-uniformity of heat flow distribution.

[0212] Step S5: Based on the particle swarm optimization algorithm, dynamically reconstruct the failure chain weights using real-time monitoring data to identify the dominant failure mode;

[0213] Step S5.1: Model the failure chain weight optimization problem; model the failure chain dynamic reconstruction problem as a multivariate optimization problem, the optimization objective of which is to find the failure chain weight allocation scheme that best explains the current degradation state of the system under the given monitoring data.

[0214] Step S5.2: Define particle position variables and constraints; define the spatial position of each particle in the particle swarm as the failure chain weight vector:

[0215] ;

[0216] in, ) indicates the interface Independent degradation weights, Represents the independent degradation weight of the sealed interface; This represents the independent degradation weight of the potting interface. These represent the independent degradation weights of the carbon brush and magnet interface, respectively. Display Interface To the interface The coupling propagation strength ( : Sealing → Potting; : Sealing → Potting; (Sealing → Magnet...), and applying normalization constraints. To ensure the physical interpretability of failure contribution allocation;

[0217] Step S5.3: Construction of Multi-Constraint Fitness Function; Construct a fitness function for particle evaluation, defined as:

[0218] ;

[0219] The prediction error term is defined as follows: ; Configure based on current weights Predicted remaining lifespan; Lifetime estimated by back-calculation of monitoring data; physical consistency penalty. The model complexity term is defined to quantify the degree of violation of physical constraints. To suppress unnecessary coupling paths;

[0220] Step S5.4: Initialize the particle swarm; at the initial time... Randomly generate a particle swarm, with the number of particles set to [number]. The initial position of the particle satisfies , Let be the initial position of the i-th particle. A uniform distribution between 0 and 1. Represents a 10-dimensional hypercube space; particle velocities are initialized to... , Let be the initial velocity of the i-th particle; Indicates the maximum speed limit; further initializes the individual optimal solution. Global optimal solution Defined as:

[0221] ;

[0222] Step S5.5: Update particle velocity and position; update the particle velocity at the (t+1)th iteration, using the following formula:

[0223] ;

[0224] in, Inertial weights; The current particle velocity; For individual learning factors; for Random numbers between; As a social learning factor; The historical best position of the i-th particle; This represents the current position of the i-th particle; It is the globally optimal position (the best position in the entire particle swarm); The physical guiding coefficient; For physical gradient guiding terms;

[0225] Then speed limit is applied. And perform location updates , This represents the truncation function, which will... Limited to between;

[0226] Step S5.6: Boundary processing and normalization constraint execution; perform boundary trimming on the updated particle positions:

[0227] ;

[0228] Further normalization is performed on the independent degradation weights to satisfy the physical constraints:

[0229] ;

[0230] in, The first four components (independent degenerate weights) of the particle position vector are: 1-sealing, 2-encapsulation, 3-carbon brush, and 4-magnet.

[0231] Step S5.7: Construction of the physical gradient guiding term; Introduction of the physical gradient guiding term The penalty function is as follows: Includes: when sealing weight Larger but monitoring penetration depth Smaller penalty items, when filling weight Larger crack propagation rate Penalty term for slow changes, and when the coupled path weights An infinite penalty term is imposed when the physical causal order is violated, in order to guide the particle to converge toward the physically plausible region;

[0232] Step S5.8: Online Dynamic Reconfiguration Triggering and Execution; When the cumulative running time reaches a preset threshold, the prediction deviation exceeds 20%, or the operating conditions change significantly, the dynamic reconfiguration process is triggered. The particle swarm optimization algorithm is run by extracting the most recent time window data, and the updated failure chain weight matrix is ​​output. The evolution history of the weights is recorded for subsequent analysis.

[0233] Step S6: Establish a three-level mapping model of material-component-system to map the interface damage state into observable performance indicators at the system level;

[0234] Step S6.1: Definition of material layer state parameters; Based on the modeling results of steps S2 and S3, the damage state quantities of each weak interface are defined as the set of material layer input parameters, including: damage variables of the magnet-casing bonding interface. Encapsulant – Crack length at the winding / core interface Contact resistance of the carbon brush-commutator contact interface With wear Penetration depth of the sealant bonding interface The above parameters characterize the degradation state of the material layer under the action of multiple physics fields, and serve as the basic input for subsequent mapping models.

