Motor health management method and system based on full life cycle monitoring

Abstract

The invention provides a motor health management method and system based on full life cycle monitoring, and the method comprises the steps: firstly collecting motor multi-source sensor data, and mapping the motor multi-source sensor data into a coordinate point in a low-dimensional health state space through a manifold learning model, so as to represent the comprehensive health state of the motor multi-source sensor data; and on the basis of the expert strategy database, a health cost function associated with the health coordinates is constructed by adopting inverse reinforcement learning and is used for quantifying health loss of different operations. And a decision model is constructed, and the model is trained through an optimal control algorithm, so that an operation strategy capable of minimizing the total health cost of the whole life cycle of the motor can be found out. In actual operation, the system determines current health coordinates in real time, generates an optimal operation strategy by using a decision model in combination with future task requirements, and analyzes the optimal operation strategy into specific control parameters to be issued and executed. According to the method, active health decline path management and closed-loop control of the motor can be realized, so that dynamic optimal balance between the performance and the service life of the motor is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of motor health management, and particularly relates to a motor health management method and system based on whole life cycle monitoring. BACKGROUND

[0002] As the core power equipment of intelligent industry, the health state of a motor is directly related to the efficiency and safety of the entire production system. Therefore, the predictive and health management (PHM) technology emerges as the times require. The mainstream method thereof is state-based monitoring and predictive maintenance. By installing various sensors such as vibration, temperature and current on the motor, a large amount of operation data is collected, and advanced signal processing and artificial intelligence algorithms are used. These systems can effectively identify early fault features and predict the remaining useful life of the motor. However, the existing predictive maintenance technology is still a relatively passive management mode in nature. When the prediction system alarms that the remaining life is insufficient, it is necessary to continue to operate according to the existing mode to ensure the production capacity, but the cost of early scrapping of the equipment needs to be borne, or the load needs to be reduced to sacrifice the yield to extend the life of the equipment. The existing prediction system is difficult to quantify the long-term health impact of different operation strategies on the equipment, so as to automatically seek the optimal balance point in the process of motor health decline, resulting in a large amount of resource waste. SUMMARY

[0003] The application provides a motor health management method and system based on whole life cycle monitoring to solve the above technical problems.

[0004] In a first aspect, the application provides a motor health management method based on whole life cycle monitoring, which comprises the following steps: Collecting multi-source sensor data containing electrical, mechanical and thermodynamic characteristics of the motor during operation, and mapping the multi-source sensor data to a motor health state space in real time through a manifold learning model that maintains the inherent physical constraints of the motor system, wherein the motor health state space represents the current comprehensive health state of the motor with a unique coordinate point; Based on a strategy database containing operation strategies in different health states and task scenarios, a health cost function associated with the coordinate point in the motor health state space is constructed by using an inverse reinforcement learning algorithm, and the health cost function is used to quantify the irreversible health loss caused by executing different operation strategies in different health states; A dynamic economic system model is constructed by virtually connecting different health-related physical components of the motor to a plurality of interrelated virtual agents, and a decision model is trained by combining the health cost function through an optimal control algorithm, wherein the decision model is used to output an optimal operation strategy that minimizes the total health cost of the motor within the whole life cycle when the given coordinate point corresponding to the current comprehensive health state and the future task demand are given; In the actual operation process of the motor, real-time multi-source sensor data is acquired and mapped to a current coordinate point in a motor health state space; A production task is received from an upper control system of the motor, the current coordinate point and the production task are combined, and a decision model is used to determine an optimal operation strategy at the current time, and the optimal operation strategy is parsed into motor control parameters and sent to a motor driver for execution.

[0005] Optionally, the multi-source sensor data at least includes three-phase current waveform data reflecting motor electrical stress, bearing seat multi-axis vibration data reflecting motor mechanical fatigue, winding end and bearing temperature data reflecting motor internal thermal stress state, and shell temperature data reflecting motor external thermal stress state.

