Data model dual-drive motor life prediction system and method thereof

Abstract

The invention discloses a data model dual-drive motor life prediction system and method. The system comprises an information acquisition module, a model fusion module, a model training module and a weight adjustment module. The information acquisition module is used for acquiring state data of different motors; the model fusion module is used for effectively fusing the data driving model and the physical information model; the model training module is used for training and optimizing the fused mechanism-data dual-drive model according to the motor state data provided by the information acquisition module; the weight adjusting module plays a key adjusting role in the whole model training process. The method not only depends on a data-driven model for prediction, but also combines a physical information model to make up for the problem that a pure data-driven method lacks interpretability. Multi-physical-field factors such as environmental stress, vibration impact, thermal management and an electromagnetic field are brought into modeling, so that the performance degradation of the motor is accurately described and evaluated, and the accuracy and physical interpretability of a prediction result are enhanced.

Description

Technical Field

[0001] The present invention belongs to the fields of Internet of Things, industrial automation and intelligent maintenance, and specifically relates to a motor life prediction system and method driven by a data model. Background Art

[0002] Special vehicle operations are subject to multiple sources of stimulation, including harsh environments (such as extreme cold, high temperatures, and rough road surfaces) and complex tasks (such as extreme operating conditions and frequent start-stop / load switching). The service state of electric drive systems is affected by the coupled effects of multiple physical fields, including mechanical, electrical, thermal, and magnetic factors. Consequently, research on service performance degradation mechanisms has shifted from traditional mechanical damage to a multi-faceted, integrated "mechanical-electrical-thermal-magnetic" damage model. As a core component of electric drive systems, the performance degradation and life evolution of the drive motor are attracting increasing attention. Existing research focuses on motor failure mode analysis, component failure mechanism analysis, life distribution types, and reliability models. However, the evolution of the drive motor's health under multi-field coupling remains unclear. Although relevant studies have demonstrated that thermal stress in the temperature field can lead to multiple forms of damage, including demagnetization of permanent magnets in the magnetic field, damage to the insulation layer of the electric winding, and deformation of the mechanical structure, the complexity of the degradation mechanism characterization model increases exponentially with the increasing number of multi-physics factors. To reduce modeling complexity, idealized assumptions about specific conditions are unavoidable, making it difficult to fully represent actual driving conditions. Furthermore, motor performance degradation is closely related to multiple stress amplitudes, frequencies, and superposition sequences. Mechanism-based modeling methods lack the ability to characterize complex mapping relationships, making it difficult to accurately reflect performance degradation under actual operating conditions.

[0003] With the development of artificial intelligence and big data technologies, data-driven approaches have become cutting-edge technologies for health state estimation and remaining life prediction. Data-driven approaches do not rely on modeling the intrinsic mechanisms of service performance degradation. Instead, they use machine learning to explore life evolution patterns based on aging test data or historical operating condition data. Deep learning, with its strong nonlinear mapping capabilities, has become the mainstream of data-driven approaches. Based on the Golmogorov judgment, it can theoretically fit arbitrary continuous functions, providing the potential for in-depth exploration of degradation mechanisms driven by the coupling of multiple physical fields under multiple environments and tasks. However, deep learning uses gradient descent for model iteration. Its random initialization and non-convex optimization lead to uncontrollable parameter optimization processes. Furthermore, its "black box" nature makes the model lacking in interpretability, making it difficult to quantify the effects of specific influencing factors on degradation trends, comparable to model-based approaches. In recent years, research has shifted from improving prediction accuracy to elucidating degradation mechanisms. However, explicit feature clustering at the data level cannot fully explore the underlying patterns of life evolution under the influence of higher-order influencing factors.

[0004] In addition, the training of data-driven methods relies on a large amount of high-quality data. The equivalent damage and accelerated aging data collected on the laboratory bench do not match the degradation trend of the service performance in the actual complex battlefield environment. The uncertainties in the actual working conditions make it difficult to meet the prior assumption that the training set and test set of neural network convergence are independently and identically distributed, which easily leads to overfitting and thus limits the generalization ability in actual working conditions. To address this issue, countries have established a simulation environment test standard system and related standards for armored vehicles. However, it is still difficult for simulation tests and reinforcement tests to fully cover the spatio-temporal coupling characteristics of service performance degradation factors in service scenarios and usage profiles. Therefore, a large number of studies have attempted to explore data-model fusion methods, either by using mechanism models to supplement the lack of data samples in extreme scenarios or by integrating data-driven and mechanism models through confidence-weighted methods. However, their essence still remains at the superficial level of data fusion and model superposition, not only failing to fully utilize the advantages of mechanism models and data-driven approaches, but also bringing about error accumulation and reducing the deduction accuracy. Summary of the Invention

[0005] The present invention provides a method for predicting the motor life driven by both data and model to solve the defects in the prior art.