[0235] Step S6.2: Component layer functional degradation modeling; mapping material layer damage parameters to component layer functional indicators, establishing the influence relationship of interface damage on the performance of key components, including: through magnet bonding damage A mapping model between magnetic flux density non-uniformity and torque pulsation amplitude was established based on the relationship with magnetic circuit offset; the crack length at the potting interface was used as a reference. The influence of winding thermal resistance and insulation performance was investigated, and a mapping relationship between winding loss and efficiency decay was established; this was achieved through carbon brush contact resistance. By varying the spark frequency, a mapping relationship between brush commutation quality and electromagnetic interference intensity was established; the penetration depth at the sealing interface was also investigated. A model is established to assess the impact of environmental media intrusion on component corrosion and electrical performance degradation in relation to changes in internal humidity. The above mapping relationship can be achieved using a semi-empirical model, lookup table function, or a calibrated simplified physical model to ensure computational efficiency and engineering feasibility.

[0236] Step S6.3: System-level performance index mapping; Based on the component-level model, further construct the system-level performance output model to map the component functional degradation to system-level observable performance indices, including: motor output torque fluctuation amplitude. System efficiency The downward trend; electromagnetic interference intensity or equivalent noise index This three-level mapping model enables the transformation from "interface damage state that cannot be directly observed" to "system performance indicators that can be monitored in real time," providing a unified benchmark for subsequent prediction verification and model correction.

[0237] Step S7: Construct a digital twin model integrating PINN and PSO weights, and use Bayesian inference to update the model parameters online to achieve quantitative propagation of uncertainty and probabilistic lifetime prediction;

[0238] Step S7.1: Construction of a multi-layered digital twin architecture; Construct a digital twin system consisting of a physical entity layer, a twin model layer, a communication layer, and a decision layer. The twin model layer integrates the PINN multi-field model and the PSO failure chain weights to reflect the system state evolution in real time.

[0239] Step S7.2: Modeling the Online Bayesian Parameter Update Problem; The uncertainty of the model parameters is modeled as a Bayesian inference problem. Based on existing historical knowledge, the uncertainty of the model parameters is updated using newly acquired data. Its posterior distribution is defined as:

[0240] ;

[0241] in, This represents the set of parameters to be estimated (including material / interface parameters in the multiphysics failure mechanism model, such as fatigue constant, diffusion coefficient, and corrosion coefficient), the time-varying coupling coefficient parameterization term learned by PINN, and the key weight parameters of the failure propagation network, etc.). This is a historical observation dataset (data that has been merged from the previous moment). For newly acquired online monitoring data (penetration depth, crack length, response signal); It is the prior distribution; To observe the likelihood function; This is evidence.

[0242] Step S7.3: Prior and Likelihood Function Construction; Set the prior distribution of parameters as follows:

[0243] ;

[0244] in, The prior mean; The prior variance reflects the uncertainty of the parameters;

[0245] The observation likelihood function is defined as:

[0246] ;

[0247] in, For the first Individual observation responses (such as penetration depth and crack length); For the prediction output of the data-driven model, For the corresponding input variables (time, load, environmental variables); To observe the noise variance;

[0248] Step S7.4: Parameter posterior sampling and statistics; The Markov chain Monte Carlo method is used to sample the parameter posterior distribution to obtain the expected value and uncertainty measure of the parameters for subsequent prediction and propagation analysis;

[0249] Step S7.5: Uncertainty Propagation and Lifetime Distribution Estimation; Calculate lifetime prediction uncertainty based on the posterior distribution of parameters:

[0250]

[0251] in, Indicates the remaining lifespan of the motor. For monitoring data / performance observation datasets; The set of parameters for the model to be estimated; It is the posterior distribution of parameters obtained by the Markov chain Monte Carlo method; For a given The lifetime condition distribution under parametric conditions is obtained by joint simulation of the failure mechanism model, failure propagation network, and PINN coupling term;

[0252] Step S7.6: Use the Monte Carlo method to perform uncertainty propagation and statistics to verify the numerical estimation of the integral in step S7.5;

[0253] First, posterior sampling is performed, drawing samples from the posterior distribution of the parameters. Group samples Then, lifetime calculation is performed for each set of parameter samples. The digital twin model is driven to extrapolate the failure evolution into the future, using the first arrival time of "reaching the failure threshold" as the lifetime output; finally, the expected lifetime and interval estimation are performed, the expected lifetime is calculated from the sample set, and the interval estimation is given by quantiles.