[0006] Optionally, the manifold learning model that maintains the inherent physical constraints of the motor system is used to map the multi-source sensor data to the motor health state space in real time, including the following steps: Acquire historical multi-source sensor data of the motor at different health stages and operating conditions, and pre-process the historical multi-source sensor data into high-dimensional feature vectors; Design an autoencoder neural network including an encoder and a decoder as the manifold learning model, the encoder is used to compress the high-dimensional feature vectors into a preset motor health state space, and the decoder is used to reconstruct the coordinate point in the motor health state space into a high-dimensional feature vector; Construct a combined loss function including a reconstruction error term and a symplectic structure penalty term, the reconstruction error term is used to measure the difference between the original high-dimensional feature vector and the reconstructed high-dimensional feature vector, and the symplectic structure penalty term is used to measure the deviation of the motor health state space from the preset inherent physical constraints; Train the manifold learning model using the historical multi-source sensor data and the combined loss function until the manifold learning model converges; Map the high-dimensional feature vector formed by the real-time collected multi-source sensor data to a coordinate point in the motor health state space using the trained and converged manifold learning model.

[0007] Optionally, the construction and calculation of the symplectic structure penalty term includes the following steps: In the process of mapping the high-dimensional feature vector to the motor health state space by the encoder in the manifold learning model, the Jacobian matrix of the mapping transformation is calculated; Define a standard symplectic form matrix corresponding to the dimension of the motor health state space; Obtain a metric matrix by matrix operation combining the Jacobian matrix and the standard symplectic form matrix; The difference between the metric matrix and the standard symplectic form matrix is taken as the quantization result of the symplectic structure penalty term; The quantification result of the octet structure penalty term is multiplied by a preset weight coefficient and added to the reconstruction error term to form a combined loss function.

[0008] Optionally, the strategy database containing strategies running in different health states and task scenarios adopts an inverse reinforcement learning algorithm to inversely deduce and construct a health cost function associated with coordinate points in the motor health state space, and the method comprises the following steps: A strategy database containing a plurality of strategy samples is established, and each strategy sample contains a health state defined by a coordinate point in the motor health state space, a task scenario, and a corresponding selected motor control parameter; The motor control parameters in the strategy database are abstracted into discrete dedicated running strategies; The coordinate points in the motor health state space are taken as states, and the dedicated running strategies are taken as actions to construct a Markov decision process; A maximum entropy inverse reinforcement learning algorithm is applied to solve the cost distribution of the state-action pairs from the Markov decision process and the strategy samples through iterative optimization; The cost distribution of the state-action pairs is fitted and generalized to construct a health cost function, which is used to input any health state coordinate point and a candidate running strategy to output a scalar health loss cost value.

[0009] Optionally, the method of constructing a dynamic economic system model virtually associating a plurality of health-related physical components of the motor with a plurality of virtual agents, and training a decision model by an optimal control algorithm in combination with the health cost function comprises the following steps: A corresponding virtual agent is defined for each of a plurality of key physical components of the motor, and an utility function is set for each virtual agent to maximize its own health capital, and the key physical components at least include a motor rotor, a motor winding, and a motor bearing; A dynamic economic system model is established, which describes the interaction between the virtual agents and the influence on the total health capital of the motor system under the influence of the running strategy through a stochastic dynamic equation; A reinforcement learning training framework is constructed in the environment of the dynamic economic system model, and a reward function of the reinforcement learning training framework is defined in combination with the health cost function and the task completion reward; An actor-critic reinforcement learning algorithm is used to simulate interactive training in the reinforcement learning training framework until the actor network in the actor-critic reinforcement learning algorithm converges, wherein the actor network in the actor-critic reinforcement learning algorithm is used to output a running strategy, and the critic network is used to evaluate the state value; The trained actor network is solidified as a decision model for outputting an optimal running strategy.

[0010] Optionally, the method further comprises the following steps: During the operation of the motor, a predicted health state coordinate point at the next time is predicted according to the current health state coordinate point and the selected optimal operation strategy by using the dynamic economic system model; A real health state coordinate point actually mapped out by the manifold learning model is acquired at the next time; A geometric distance between the predicted health state coordinate point and the real health state coordinate point in the motor health state space is calculated as a model prediction error; If the model prediction error continuously exceeds a preset error threshold value within multiple time steps, it is determined that a model mismatch event occurs; When the model mismatch event occurs, data including multi-source sensor data, the executed optimal operation strategy and the real health state coordinate point collected in the recent period are stored in a preset online learning database, and the online learning database is used to perform online fine-tuning on the decision model.