[0006] The present invention is achieved through the following technical solutions:

[0007] A motor life prediction system driven by both data and model includes an information acquisition module, a model fusion module, a model training module, and a weight adjustment module;

[0008] The information acquisition module is used to collect the state data of different motors and comprehensively analyze and process these data;

[0009] The model fusion module is used to effectively fuse the data-driven model and the physical information model;

[0010] The model training module is used to train and optimize the fused mechanism-data dual-driven model according to the motor state data provided by the information acquisition module;

[0011] The weight adjustment module plays a key regulatory role throughout the model training process.

[0012] A method for predicting the motor life driven by both data and model, characterized in that:

[0013] Step 1: The information acquisition module collects the signals generated during the operation of the motor in real time through various sensors deployed on the motor and its surrounding environment;

[0014] Step 2: The information acquisition module performs preliminary processing and analysis on various types of raw signals collected, including signal denoising, distortion correction, and outlier detection operations. At the same time, common statistical features (such as mean, variance, peak value, kurtosis, impact factor, waveform factor, etc.) are extracted or calculated;

[0015] Step 3: Transmit the processed data to the training after the adjustment of the weight coefficients of the loss function and model fusion;

[0016] Step 4: Integrate the physical model and the data-driven model to obtain a mechanism-data dual-driven model;

[0017] Step 5: For the training process of multi-objective and multi-loss functions, dynamically adjust the weight coefficient λ of different loss terms through an adaptive method; this adjustment mechanism is usually based on information such as gradient statistics and error distribution, so that the model can balance the physical constraints and data fitting effect while maintaining control over numerical stability;

[0018] Step 6: Put the mechanism-data dual-driven model that has completed training or optimization into practical applications to predict the motor health status, remaining life, or key performance degradation trend.

[0019] A method for predicting the motor life with dual drive of data models as described above, the construction of the mechanism-data dual-driven model includes the following steps:

[0020] Step 1: Use deep learning to train a "black box" model based on a large amount of collected motor operation data; this model can capture complex non-linear relationships and provide high-precision prediction capabilities, but its internal mechanism usually lacks physical interpretability;

[0021] Step 2: Embed the physical mechanism of motor degradation into the model; since it is difficult to fully obtain explicit physical equations, a generalized function approximator is used to represent complex non-linear dynamic processes to form a physical information neural network;

[0022] Step 3: Through the function approximator, fuse the physical information with the data-driven model to form a deeply hidden physical model;

[0023] Step 4: During the training process, the model simultaneously considers the regression error of the data-driven and the constraints of the physical information model, and through the method of joint optimization, realizes the deep integration of mechanism and data.

[0024] A method for predicting the motor life with dual drive of data models as described above, the construction method of the loss function includes the following steps:

[0025] Step (1): Construct a test plan and collect test data, and analyze and process the collected data; the data should be targeted at the external excitation and environmental stress changes in the special vehicle mission profile and service environment, focusing on data in typical environmental scenarios such as high temperature, low temperature, dust, mud, snow, etc. and multi-scenarios of usage tasks such as low-temperature start-up and load impact;

[0026] Step (2): When the vehicle is driving under complex road conditions, road excitation and on-vehicle load impact will generate significant vibration responses on the drive motor. In this context, focus on collecting relevant data of the drive motor under the action of the electromagnetic field to deeply analyze the impact of the vibration response on the degradation of the motor service performance;

[0027] Step (3): Regarding the influence of the outdoor environmental temperature on the heat dissipation process of the drive motor, calculate the power loss during the motor operation and the heat generation caused by the electromagnetic field magnetocaloric effect, establish a temperature field model to estimate the change trend of the temperature rise of the drive motor at a specific environmental temperature, and then calculate the temperature distribution of the insulation winding and the permanent magnet, and estimate the influence degree of the environmental temperature excitation on the performance degradation of the electrical conductivity of the insulation layer and the remanent magnetic density of the permanent magnet;

[0028] Step (4): According to the motor service performance degradation data, extract the key parameters that dominate the service performance degradation, obtain the historical data of the service performance degradation, and construct a data set.