[0254] Step S8: By comparing the prediction results with the measured data in a closed loop, the model parameters are adaptively updated and the failure chain is reconstructed, forming a continuously self-evolving prediction-verification-correction mechanism;

[0255] Step S8.1: Prediction results are generated; in each prediction period Within the digital twin model based on step S7, using the monitoring data at the current moment and the updated model parameters, the system-level performance indicators and remaining lifetime prediction results are generated, including: system performance prediction values, probability distribution of remaining lifetime and confidence intervals.

[0256] Step S8.2: Compare the prediction results with the measured data; compare the predicted system-level performance indicators with the actual monitored values, and calculate the prediction error indicators, including mean square error, relative error or probability coverage, to quantify the model prediction accuracy.

[0257] Step S8.3: Adaptive Trigger Criterion Determination; Based on the preset error threshold or operating condition change criterion, determine whether to trigger the model self-correction process; The self-evolution mechanism is triggered when any of the following conditions are met:

[0258] The prediction error exceeds the set threshold for multiple consecutive periods;

[0259] The confidence interval for predicted remaining lifespan deviates significantly from the measured degradation trend.

[0260] The operating conditions have changed significantly, causing the original failure mode assumptions to no longer hold.

[0261] Step S8.4: Parameter update and failure chain reconstruction execution; When the triggering condition is met, the following operations are performed automatically: call the Bayesian inference method in step S7 to update the posterior distribution of key model parameters online; call the particle swarm optimization algorithm in step S5 to dynamically reconstruct the failure chain weights and adjust the dominant failure path and its propagation intensity.

[0262] Step S8.5: Closed-loop update and iterative operation; synchronize the updated model parameters and failure chain structure to the digital twin model, and enter the next prediction cycle. This forms a closed-loop self-evolutionary operation mode of "prediction-verification-correction-re-prediction", enabling the motor failure prediction model to continuously adapt and improve its performance throughout the entire service life.

[0263] In the description of this invention, it should be understood that the indicated orientation or positional relationship is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing this invention and simplifying the description, and is not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention.

[0264] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A motor adaptive failure prediction method based on digital twin fusion intelligent optimization, characterized in that: Includes the following steps: Step S1: Deploy a multimodal sensor array to collect multi-source data and perform preprocessing and feature extraction; Step S2: Identify the failure mechanisms of the four weak interfaces inside the motor: magnet-housing, potting compound-winding / core, carbon brush-commutator, and sealing material, and establish the corresponding physical damage evolution equations and failure propagation networks. Step S3: Optimize the accelerated life test scheme using a multi-objective genetic algorithm; Step S4: Construct a physical information neural network and automatically learn the time-varying coupling coefficient between state variables and multiphysics fields by embedding a hybrid loss function of physical control equations, boundary conditions and causal constraints; Step S5: Based on the particle swarm optimization algorithm, dynamically reconstruct the failure chain weights using real-time monitoring data to identify the dominant failure mode; Step S6: Establish a three-level mapping model of material-component-system to map the interface damage state into observable performance indicators at the system level; Step S7: Construct a digital twin model integrating PINN and PSO weights, and use Bayesian inference to update the model parameters online to achieve quantitative propagation of uncertainty and probabilistic lifetime prediction; Step S8: By comparing the prediction results with the measured data in a closed loop, the model parameters are adaptively updated and the failure chain is reconstructed, forming a continuously self-evolving prediction-verification-correction mechanism.