[0011] Optionally, the method of receiving a production task from an upper control system of the motor, combining the current coordinate point and the production task and determining the optimal operation strategy at the current time by using the decision model, and parsing the optimal operation strategy into motor control parameters and issuing the motor control parameters to the motor driver for execution comprises the following steps: A mapping table between operation strategies and motor control parameter sets is established in advance, each of the operation strategies corresponds to a set of motor control parameter sets including pulse width modulation frequency, current loop controller parameters, torque change rate limit and cooling system control level; The current coordinate point and the production task are input into the decision model, and when the decision model outputs the optimal operation strategy, the optimal motor control parameter set corresponding to the optimal operation strategy is retrieved through the mapping table; The optimal motor control parameter set is packaged into a data frame conforming to a preset industrial bus protocol; The data frame is sent to the motor driver through a high-speed real-time industrial bus; The data frame is received and parsed by the motor driver, and the optimal motor control parameters in the data frame are updated in real time to the corresponding control registers in the internal motor driver.

[0012] In a second aspect, the application further provides a motor health management system based on whole life cycle monitoring, comprising a memory, a processor and a computer program stored in the memory and executable on the processor, and the processor implements the motor health management method based on whole life cycle monitoring as described in any one of the above aspects when executing the computer program.

[0013] In a third aspect, the present application also provides a computer readable storage medium, which stores instructions configured to cause a processor to perform the motor health management method based on whole life cycle monitoring according to any one of the above aspects when the instructions are executed by the processor.

[0014] The present application has the following beneficial effects: The present application condenses complex multi-source data into a unique health state coordinate through manifold learning, providing a solid foundation for accurate quantification and decision-making. The innovation lies in using inverse reinforcement learning to infer and construct a health cost function from expert operations, which is equivalent to giving the motor a value perception ability, enabling it to quantify the irreversible damage to its life caused by different operating strategies. Based on this, the decision-making model trained by the dynamic economic system model and the optimal control algorithm no longer simply pursues the fastest completion of the current task, but carefully weighs the current output and future health cost in each control decision, thereby minimizing the total cost of the motor's whole life cycle. Ultimately, through real-time closed-loop control, the motor's health degradation path is actively intervened and managed, enabling the motor to intelligently adjust its working mode and fundamentally achieving a perfect balance between performance, efficiency, and life. BRIEF DESCRIPTION OF DRAWINGS

[0015] Figure 1 FIG. 1 is a flowchart of a motor health management method based on whole life cycle monitoring according to an embodiment of the present application.

[0016] Figure 2 FIG. 3 is a flowchart of real-time mapping of multi-source sensor data to a motor health state space according to an embodiment of the present application.

[0017] Figure 3 FIG. 4 is a flowchart of determining an optimal operating strategy and resolving the optimal operating strategy into motor control parameters for execution by a motor driver according to an embodiment of the present application. DETAILED DESCRIPTION

[0018] The technical solutions in the embodiments of the present application will be described clearly below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art belong to the scope of protection of the present application.

[0019] The terms "first", "second", etc. in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second", etc. are generally a class and do not limit the number of objects, for example, the first object can be one or more. In addition, "and / or" in the specification and claims indicates at least one of the connected objects, and the character " / " generally indicates that the objects before and after are in an "or" relationship.

[0020] Figure 1 A flowchart of an embodiment of a motor health management method based on full life cycle monitoring. It should be understood that, although Figure 1 The steps in the flowchart are displayed in sequence according to the direction of the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, the execution of these steps is not strictly limited in sequence, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps in the flowchart can include multiple sub-steps or multiple stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be alternately executed with other steps or at least part of the sub-steps or stages of other steps. As Figure 1 As shown in the figure, the motor health management method based on full life cycle monitoring disclosed by the present application specifically includes the following steps: S101. Collecting multi-source sensor data containing electrical, mechanical and thermodynamic characteristics during motor operation, and mapping the multi-source sensor data to the motor health state space in real time through a manifold learning model that maintains the inherent physical constraints of the motor system.