[0029] A method for predicting the motor life driven by a dual data model as described above, the method for predicting the health status of the motor is as follows:

[0030] Step 1): The motor decline is affected by multiple factors such as heat management and temperature rise, wear and fatigue of mechanical components, electrical stress and insulation aging, stability and protection of permanent magnet materials, environmental and working conditions, etc., so it is modeled as a multivariate function;

[0031] Step 2): According to the model describing the motor decline mechanism, combined with the self-learning ability of the neural network, define a neural network with physical information embedding of motor aging; to balance accuracy and computational complexity, only consider the influence of the first derivative;

[0032] Step 3): The motor decline dynamics model requires a continuous description of the health state. Although the output of the neural network can be approximately continuous, due to the fact that data collection is usually discrete, it is difficult to fully meet the continuity assumption of the dynamics model. On this basis, a calculation formula suitable for describing the health state of the motor is derived through mathematical derivation;

[0033] Step 4): To reduce the training error, train the model by minimizing the mean square error loss, pay attention to the loss function, and propose an adaptive weight coefficient adjustment method to finally obtain the health status of the motor.

[0034] A method for predicting the service life of an electric motor driven by a dual data model as described above, wherein the multivariate function is:

[0035] u = f(t, x) (13)

[0036] where t represents time and x represents a vector composed of temperature, motor current, motor voltage, magnetic flux density, and other key factors.

[0037] A method for predicting the service life of an electric motor driven by a dual data model as described above, wherein the construction method of a neural network that describes motor aging and has learnability by fusing a physical model and a data-driven model includes the following steps:

[0038] Step a: Set the partial differential equation to be parameterized by θ. To describe the degradation dynamics of the motor, its decay rate is

[0039]

[0040] In the formula, g(g) represents a nonlinear function of t, x, and u. The function g(g) characterizes the degradation dynamics inside the motor. By changing this nonlinear function, various forms of degradation can be characterized. When only considering time, it can be regarded as a special case of formula (2);

[0041] Step b: Under the conditions of step a, since the explicit form of g(g) is unknown, a more generalized function approximator g'(g) with parameter θ' is used to represent the nonlinear dynamics g(g). Therefore, equation (2) becomes:

[0042] u , ,

[0043] , xx , ,

[0042] , xx ,

[0045] ,

[0041] , t ,

[0044] ,

[0040] , ,

[0039] , , t , , x ,

[0046] , , t , , , , ≈ g'(t, ×, u, u t , u x , u xx ,...; θ') (15)

[0043] In the formula, Use a neural network F(t, x; Φ) with learnable parameters to model f(t, x) and calculate u using automatic differentiation t , represents the first-order partial derivative of u with respect to x, and u xx represents the second-order partial derivative of u with respect to x;

[0044] Step c: Under the conditions of satisfying step b, instead of specifying a candidate basis function set for g(g), a more generalized function approximator g'(g) is used. To balance accuracy and computational complexity, only the influence of the first-order partial derivative is considered, and higher-order derivatives are discarded. Therefore, the physics-informed neural network is:

[0045]

[0046] In the formula, denotes the partial derivative of the neural network F(g) with respect to t, and G(g) represents the motor degradation dynamics equation modeled by the neural network.

[0047] A method for predicting the remaining useful life of a motor driven by a dual data model as described above, the calculation formula for the motor health status:

[0048]

[0049] In the formula, Q m represents the life cycle of cycle m, and Q 0 represents the nominal life cycle. When t = m, the health status value u m coincides with the point on the degradation trajectory f(g).

[0050] A method for predicting the remaining useful life of a motor driven by a dual data model as described above, the physics-informed neural network and its dynamic model G(g) can be trained by minimizing the mean squared error loss, and its training formula is:

[0051]

[0052] where

[0053]

[0054] In the formula, equation (6) represents multi-task learning, and the loss items in the training process can be balanced by adjusting the weight coefficients λ u , λ f and ; L u corresponds to the regression fitting error at the observed values collected, and L f and strengthen the structure of the physics-informed neural network imposed by equation (3) and its first-order partial derivative with respect to t; here represents the training data, and u i represents the predicted label corresponding to the input data {x i , t i};

[0055] Taking the output corresponding to the loss L u as an example, other objectives in equation (6) can be defined accordingly, and the observation noise scalar is related to the weight coefficient λ corresponding to the output in the loss function, and the expression is as follows:

[0056]

[0057] Based on equations (7), (8), (9), and (10), the regularization process of the loss function:

[0058]

[0059] The loss function formula (6) is regularized as:

[0060]

[0061] Through the above operations, the relative weight coefficients λ u , λ f and

[0062] The advantages of the present invention are:

[0063] The present invention not only relies on a data-driven model for prediction, but also combines a physical information model to make up for the lack of interpretability of pure data-driven methods. By incorporating multi-physical field factors such as environmental stress, vibration shock, thermal management, and electromagnetic fields into the modeling, accurate characterization and evaluation of the motor performance degradation are achieved, enhancing the accuracy and physical interpretability of the prediction results;