2. The adaptive failure prediction method for motors based on digital twin fusion intelligent optimization according to claim 1, characterized in that: Step S2: Physical modeling of coupled failure based on the identification of four weak links includes the following steps: Step S2.1: Identify systemic weaknesses; weaknesses include: The bonding interface between the magnet and the housing is prone to fatigue cracking of the bonding layer under thermal-vibration coupling stress, which can help monitor torque fluctuations. With vibration acceleration ; The potting compound at the winding / core interface is prone to crack propagation due to thermal cycling and hydrolysis, which can be addressed by measuring insulation resistance. With efficiency Conduct monitoring; The carbon brush-commutator contact interface suffers from a complex failure resulting from arc erosion, mechanical wear, and electrochemical corrosion, involving multi-field coupling across electro-thermal-mechanical-chemical fields. Spark voltage can be monitored. Contact resistance With electromagnetic interference intensity; The sealing performance of the bonding interface of the sealing material deteriorates due to thermo-oxidative aging and stress relaxation; its condition can be monitored by internal humidity. and penetration depth based on impedance spectrum inversion To characterize; Step S2.2: Based on the four identified weak links and their dominant failure mechanisms, establish a quantitative physical model for each interface to describe its damage evolution or performance degradation process; Step S2.3: Establish the failure propagation network among the four weak links of the motor and reveal their interaction and cascade effect.

3. The adaptive failure prediction method for motors based on digital twin fusion intelligent optimization according to claim 2, characterized in that: Step S3: Optimizing the accelerated life test scheme using a multi-objective genetic algorithm includes the following processes: Based on the multiphysics field coupled failure mechanism modeling and failure chain dynamic reconstruction, a multi-objective optimization model for accelerated life testing oriented towards the failure of key interfaces of motors is constructed. Using test temperature, mechanical load, electromagnetic excitation intensity, and environmental stress as decision variables, an accelerated stress feasible region is formed under equipment capacity and safety constraints. With minimum test time, maximum failure mode coverage, and minimum accelerated equivalent deviation as optimization objectives, the non-dominated sorting genetic algorithm NSGA-II is used to solve the problem and obtain the Pareto optimal solution set corresponding to multiple test conditions. The Pareto solutions are screened and evaluated to predict the interface damage evolution and failure chain activation characteristics. The accelerated life test scheme with the best overall efficiency, coverage and equivalence is selected to provide high-quality test data support for model calibration and life prediction.