[0021] Among them, the multi-source heterogeneous data reflecting the comprehensive operation state of the motor are comprehensively collected. These data cover multiple physical domains of the motor, for example, three-phase current waveforms reflecting electrical stress are obtained through current sensors, multi-axis bearing seat vibration signals reflecting mechanical fatigue are collected through accelerometers, and winding, bearing and shell temperatures representing internal and external thermal stress states are monitored using thermocouples or infrared sensors. The collected original time series data are first standardized, filtered and denoised, and feature extracted, and are converted into uniform high-dimensional feature vectors. Subsequently, a structure based on a self-encoder neural network can be used as a manifold learning model to nonlinearly map the high-dimensional feature vectors to a pre-set low-dimensional motor health state space. The core of this model lies in the construction of its loss function, which not only includes a reconstruction error term for minimizing the difference between the input and the reconstructed output, but also introduces a symplectic structure penalty term. The symplectic structure penalty term ensures that the transformation process from the high-dimensional observation space to the low-dimensional health space can maintain the inherent physical constraints such as energy conservation of Hamiltonian systems of the motor system by constraining the Jacobian matrix of the mapping process. Finally, by fully training the model containing the combined loss function on the historical data set, a mapping tool can be obtained which can efficiently and accurately compress real-time multi-source data into a unique coordinate point in the health state space. The coordinate point not only integrates multi-source information, but also has a high degree of interpretability due to its embedded physical meaning, providing a solid foundation for subsequent health assessment and decision-making.

[0022] S102. Based on the strategy database containing the operation strategies under different health states and task scenarios, a health cost function associated with the coordinate point in the motor health state space is inferred and constructed in reverse using a reinforcement learning algorithm.

[0023] Wherein, it is extremely difficult to directly establish a loss function based on physical model for complex motor system, therefore inverse reinforcement learning algorithm is adopted to infer from expert experience data. Firstly, a comprehensive strategy database needs to be established, which stores a large number of strategy samples. Each sample records the specific motor control parameters selected by the motor when it is in a certain specific health state (defined by the health space coordinate point obtained in the foregoing step) and faces a specific task scenario. Then, the specific control parameters in the database are abstracted into a series of independent dedicated operation strategies, such as high-efficiency strategy, high-torque strategy or low-temperature operation strategy. Based on this, a Markov decision process is constructed, wherein the coordinate point of the health state space is defined as the state, and the abstracted dedicated operation strategy is defined as the action. With the help of maximum entropy inverse reinforcement learning algorithm, the optimal state-action pair cost distribution can be solved from the expert samples in the strategy database through an iterative optimization process under the framework of the Markov decision process. The basic principle of the algorithm is to select the cost function that maximizes the entropy of the expert strategy probability distribution among all the cost functions that can explain the expert behavior, which ensures the best generalization of the model in unknown areas. Finally, the solved discrete cost distribution is constructed into a continuous health cost function through function fitting and generalization techniques. This function can input any health state coordinate point and alternative operation strategy, and output a scalar health loss cost value.

[0024] S103. Construct a dynamic economic system model that virtually associates different health-related physical components of the motor with a plurality of virtual agents, and train a decision-making model by combining the health cost function and an optimal control algorithm.

[0025] In the dynamic economic system model, the physical components of the motor, such as the rotor, winding and bearing, which are closely related to the health of the motor, are virtually represented as a plurality of interrelated intelligent agents, forming a dynamic economic system. Each virtual agent is endowed with a utility function aimed at maximizing its own health capital. The entire dynamic economic system is described by a set of stochastic dynamic equations, which depict how the health capital of each virtual agent interacts with each other and how they collectively affect the total health capital of the motor system under the influence of different operating strategies. Subsequently, a reinforcement learning training framework is constructed in the simulation environment of the dynamic economic system model. The reward function of this framework is carefully designed, which not only includes the direct benefits brought by the execution of the task, but also integrates the health cost function constructed in the previous step, taking the health loss caused by the execution strategy as a negative reward. Advanced actor-critic reinforcement learning algorithms are used in this framework for extensive simulation interaction training. During the training process, the actor network is responsible for exploration and output of the operating strategy, while the critic network is responsible for evaluating the long-term cumulative return brought by the execution of the strategy under a certain health state. Through continuous trial and error and optimization, the actor network converges, at which time the network has the ability to output the optimal operating strategy given any health state and task demand. Finally, this trained and converged actor network is solidified to form the final decision model, which is used to guide the actual operation of the motor.