[0064] The present invention has high adaptability and generality. A more generalized non-linear function approximator (such as a neural network) is used to represent complex degradation dynamics, avoiding excessive dependence on motor physical formulas. Since there is no need to specify a candidate basis function set for the non-linear function, this method can adapt to the multi-field coupling degradation process of motors of different types and under different working conditions, providing a general solution for the life evolution assessment of multi-task and multi-environment motor systems;

[0065] The present invention takes into account both accuracy and computational efficiency. By only considering the first-order partial derivatives of the dynamic equation, a physics-informed neural network (PINN) is constructed, effectively balancing accuracy and computational complexity. While satisfying the core physical constraints, it reduces the training overhead brought by higher-order derivatives, reduces the risk of numerical decoupling and gradient imbalance, and improves the feasibility and robustness of the prediction model;

[0066] During the training process of the present invention, when minimizing the mean square error loss for multi-task and multi-loss terms, the present invention proposes to automatically and adaptively adjust the weight coefficients of each loss in the network. By using gradient statistics to dynamically adjust the weights, the workload and uncertainty of manual parameter tuning are reduced, and comprehensive and multi-objective optimization of the health state prediction and degradation dynamic equation is achieved, making the model more adaptable to complex environmental excitations. Description of the Drawings

[0067] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0068] Figure 1 It is a schematic diagram of the physical information neural network proposed by the present invention;

[0069] Figure 2 It is an architecture diagram of the control method for the degradation mechanism of the service performance of a drive motor under the coupling action of multiple physical fields and multiple factors of the present invention. Specific embodiments

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts belong to the scope of protection of the present invention.

[0071] The research of the present invention is based on the following assumptions as premises:

[0072] 1) The motor is a permanent magnet synchronous motor mostly used in complex environments; 2) Assume that the degradation process of the drive motor under the excitation of complex environments and multi-physical fields (such as thermal fields, electromagnetic fields, mechanical fields, etc.) can be decomposed into several relatively independent and mutually coupled factors (such as temperature, vibration, load impact, electrical stress, insulation aging, environmental dust, etc.), and data can be observed or obtained through certain testing and sensor monitoring means; 2) The degradation trajectory of the motor shows a certain continuity and differentiability under the combined action of time and various external environmental factors, and can be approximated or described by appropriate functions (such as partial differential equations, neural networks). Especially when the degradation is modeled as a dynamic equation, it is assumed that the evolution of the motor health state is relatively smooth and has a certain physical interpretability; 3) Only considering the first derivative (or some low-order derivatives) can better characterize the main degradation trends and mechanisms of the motor during service; the influence of higher-order derivatives can be ignored within an acceptable range, so as to achieve a balance between model accuracy and computational complexity. 4) There are deficiencies in purely physical mechanism-based or pure data-driven approaches: physical models are difficult to obtain a complete and accurate explicit model in practical applications due to the changing environmental scenarios and unclear mechanisms; data-driven models can fit complex non-linear relationships, but lack the interpretation and constraints of actual physical meanings; 5) Assume that the collected or simulated environmental stresses (high temperature, low temperature, dust, mud, snow, impact load, low temperature start, etc.) can represent the main excitation characteristics encountered by the motor in the actual service scenario, and the collected data samples have a certain coverage and credibility, which can provide sufficient support for model training; 6) Assume that a reasonable health index and other key characterization parameters (such as current, voltage, temperature, vibration signal, etc.) can be constructed or selected, and they are incorporated into a multivariate function together with time t and environmental stress parameters, so as to quantify the health status, remaining life or performance degradation degree of the motor.

[0073] A motor life prediction system driven by a dual data model includes an information collection module, a model fusion module, a model training module, and a weight adjustment module;

[0074] The described information acquisition module is used to collect the status data of different motors and comprehensively analyze and process this data; specifically, this module real-time monitors the operating parameters of the motors through a variety of sensors, including key indicators such as motor current, motor speed, and error signals. The collected data information not only covers basic operating parameters but also includes various statistical features such as mean, variance, peak value, kurtosis, peak factor, impact factor, and waveform factor. These statistical features can deeply reflect the operating status and potential performance changes of the motors under different working conditions. In addition, the information acquisition module is also responsible for recording the serial number and timestamp of each data acquisition to ensure that the data of different motors at different times can be accurately traced and correlated, thereby providing a reliable basis for subsequent data analysis and modeling;

[0075] The described model fusion module is used to effectively fuse the data-driven model with the physical information model. Although the traditional pure data-driven "black box" model has advantages in prediction accuracy, its internal mechanism lacks transparency and physical interpretability, which limits its trust and application scope in practical engineering applications. To make up for this deficiency, the model fusion module introduces the physical information model and improves the interpretability and reliability of the overall model by fusing physical constraints in the model. However, the motor degradation process involves complex physical phenomena such as thermal management, electromagnetic field effects, and mechanical vibrations, and the explicit physical formulas used to accurately describe these phenomena are very limited and difficult to obtain. Therefore, the present invention uses a more generalized function approximator to represent the nonlinear dynamic process of motor degradation, thereby constructing a deeply hidden physical model. This method can not only flexibly adapt to various complex degradation mechanisms but also effectively fuse physical information with data-driven methods in the absence of complete physical formulas, achieving the training results of mechanism-data dual drive, significantly improving the prediction performance and physical consistency of the model.