4. The adaptive failure prediction method for motors based on digital twin fusion intelligent optimization according to claim 3, characterized in that: Step S4: Construct a physical information neural network and automatically learn the time-varying coupling coefficient between state variables and multiphysics fields by embedding a hybrid loss function of physical control equations, boundary conditions and causal constraints. This includes the following steps: Step S4.1: Construct the input layer variables of the physical information neural network to characterize the multiphysics operating conditions of the system in the time dimension. The input vector is defined as: ； in, For temperature parameters, For vibration stress parameters, For current parameters, This refers to the ambient relative humidity parameter. For input voltage parameters, It is a time variable; Step S4.2: Construct the output layer variables of the neural network to simultaneously characterize the system state evolution process and the coupling relationship between multiple physics fields. The output vector is defined as: ； in, , , , These represent damage variables, crack length, contact resistance, and penetration depth, respectively. , , The time-varying coupling coefficients to be learned are used to describe the dynamic interaction strength between different physical fields; Step S4.3: Construct a hybrid fully connected neural network, the network structure of which is as follows: ； Specifically, the Sigmoid activation function is used for the output nodes of state variables to constrain their physical value range, and a linear mapping method is used for the output nodes of coupling coefficients, thereby realizing continuous modeling of time-varying parameters. Step S4.4: Construct a joint loss function by introducing data constraints, physical constraints, boundary constraints, and causal consistency constraints: ； in, , , and These are the weight coefficients for each loss term, used to adjust the relative importance of different constraints during the training process; Step S4.5: Introduce supervised fitting loss to the state variables with measured data, which is defined as: ； in, The data constraint loss term, also known as the supervised fitting loss function, improves the model's ability to fit the real system behavior by minimizing the squared error between the predicted and measured values. This represents the total number of samples for the measured data. The predicted values ​​are for the damage variables at the magnet-casing bonding interface; The measured values ​​represent the damage variables at the magnet-casing bonding interface. The predicted value for the crack length at the potting compound-winding / core interface; This is the measured value of the crack length at the potting compound-winding / core interface; Step S4.6: Embed the physical control equations, including damage evolution, crack propagation, wear, and diffusion, into the loss function in residual form to construct the physical constraint loss: ； The collocation points are generated in the spatiotemporal domain using the Latin hypercube sampling method to ensure that the physical constraints are effective throughout the entire domain. Loss due to physical constraints; This represents the total number of points; Let D be the derivative of the damage variable with respect to time. The physical model function for damage evolution depends on: Current damage variable, Surface temperature of the magnet Equivalent cyclic stress amplitude; Let be the derivative of the crack length 'a' with respect to the number of load cycles 'N'; For the physical model of crack propagation, Current crack length, Effective stress intensity factor range; The time derivative of the wear amount W; This is a wear-corrosion physical model, corresponding to the extended Arcard's law in step S2.2, where W is the cumulative wear volume and I is the contact current. For dynamic contact resistance; The time partial derivative of the concentration C of the permeating medium; The effective diffusion coefficient; The spatial second partial derivative of concentration C; Step S4.7: Introduce initial state and boundary condition constraints of the system, and construct a boundary loss term to ensure that the model prediction results satisfy the known physical conditions at the initial time and boundary position: ； in, Loss is due to boundary condition constraints; The values ​​of the damage variables at the initial time t=0, as predicted by the neural network; The initial crack length predicted by the neural network; The concentration of the medium on the outer surface of the sealing layer is predicted by the neural network; 0 indicates that the damage to the brand-new motor is zero, and it is in a damage-free state. The initial concentration of the medium within the entire sealing layer as predicted by the neural network; Step S4.8: Introduce causal consistency loss to constrain the network's learning results from violating known physical causal relationships. Its definition is: ； in, The total loss for causal consistency; The total number of causal constraint rules; The penalty for the thermo-mechanical coupling coefficient under low-temperature conditions is: ； in, It is the thermo-mechanical coupling coefficient; For temperature, Temperature threshold; Indicates the current stress level; The causal relationship between penetration depth and internal state is as follows: ； in, For internal physical quantities; This refers to the penetration depth of the sealing layer. This indicates the sensitivity of internal quantities to penetration depth; The damage evolution constraint under low stress conditions is the penetration depth of the sealing layer: ； Where D is the damage variable; This represents the rate of damage accumulation. This represents the equivalent cyclic stress amplitude. The stress threshold; It is a tiny positive number; The arc erosion penalty under low current conditions is: ； in, The cumulative material loss caused by arc erosion; Arc erosion rate; For current; The current threshold; Finally, the monotonicity constraint of the temperature-diffusion coefficient is defined as follows: ； in, The effective diffusion coefficient; This is the partial derivative of the diffusion coefficient with respect to temperature; The above constraints help avoid generating unreasonable strong coupling relationships under low-excitation conditions; Step S4.9: Optimize the model according to the phased training strategy, initially using only... and Perform physical pre-training, and then introduce Make minor adjustments to the data, and finally add... The L-BFGS optimization algorithm is used for fine convergence, and Dropout, gradient pruning and early stopping strategies are combined to improve the stability and generalization performance of model training. Step S4.10: After the model training is complete, the multiphysics time-varying coupling coefficient is directly obtained from the neural network output, and its expression is: ;in, is the thermo-mechanical coupling coefficient, which characterizes the coupling strength between the temperature field and the mechanical field; The function mapping relationship learned by the neural network; The vibration stress field specifically refers to the equivalent stress amplitude caused by vibration. The chemical-mechanical coupling coefficient characterizes the synergistic effect between chemical corrosion and mechanical damage. This refers to the concentration of corrosive ions. Plastic strain rate, characterizing the deformation rate of a material; is the electro-thermal coupling coefficient, which characterizes the coupling between current effects and thermal effects; These are contact current, dynamic contact resistance, reflecting contact quality, and temperature gradient, respectively, characterizing the non-uniformity of heat flow distribution.