[0026] S104. In the actual operation of the motor, real-time acquisition of multi-source sensor data and mapping into the current coordinate point in the motor health state space.

[0027] In the actual operation stage of the motor, the health perception capability constructed needs to be put into practice to realize continuous and accurate control of the motor state. The sensor network deployed on site, including current transformers, vibration sensors, and temperature probes, continuously collects multi-source data streams reflecting the motor operating conditions at a high sampling rate. These real-time data streams are first sent to an edge computing unit or data acquisition system for preliminary data alignment and preprocessing. The processing flow is identical to the model training stage to ensure consistency of data features. After preprocessing, the multi-source data are integrated into a high-dimensional feature vector. Then, the encoder part of the manifold learning model trained and converged in the first step is called to input the real-time generated high-dimensional feature vector. The encoder network then performs an efficient forward propagation calculation, which is a deterministic mathematical transformation process, which can be represented as z = fenc(x), where x is the high-dimensional sensor data feature vector, fenc represents the solidified encoder mapping function, and z is the output low-dimensional health state space coordinate point. The mapping process has low computational complexity and can meet the real-time requirements of industrial sites. Finally, this method can convert complex and massive raw sensor data into a health coordinate point that is intuitive and rich in physical meaning. This coordinate point constitutes the core state variable of the motor digital twin and provides a clear and quantitative snapshot of the current motor health status for the upper control system, which is the direct input and key basis for subsequent optimal health management decisions.

[0028] S105. Receive production tasks from the upper control system of the motor, combine the current coordinate point and the production tasks, and determine the optimal operation strategy at the current time using the decision model, and parse the optimal operation strategy into motor control parameters and issue them to the motor driver for execution.

[0029] wherein a specific production task instruction is received from an upper control system, such as a manufacturing execution system or a process control system, which defines the performance target that the motor needs to achieve in the future, such as a specific speed, torque or production rhythm. At the same time, a coordinate point representing the current comprehensive health status of the motor is obtained through the aforementioned real-time monitoring step. The future task requirement and the current health coordinate point are simultaneously input into the mature decision model that has been trained. Based on its internal complex neural network structure, the decision model quickly calculates and outputs an operation strategy that is judged as the current optimal one. This strategy is an abstract instruction, such as a balance mode or a durability mode. In order to enable it to be understood and executed by the motor driver, a pre-established mapping table is queried. The mapping table defines in detail the correspondence between each abstract operation strategy and a specific set of motor control parameters, which usually includes the pulse width modulation frequency, the proportional integral parameters of the current loop controller, the amplitude limit value of the torque change rate, and the control level of the cooling system (such as the fan or the water pump). After finding the corresponding optimal set of motor control parameters, the system packages it into a data frame conforming to the preset industrial bus protocol, and issues the data frame to the motor driver through the high-speed real-time industrial bus network. The motor driver receives and parses the data frame, and updates the optimal control parameters in it to the internal corresponding control registers in real time and seamlessly, thereby changing the operation mode of the motor, and finally realizing an intelligent, closed-loop health management process that can dynamically adjust the operation mode according to the task requirement and the health status of the motor.

[0030] In one embodiment, referring to Figure 2 , the real-time mapping of multi-source sensor data to the motor health state space by a manifold learning model that maintains the intrinsic physical constraints of the motor system includes the following steps: S201. Obtain historical multi-source sensor data of the motor at different health stages and operating conditions, and pre-process the historical multi-source sensor data into high-dimensional feature vectors; S202. Design an autoencoder neural network containing an encoder and a decoder as the manifold learning model, wherein the encoder is used to compress the high-dimensional feature vectors into a pre-set motor health state space, and the decoder is used to reconstruct the coordinate points in the motor health state space into high-dimensional feature vectors; S203. Construct a combined loss function containing a reconstruction error term and a symplectic structure penalty term, wherein the reconstruction error term is used to measure the difference between the original high-dimensional feature vectors and the reconstructed high-dimensional feature vectors, and the symplectic structure penalty term is used to measure the degree of deviation of the motor health state space from the pre-set intrinsic physical constraints; S204. Train the manifold learning model using the historical multi-source sensor data and the combined loss function until the manifold learning model converges; S205. The manifold learning model trained to convergence is used to map the high-dimensional feature vector formed by the real-time collected multi-source sensor data into a coordinate point in the motor health state space.