[0076] The described model training module is used to train and optimize the fused mechanism-data dual drive model according to the motor status data provided by the information acquisition module. Specifically, this module uses the root mean square error (RMSE) and the root mean square percentage error (RMSPE) as the main measurement indicators to evaluate the performance of the model in predicting the motor health status and remaining useful life. By minimizing these error indicators, the model training module can continuously adjust and optimize the model parameters to improve the accuracy and robustness of the prediction. In addition, the model training module also needs to comprehensively consider the fitting ability of the data-driven model and the constraint requirements of the physical information model to ensure that the final model not only performs excellently in data fitting but also is physically reasonable and consistent;

[0077] The described weight adjustment module plays a crucial regulatory role throughout the model training process; since the Physics-Informed Neural Network (PINN) needs to handle regression tasks from the dataset while satisfying the given partial differential equation (PDE), the model is very sensitive to the relative weights of different objective functions during training. Manually adjusting these weight coefficients is not only time-consuming and laborious but also difficult to find the optimal weight combination in multi-task learning. Especially when facing complex numerical stiffness problems, the gradient imbalance during the backpropagation process further increases the difficulty of weight adjustment. To overcome these challenges, the present invention designs an adaptive weight adjustment method based on gradient statistics. This method dynamically adjusts the weight coefficients of each loss term by monitoring the gradient changes of each loss term in real-time and according to statistical characteristics such as the mean and variance of the gradient, thereby automatically balancing the physical constraints and data fitting effect during training. This not only significantly reduces the workload and uncertainty of manual parameter tuning but also effectively alleviates the gradient imbalance problem caused by numerical stiffness, ensuring that the model has stronger adaptability and stability in complex environments.

[0078] As Figure 1 and Figure 2 shown, a method for predicting the motor life driven by a data model dual-drive includes the following steps:

[0079] Step 1: The information acquisition module collects the signals generated during the operation of the motor in real-time through various sensors deployed on the motor and its surrounding environment; the collected content includes motor current, motor speed, temperature, vibration, noise, and other indicators related to the health status of the motor, providing a necessary source of raw data for subsequent data preprocessing and modeling.

[0080] Step 2: The information acquisition module performs preliminary processing and analysis on the collected various raw signals, including signal denoising, distortion correction, and outlier detection operations. At the same time, common statistical features (such as mean, variance, peak value, kurtosis, impact factor, waveform factor, etc.) are extracted or calculated.

[0081] Step 3: Transmit the processed data to the training after the loss function weight coefficient adjustment and model fusion.

[0082] Step 4: Integrate the physical model and the data-driven model to obtain a mechanism-data dual-drive model; although traditional data-driven models have strong non-linear fitting capabilities, their interpretability is relatively insufficient; while pure physical models often have difficulty fully depicting the degradation process of the motor in complex environments. By integrating the two, it is possible to ensure a certain degree of physical interpretability while achieving a high prediction accuracy, forming a hybrid intelligent model in which the deep hidden physical model and the data-driven model work together.

[0083] Step 5: For the training process with multiple objectives and multiple loss functions, dynamically adjust the weight coefficients λ of different loss terms through an adaptive method; this adjustment mechanism is usually based on information such as gradient statistics and error distribution, enabling the model to balance physical constraints and data fitting effects while maintaining control over numerical stability;

[0084] Step 6: Put the mechanism-data dual-driven model after training or optimization into practical applications to predict the health status, remaining useful life, or key performance degradation trend of the motor.

[0085] Specifically, the construction of the mechanism-data dual-driven model described in this embodiment includes the following steps:

[0086] Step 1: Use deep learning to train a "black box" model based on a large amount of collected motor operation data; this model can capture complex non-linear relationships and provide high-precision prediction capabilities, but its internal mechanism usually lacks physical interpretability;

[0087] Step 2: Embed the physical mechanism of motor degradation into the model; since explicit physical equations are difficult to obtain completely, use a generalized function approximator to represent complex non-linear dynamic processes to form a physics-informed neural network;

[0088] Step 3: Through the function approximator, fuse the physical information with the data-driven model to form a deeply hidden physical model;

[0089] Step 4: During the training process, the model simultaneously considers the regression error driven by data and the constraints of the physical information model, and through a joint optimization method, realizes the deep integration of mechanism and data.