5. The adaptive failure prediction method for motors based on digital twin fusion intelligent optimization according to claim 4, characterized in that: Step S5: Based on the particle swarm optimization algorithm, the failure chain weights are dynamically reconstructed using real-time monitoring data to identify the dominant failure mode, including the following process: Step S5.1: Model the dynamic reconstruction problem of the failure chain as a multivariate optimization problem. Its optimization objective is to find the failure chain weight allocation scheme that best explains the current degradation state of the system under the given monitoring data. Step S5.2: Define the spatial position of each particle in the particle swarm as the failure chain weight vector: ； in, ) indicates the interface Independent degradation weights, Represents the independent degradation weight of the sealed interface; This represents the independent degradation weight of the potting interface. These represent the independent degradation weights of the carbon brush and magnet interface, respectively. Display Interface To the interface The coupling propagation strength is determined, and a normalization constraint is applied. To ensure the physical interpretability of failure contribution allocation; Step S5.3: Construct the fitness function for particle evaluation, which is defined as: ； The prediction error term is defined as follows: ; Configure based on current weights Predicted remaining lifespan; Lifetime estimated by back-calculation of monitoring data; physical consistency penalty. The model complexity term is defined to quantify the degree of violation of physical constraints. To suppress unnecessary coupling paths; Step S5.4: At the initial time Randomly generate a particle swarm, with the number of particles set to [number]. The initial position of the particle satisfies , Let be the initial position of the i-th particle. A uniform distribution between 0 and 1. Represents a 10-dimensional hypercube space; particle velocities are initialized to... , Let be the initial velocity of the i-th particle; Indicates the maximum speed limit; further initializes the individual optimal solution. Global optimal solution Defined as: ； Step S5.5: Update the particle velocity at the (t+1)th iteration, using the following formula: ； in, Inertial weights; The current particle velocity; For individual learning factors; for Random numbers between; As a social learning factor; The historical best position of the i-th particle; This represents the current position of the i-th particle; The globally optimal position; The physical guiding coefficient; For physical gradient guiding terms; Then speed limit is applied. And perform location updates , This represents the truncation function, which will... Limited to between; Step S5.6: Trim the boundaries of the updated particle positions: ； Further normalization is performed on the independent degradation weights to satisfy the physical constraints: ； in, The first four components of the particle position vector are: 1-sealing, 2-encapsulation, 3-carbon brush, and 4-magnet. Step S5.7: Introduce the physical gradient guiding term The penalty function is as follows: Includes: when sealing weight Larger but monitoring penetration depth Smaller penalty items, when filling weight Larger crack propagation rate Penalty term for slow changes, and when the coupled path weights An infinite penalty term is imposed when the physical causal order is violated, in order to guide the particle to converge toward the physically plausible region; Step S5.8: When the cumulative running time reaches a preset threshold, the prediction deviation exceeds 20%, or the operating conditions change significantly, the dynamic reconstruction process is triggered. The particle swarm optimization algorithm is run by extracting the most recent time window data, and the updated failure chain weight matrix is ​​output. The evolution history of the weights is recorded for subsequent analysis.

6. The adaptive failure prediction method for motors based on digital twin fusion intelligent optimization according to claim 5, characterized in that: Step S6: Establish a three-level mapping model of material-component-system, mapping the interface damage state to system-level observable performance indicators, including the following steps: Step S6.1: Definition of material layer state parameters; Based on the modeling results of steps S2 and S3, the damage state quantities of each weak interface are defined as the set of material layer input parameters, including: damage variables of the magnet-casing bonding interface. Encapsulant – Crack length at the winding / core interface Contact resistance of the carbon brush-commutator contact interface With wear Penetration depth of the sealant bonding interface The above parameters characterize the degradation state of the material layer under the action of multiple physics fields, and serve as the basic input for subsequent mapping models. Step S6.2: Component layer functional degradation modeling; mapping material layer damage parameters to component layer functional indicators, establishing the influence relationship of interface damage on the performance of key components, including: through magnet bonding damage A mapping model between magnetic flux density non-uniformity and torque pulsation amplitude was established based on the relationship with magnetic circuit offset; the crack length at the potting interface was used as a reference. The influence of winding thermal resistance and insulation performance was investigated, and a mapping relationship between winding loss and efficiency decay was established; this was achieved through carbon brush contact resistance. By varying the spark frequency, a mapping relationship between brush commutation quality and electromagnetic interference intensity was established; the penetration depth at the sealing interface was also investigated. A model is established to assess the impact of environmental media intrusion on component corrosion and electrical performance degradation in relation to changes in internal humidity. The above mapping relationship can be achieved using a semi-empirical model, lookup table function, or a calibrated simplified physical model to ensure computational efficiency and engineering feasibility. Step S6.3: System-level performance index mapping; Based on the component-level model, further construct the system-level performance output model to map the component functional degradation to system-level observable performance indices, including: motor output torque fluctuation amplitude. System efficiency The downward trend; electromagnetic interference intensity or equivalent noise index This three-level mapping model enables the transformation from "interface damage state that cannot be directly observed" to "system performance indicators that can be monitored in real time," providing a unified benchmark for subsequent prediction verification and model correction.