[0031] In this embodiment, multi-source sensor signals are continuously collected throughout the life cycle of the motor, from brand new to failure, or under various operating conditions including different loads, speeds, and ambient temperatures. These signals should cover the key physical domains of the motor, for example, three-phase current data acquired using Hall sensors or current transformers to represent electrical stress; vibration data collected by multi-axis accelerometers mounted on the bearing seat to reflect mechanical wear and fatigue; and winding and casing temperature data monitored by embedded thermocouples or external infrared probes to represent internal and external thermal stress. After obtaining the raw time series data, a series of preprocessing operations must be performed to convert them into a standardized format suitable for machine learning model input. The specific implementation includes denoising filtering of the original signals, followed by normalization to eliminate the influence of different sensor dimensions, and then extracting a series of features that can sensitively reflect state changes through short-time Fourier transform, wavelet packet decomposition, or direct calculation of statistical quantities (such as root mean square, kurtosis, and margin factor) in the time domain. All these features are combined into a high-dimensional feature vector, each corresponding to the characteristics of the motor at a certain time, ultimately forming a large and complete historical data set.

[0032] The manifold learning model can efficiently compress the preprocessed high-dimensional feature vector into a low-dimensional and physically meaningful health state space. An effective implementation is to build an autoencoder neural network composed of an encoder and a decoder. The function of the encoder part is to perform nonlinear dimensionality reduction, which is stacked by multiple neural network layers, with the number of neurons decreasing layer by layer to form an information bottleneck. The high-dimensional feature vector input is continuously compressed and refined as it passes through the encoder, and finally outputs a low-dimensional vector at the bottleneck layer, which is the coordinate point of the motor in the health state space. The decoder part has a symmetrical structure opposite to the encoder, with the number of neurons in the network layers increasing layer by layer, responsible for receiving the low-dimensional coordinate point from the bottleneck layer and attempting to reconstruct it into the original high-dimensional feature vector without loss. The entire mapping and reconstruction process can be described by the formula and where x is the input high-dimensional feature vector, z is the low-dimensional health state space coordinate point output by the encoder, represents the nonlinear mapping function of the encoder, is the reconstructed high-dimensional feature vector output by the decoder, and represents the nonlinear reconstruction function of the decoder.

[0033] To ensure that the health state space constructed by manifold learning model not only accurately reflects the characteristics of the data, but also follows the inherent physical laws of the motor system, a special combination loss function needs to be designed to guide the training of the model. The loss function consists of two key parts: reconstruction error term and symplectic structure penalty term. The reconstruction error term is a standard component of the autoencoder, which is used to measure the difference between the original high-dimensional feature vector and the reconstructed high-dimensional feature vector after the encoding and decoding process, usually quantified by mean square error. The existence of this term ensures that the low-dimensional coordinate points retain enough information to restore the original data. More importantly, the symplectic structure penalty term introduces physical constraints into the model. As a dynamic system, the energy evolution process of the motor has the property of symplectic structure preservation under the Hamiltonian mechanics framework. This penalty term calculates the Jacobian matrix of the encoder mapping process and measures the extent to which the transformation deviates from the mathematical requirements of preserving symplectic structure, thereby penalizing the mapping that does not comply with the physical laws.

[0034] Model training is the process of combining the network structure and loss function designed in the preceding with the prepared historical multi-source sensor data to determine the specific values of all connection weights in the neural network through iterative optimization. The core of the training is to use a gradient-based optimization algorithm, such as the adaptive moment estimation algorithm, to continuously adjust the model parameters to minimize the value of the combined loss function. In each iteration, a small batch of high-dimensional feature vectors is randomly selected from the historical data set and input into the autoencoder model. The model performs a complete forward propagation calculation to obtain the reconstruction vector and low-dimensional coordinate point. Subsequently, according to the definition of the combined loss function, the reconstruction error and the symplectic structure penalty of the current batch of data are calculated, and their weighted sum is obtained to obtain the total loss. Next, the backpropagation algorithm is used to calculate the gradient of the total loss with respect to each parameter (weight and bias) in the model, which indicates the direction of parameter adjustment. The optimization algorithm updates the parameters slightly based on the gradient information, and the update rule can be abstracted as where represents the parameter set of the model at the t-th iteration, is the learning rate, which controls the step size of each update, C is the total loss function, is the gradient of the loss function with respect to the parameter.