[0090] Specifically, the method for constructing the loss function described in this embodiment includes the following steps:

[0091] Step (1): Construct an experimental scheme and collect test data, and analyze and process the collected data; the data should target the external excitation and environmental stress changes in the special vehicle mission profile and service environment, focusing on data in typical environmental scenarios such as high temperature, low temperature, dust, mud, snow, etc. and multi-scenarios of use tasks such as low-temperature start-up and load impact;

[0092] Step (2): When the vehicle is driving on complex road conditions, road excitation and on-vehicle load impact will generate significant vibration responses on the drive motor. In this context, focus on collecting relevant data of the drive motor under the action of electromagnetic fields to deeply analyze the impact of vibration responses on the degradation of the motor service performance;

[0093] Step (3): Regarding the influence of the ambient outdoor temperature on the heat dissipation process of the drive motor, calculate the power loss during the motor operation and the heat generated by magneto-thermal effect of the electromagnetic field, establish a temperature field model to estimate the change of the temperature rise trend of the drive motor at a specific ambient temperature, and then calculate the temperature distribution of the insulation winding and the permanent magnet, and estimate the influence degree of the ambient temperature excitation on the performance degradation of the conductivity of the insulation layer and the remanence density of the permanent magnet;

[0094] Step (4): According to the motor service performance degradation data, extract the key parameters of the service performance degradation to obtain the historical data of the service performance degradation and construct a data set.

[0095] More specifically, the method for predicting the health status of the motor in this embodiment is as follows:

[0096] Step 1): The motor degradation is affected by multiple factors such as heat management and temperature rise, wear and fatigue of mechanical components, electrical stress and insulation aging, stability and protection of permanent magnet materials, environment and working conditions, etc., so it is modeled as a multivariate function;

[0097] Step 2): According to the motor degradation mechanism model, combined with the self-learning ability of the neural network, define a neural network with physical information embedding of motor aging; to balance accuracy and computational complexity, only consider the influence of the first derivative;

[0098] Step 3): The motor degradation dynamics model requires a continuous description of the health state. Although the output of the neural network can be approximately continuous, due to the fact that data acquisition is usually discrete, it is difficult to fully meet the continuity assumption of the dynamics model. On this basis, a calculation formula applicable to describing the health state of the motor is derived through mathematical derivation;

[0099] Step 4): To reduce the training error, train the model by minimizing the mean square error loss, focus on the loss function, and propose an adaptive weight coefficient adjustment method to finally obtain the health status of the motor.

[0100] Further, the multivariate function in this embodiment is: u = f(t, x) (25)

[0101] Where t represents time, and x represents a vector composed of temperature, motor current, motor voltage, magnetic flux density and other key factors.

[0102] Furthermore, the construction method of the neural network for describing motor aging with learnability by fusing the physical model and the data-driven model in this embodiment includes the following steps:

[0103] Step a: Set the partial differential equation to be parameterized by θ. To describe the degradation dynamics of the motor, its decay rate is

[0104]

[0105] In the formula, \(g(g)\) represents a non - linear function of \(t\), \(x\), and \(u\). The function \(g(g)\) characterizes the degradation dynamics inside the motor. By changing this non - linear function, various forms of degradation can be characterized. When only considering time, it can be regarded as a special case of formula (2).

[0106] Step b: Under the conditions of step a, since the explicit form of \(g(g)\) is unknown, a more general function approximator \(g'(g)\) with parameter \(\theta'\) is used to represent the non - linear dynamics \(g(g)\). Therefore, equation (2) becomes:

[0107] u t \(\approx g'(t,\times,u,u t ,u x ,u xx ,...;\theta') (27)

[0108] In the formula, A neural network \(F(t,x;\varPhi)\) with learnable parameters is used to model \(f(t,x)\) and automatic differentiation is used to calculate \(u t , represents the first - order partial derivative of \(u\) with respect to \(x\), and \(u xx represents the second - order partial derivative of \(u\) with respect to \(x\);

[0109] Step c: Under the conditions satisfying step b, instead of specifying a candidate basis function set for \(g(g)\), a more general function approximator \(g'(g)\) is used. To balance accuracy and computational complexity, only the influence of the first - order partial derivative is considered and the higher - order derivatives are discarded. Therefore, the physics - informed neural network is:

[0110]

[0111] In the formula, represents the partial derivative of the neural network \(F(g)\) with respect to \(t\), and \(G(g)\) represents the motor degradation dynamics equation modeled by the neural network.