7. The adaptive failure prediction method for motors based on digital twin fusion intelligent optimization according to claim 6, characterized in that: Step S7: Construct a digital twin model integrating PINN and PSO weights, and use Bayesian inference to update the model parameters online to achieve quantified propagation of uncertainty and probabilistic lifetime prediction, including the following steps: Step S7.1: Construct a digital twin system consisting of a physical entity layer, a twin model layer, a communication layer, and a decision layer. The twin model layer integrates the PINN multi-field model and the PSO failure chain weights to reflect the system state evolution in real time. Step S7.2: Model the uncertainty of the model parameters as a Bayesian inference problem. Based on existing historical knowledge, update the uncertainty of the model parameters using newly acquired data. The posterior distribution is defined as: ； in, Represents the set of parameters to be estimated; This is a historical observation dataset; This refers to newly acquired online monitoring data; It is the prior distribution; To observe the likelihood function; This is evidence. Step S7.3: Set the prior distribution of the parameters as follows: ； in, The prior mean; The prior variance reflects the uncertainty of the parameters; The observation likelihood function is defined as: ； in, For the first One observation response; For the prediction output of the data-driven model, For the corresponding input variables; To observe the noise variance; Step S7.4: Use the Markov chain Monte Carlo method to sample the posterior distribution of the parameters to obtain the expected value of the parameters and the uncertainty measure, which will be used for subsequent prediction and propagation analysis; Step S7.5: Calculate the lifetime prediction uncertainty based on the posterior distribution of the parameters: ； in, Indicates the remaining lifespan of the motor. For monitoring data / performance observation datasets; The set of parameters for the model to be estimated; It is the posterior distribution of parameters obtained by the Markov chain Monte Carlo method; In a given The lifetime condition distribution under parametric conditions is obtained by joint simulation of the failure mechanism model, failure propagation network, and PINN coupling term; Step S7.6: Use the Monte Carlo method to perform uncertainty propagation and statistics to verify the numerical estimation of the integral in step S7.5; First, posterior sampling is performed, drawing samples from the posterior distribution of the parameters. Group samples Then, lifetime calculation is performed for each set of parameter samples. The digital twin model is driven to extrapolate the failure evolution into the future, using the first arrival time of "reaching the failure threshold" as the lifetime output; finally, the expected lifetime and interval estimation are performed, the expected lifetime is calculated from the sample set, and the interval estimation is given by quantiles.

8. The adaptive failure prediction method for motors based on digital twin fusion intelligent optimization according to claim 7, characterized in that: Step S8: By comparing the predicted results with the measured data in a closed loop, adaptively triggering model parameter updates and failure chain reconstruction, a continuously self-evolving prediction-verification-correction mechanism is formed, including the following steps: Step S8.1: In each prediction period Within the digital twin model based on step S7, using the monitoring data at the current moment and the updated model parameters, the system-level performance indicators and remaining lifetime prediction results are generated, including: system performance prediction values, probability distribution of remaining lifetime and confidence intervals. Step S8.2: Compare the predicted system-level performance indicators with the actual monitored values, and calculate the prediction error indicators, including mean square error, relative error or probability coverage, to quantify the model prediction accuracy. Step S8.3: Based on the preset error threshold or operating condition change criterion, determine whether to trigger the model self-correction process; the self-evolution mechanism is triggered when any of the following conditions are met: The prediction error exceeds the set threshold for multiple consecutive periods; The confidence interval for predicted remaining lifespan deviates significantly from the measured degradation trend. The operating conditions have changed significantly, causing the original failure mode assumptions to no longer hold. Step S8.4: When the triggering condition is met, automatically perform the following operations: call the Bayesian inference method in step S7 to update the posterior distribution of the key model parameters online; call the particle swarm optimization algorithm in step S5 to dynamically reconstruct the failure chain weights and adjust the dominant failure path and its propagation intensity. Step S8.5: Synchronize the updated model parameters and failure chain structure to the digital twin model and proceed to the next prediction cycle. This forms a closed-loop self-evolutionary operation mode of "prediction-verification-correction-re-prediction", enabling the motor failure prediction model to continuously adapt and improve its performance throughout the entire service life.

Images