[0035] When the manifold learning model is trained to converge, it can be deployed in the actual motor health management system for real-time state monitoring and evaluation of the motor. In the actual application stage, the core component is the trained encoder part, and the decoder part is no longer needed after completing its auxiliary task in the training stage. The specific implementation process is as follows: first, the multi-source sensors installed on the running motor collect real-time current, vibration, temperature and other raw data streams. These data streams are sent to the data acquisition and preprocessing module, which performs the same denoising, normalization and feature extraction operations as in the training stage to convert the instantaneous sensor readings into a high-dimensional feature vector. Subsequently, the high-dimensional feature vector is directly input into the solidified encoder network. The encoder network performs a fast forward propagation calculation, which is a deterministic mapping process, which can be represented as y = h(v), where v is the real-time collected and preprocessed high-dimensional feature vector, h represents the trained and converged encoder mapping function, and y is the output low-dimensional coordinate point. Due to the high computational efficiency of neural network forward propagation, the entire mapping process can be completed within milliseconds, fully meeting the real-time requirements of industrial sites. Finally, the complex and multi-dimensional sensor data is converted into a unique coordinate point in the health state space in real time and stably, which directly represents the current comprehensive health state of the motor.

[0036] In one embodiment, the construction and calculation of the symplectic structure penalty term includes the following steps: In the process of mapping the high-dimensional feature vector to the motor health state space by the encoder in the manifold learning model, the Jacobian matrix of the mapping transformation is calculated; A standard symplectic form matrix corresponding to the dimension of the motor health state space is defined; The metric matrix is obtained by matrix operation combining the Jacobian matrix and the standard symplectic form matrix; The difference between the metric matrix and the standard symplectic form matrix is taken as the quantization result of the symplectic structure penalty term; The quantization result of the symplectic structure penalty term is multiplied by a preset weight coefficient and added to the reconstruction error term to form a combined loss function.

[0037] In this embodiment, the encoder is essentially a complex vector function that maps a high-dimensional input feature vector to a low-dimensional health state space coordinate point. The Jacobian matrix is the first-order derivative matrix of this vector function, which accurately describes the instantaneous rate of change of each dimension of the output coordinate with respect to each dimension of the input feature. Specifically, if the encoder mapping function is represented as where x is an m-dimensional input feature vector, z is a 2n-dimensional output health coordinate point, 2n < m, and the dimension is even to adapt to the Hamiltonian mechanics framework, then the Jacobian matrix J is a The Jacobian matrix is ​​a matrix that, for encoders implemented using neural networks, can be efficiently computed using backpropagation and the chain rule. Modern deep learning frameworks can automate this process with automatic differentiation. For each high-dimensional feature vector input into the encoder, a corresponding Jacobian matrix is ​​calculated. This matrix is ​​not a fixed model parameter but dynamically changes with the input data. It captures information about local geometric transformations such as stretching, rotation, and shearing in the neighborhood of a specific data point in high-dimensional space, mapping to low-dimensional space. It forms the basis for evaluating whether the transformation preserves a specific physical structure.

[0038] To measure the physical compliance of encoder mappings, a clear mathematical benchmark is needed to represent the ideal physical structure. In the Hamiltonian framework describing passive or weakly dissipative dynamic systems such as electric motors, this ideal structure is defined by a standard symplectic matrix. This matrix is ​​a fixed square matrix corresponding to the dimension of the motor's health state space. It is not calculated but rather a predefined form based on mathematical axioms. In practice, if the dimension of the motor's health state space is set to 2n, this space can be understood as consisting of n generalized position coordinates and n generalized momentum coordinates. In this case, the standard symplectic matrix S is... A block matrix. The matrix has the form of... ,in yes The identity matrix, where 0 is... The zero matrix. This matrix is ​​antisymmetric. and satisfaction The importance of this property. In a physical sense, preserving the symplectic structure means that the phase space volume of the system is conserved during evolution, which is closely related to the law of conservation of energy. Therefore, defining this standard symplectic form matrix essentially provides a lossless mathematical template that completely follows the evolutionary laws of Hamiltonian systems for subsequent comparisons. Any transformation deviating from this template will be considered a violation of the intrinsic physical constraints.