[0112] Furthermore, the formula for calculating the health status of the motor in this embodiment is:

[0113]

[0114] In the formula, \(Q m represents the life cycle of cycle \(m\), and \(Q 0 represents the nominal life cycle. When \(t = m\), the health status value \(u m coincides with the point on the degradation trajectory \(f(g)\).

[0115] As described above, a method for predicting the motor life driven by a data - model dual - drive, the physics - informed neural network It and its dynamic model G(g) can be trained by minimizing the mean square error loss, and its training formula is:

[0116]

[0117] where

[0118]

[0119] In the formula, Equation (6) represents multi-task learning, and the loss items in the training process can be balanced by adjusting the weight coefficients λ u , λ f and ; L u corresponds to the regression fitting error at the collected observation values, and L f and strengthen the physical information neural network structure imposed by Equation (3) and its first-order partial derivative with respect to t; here represents the training data, and u i represents the predicted label corresponding to the input data {x i , t i};

[0120] Taking the output corresponding to the loss L u as an example, other objectives in Equation (6) can be defined accordingly, and the observation noise scalar is related to the weight coefficient λ corresponding to the output in the loss function, and the expression is as follows:

[0121]

[0122] Based on Equation (7), Equation (8), Equation (9), and Equation (10), the regularization process of the loss function:

[0123]

[0124] The loss function Equation (6) is regularized to:

[0125]

[0126] Through the above operations, the relative weight coefficients λ u , λ f and

[0127] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A motor life prediction system driven by dual data models, characterized in that: It includes an information acquisition module, a model fusion module, a model training module, and a weight adjustment module; The described information acquisition module is used to collect the status data of different motors and conduct comprehensive analysis and processing on this data; The described model fusion module is used to effectively fuse the data-driven model and the physical information model; The described model training module is used to train and optimize the fused mechanism-data dual-driven model according to the motor status data provided by the information acquisition module; The described weight adjustment module plays a key regulatory role throughout the model training process.

2. A method for predicting the service life of a motor driven by a dual data model, characterized in that: It includes the following steps: Step 1: The information acquisition module collects the signals generated during the operation of the motor in real time through various sensors deployed on the motor and its surrounding environment; Step 2: The information acquisition module conducts preliminary processing and analysis on the collected various original signals, including signal denoising, distortion correction, and outlier detection operations. At the same time, common statistical features are extracted or calculated; Step 3: Transmit the processed data to the loss function weight coefficient adjustment neutralization and the training after model fusion; Step 4: Fuse the physical model and the data-driven model to obtain a mechanism-data dual-driven model; Step 5: For the training process with multiple objectives and multiple loss functions, dynamically adjust the weight coefficient λ of different loss terms through an adaptive method; this adjustment mechanism is usually based on information such as gradient statistics and error distribution, enabling the model to balance the physical constraints and data fitting effect while maintaining control over numerical stability; Step 6: Put the trained or optimized mechanism-data dual-driven model into practical application to predict the motor health status, remaining life, or key performance degradation trend.

3. A method for predicting the service life of an electric motor driven by a dual data model according to claim 2, characterized in that: The construction of the described mechanism-data dual-driven model includes the following steps: Step 1: Adopt deep learning to train a "black box" model based on a large amount of collected motor operation data; this model can capture complex nonlinear relationships and provide high-precision prediction capabilities, but its internal mechanism usually lacks physical interpretability; Step 2: Embed the physical mechanism of motor degradation into the model; since it is difficult to fully obtain explicit physical equations, a generalized function approximator is used to represent the complex nonlinear dynamic process to form a physical information neural network; Step 3: Through the function approximator, fuse the physical information and the data-driven model to form a deeply hidden physical model; Step 4: During the training process, the model simultaneously considers the regression error of the data-driven and the constraints of the physical information model, and realizes the deep fusion of mechanism and data through a joint optimization method.

4. A method for predicting the service life of a motor driven by a dual data model according to claim 3, characterized in that: The construction method of the described loss function includes the following steps: Step (1): Construct an experimental plan and collect test data, and conduct analysis and processing on the collected data; the described data should target the external excitation and environmental stress changes in the special vehicle mission profile and service environment, focusing on data in typical environmental scenarios such as high temperature, low temperature, dust, mud, snow, etc. and multiple usage task scenarios such as low-temperature start-up and load impact; Step (2): When the vehicle is driving on complex road conditions, the road excitation and the impact of the upper load will generate significant vibration responses on the drive motor. Under this background, relevant data of the drive motor under the action of the electromagnetic field are mainly collected to deeply analyze the influence of the vibration response on the degradation of the motor service performance; Step (3): Regarding the influence of the outdoor ambient temperature on the heat dissipation process of the drive motor, calculate the power loss during the operation of the motor and the heat generated by the magneto-thermal effect of the electromagnetic field, establish a temperature field model to estimate the change of the temperature rise trend of the drive motor at a specific ambient temperature, and then calculate the temperature distribution of the insulation winding and the permanent magnet, and estimate the degree of influence of the ambient temperature excitation on the performance degradation of the conductivity of the insulation layer and the remanence density of the permanent magnet; Step (4): According to the data of the degradation of the motor service performance, extract the key parameters of the degradation of the table conquest service performance, obtain the historical data of the degradation of the service performance, and construct a data set.

5. A method for predicting the service life of a motor driven by a dual drive of a data model according to claim 2, characterized in that: The prediction method for the health status of the motor is as follows: Step 1): The decline of the motor is affected by multiple factors such as heat management and temperature rise, wear and fatigue of mechanical components, electrical stress and insulation aging, stability and protection of permanent magnet materials, environment and working conditions, etc., so it is modeled as a multivariate function; Step 2): According to the description of the motor degradation mechanism model, combined with the self-learning ability of the neural network, define a neural network with physical information embedding of motor aging; to balance accuracy and computational complexity, only consider the influence of the first derivative; Step 3): The motor degradation dynamics model requires a continuous description of the health state. Although the output of the neural network can be approximately continuous, due to the fact that data acquisition is usually discrete, it is difficult to fully meet the continuity assumption of the dynamics model. On this basis, a calculation formula suitable for describing the health state of the motor is derived through mathematical derivation; Step 4): To reduce the training error, use the mean square error loss minimization to train the model, pay attention to the loss function, and propose an adaptive weight coefficient adjustment method to finally obtain the health status of the motor.

6. A method for predicting the service life of an electric motor driven by a dual data model according to claim 5, characterized in that: The multivariate function is: u = f(t, x) (1) where t represents time, x represents a vector composed of temperature, motor current, motor voltage, magnetic flux density, and other key factors.

7. A method for predicting the service life of a motor driven by a dual data model according to claim 2, characterized in that: The construction method of the neural network for describing motor aging with learnability of the fusion of the physical model and the data-driven model includes the following steps: Step a: Set the explicit partial differential equation to be parameterized by θ. To describe the degradation dynamics of the motor, its decay rate is In the formula, g(g) represents a nonlinear function of t, x, and u. The function g(g) characterizes the degradation dynamics inside the motor. By changing this nonlinear function, various forms of degradation can be characterized. When only considering time, it can be regarded as a special case of formula (2); Step b: Under the conditions of step a, since the explicit form of g(g) is unknown, a more general function approximator g'(g) with parameter θ' is used to represent the nonlinear dynamics g(g). Therefore, equation (2) becomes: u t ≈g'(t,×,u,u t ,u x ,u xx ,...;θ') (3) wherein, using a neural network F(t, x; with learnable parameters Φ) to model f(t,x) and use automatic differentiation to calculate u t , represents the first-order partial derivative of u with respect to x, u xx represents the second-order partial derivative of u with respect to x; Step c: Under the conditions that satisfy step b, instead of specifying a candidate basis function set for g(g), a more general function approximator g'(g) is used. To balance accuracy and computational complexity, only consider the influence of the first partial derivative and discard the higher-order derivatives. Therefore, the physical information neural network is: In the formula, represents the partial derivative of the neural network F(g) with respect to t, and G(g) represents the motor degradation dynamics equation modeled by the neural network.

8. The motor life prediction method based on data and model dual drive according to claim 5, characterized in that: The calculation formula for the health status of the motor: Where Q m represents the life cycle of cycle m, and Q 0 represents the nominal life cycle. When t = m, the health state value u m coincides with the point on the degradation trajectory f(g).

9. The method for predicting the health status of an electric motor in a complex environment according to claim 2, wherein, The described physical information neural network and its dynamic model G(g) can be trained by minimizing the mean square error loss, and its training formula is as follows: Where In the formula, Equation (6) represents multi-task learning, and the loss items of the training process can be balanced by adjusting the weight coefficients λ u , λ f and ; L u corresponds to the regression fitting error at the collected observation values, and L f and strengthen the physics-informed neural network structure imposed by Equation (3) and its first-order partial derivative with respect to t; here represents the training data, and u i represents the predicted label corresponding to the input data {x i , t i}; Take the loss L u As an example of the corresponding output, other objectives in Equation (6) can be defined accordingly. The observation noise scalar is related to the weight coefficient λ corresponding to the output in the loss function, and the expression is as follows: Based on equations (7), (8), (9), and (10), the loss function regularization process: The loss function formula (6) is regularized as follows: Through the above operations, the relative weight coefficients λ u , λ f and

Images