[0039] Next, the Jacobian matrix representing the actual transformation is combined with the standard symplectic form matrix representing the ideal structure to generate a metric matrix. The calculation of this metric matrix is ​​a pure matrix operation, reflecting the form of the standard symplectic form after the coordinate transformation described by the Jacobian matrix. Specifically, combining the Jacobian matrix J calculated in the previous steps and the defined standard symplectic form matrix S, the metric matrix M is obtained through matrix operations of congruent transformation. The mathematical expression for this operation is: In this expression, It is the transpose of the Jacobian matrix J. The logic of this operation lies in transforming the standard symplectic structure S in the low-dimensional healthy state space through the inverse process of mapping (from...). Approximately) back to the high-dimensional space, and then transformed back to the low-dimensional space by the forward mapping (approximated by J). If the mapping transformation of the encoder is a perfect symplectic mapping, i.e. completely preserves the intrinsic physical structure of the Hamiltonian system, then after this series of transformations, its structure should remain unchanged. In other words, in this ideal case, the calculated metric matrix M will be exactly equal to the original standard symplectic form matrix S. Therefore, the calculation result of the metric matrix M provides a direct comparison object for the subsequent quantization bias, and its form directly reflects the degree to which the mapping transformation follows the physical constraints under the current neural network model parameters.

[0040] After obtaining the metric matrix M which can reflect the actual transformation structure and the standard symplectic form matrix S which represents the ideal physical structure, the next step is to convert the difference between them into a scalar penalty value that can be used for optimization. This process aims to quantify the inconsistency between the two matrix structures into a single non-negative numerical value, which is directly proportional to the degree to which the mapping transformation deviates from the physical constraints. An effective and numerically stable implementation is to calculate the norm of the difference between the metric matrix and the standard symplectic form matrix. Specifically, the Frobenius norm of the matrix can be used because it is equivalent to taking the square root of the sum of the squares of all elements of the matrix in calculation, and has good differentiability, which is convenient for use in the backpropagation of the neural network. Therefore, the quantization result of the symplectic structure penalty term can be calculated by the following formula: . Where, represents the Frobenius norm, and the superscript 2 indicates that the norm is squared, which can simplify the complexity of derivative calculation. The factor 1 / 2 is also for mathematical convenience and can be canceled out after differentiation. When the encoder's mapping completely preserves the symplectic structure, M is equal to S, the difference between the matrices is a zero matrix, and the value of the penalty term is zero. As the degree of deviation of the mapping from the physical constraints increases, the element difference between M and S will become larger, causing the value of to also increase accordingly.

[0041] Finally, the symplectic structure penalty term quantified in the previous steps is combined with the standard reconstruction error term to form a unified combined loss function. This combined loss function will serve as the final objective function for parameter optimization of the neural network model. In specific implementation, the quantization result of the symplectic structure penalty term is scaled by a pre-set weight coefficient, and then linearly summed with the reconstruction error term. The reconstruction error term is usually calculated using the mean square error to measure the difference between the original input and the decoder reconstruction output. The expression of the final combined loss function L is: . In this formula, is a pre-set non-negative weight coefficient, and When the value is large, model training will focus more on minimizing the symplectic structure penalty, that is, forcing the model to strictly adhere to physical constraints in the learned mappings, which may sacrifice some reconstruction accuracy; conversely, when When the value is small, the model prioritizes ensuring the quality of data reconstruction. By minimizing this combined loss function L, optimization algorithms (such as gradient descent) drive the neural network weights to update in a direction that simultaneously reduces reconstruction error and symplectic structure penalty. The resulting trained model is not only an efficient data compressor but also a physically self-consistent feature extractor, and its constructed healthy state space has stronger generalization ability and better interpretability.

[0042] In one implementation, based on a policy database containing operating strategies under different health states and task scenarios, the inverse reinforcement learning algorithm is used to infer and construct a health cost function associated with coordinate points in the motor health state space, including the following steps: Establish a strategy database containing multiple strategy samples. Each strategy sample includes a health state, a task scenario, and corresponding selected motor control parameters defined by coordinate points in the motor health state space. The motor control parameters in the strategy database are abstracted into discrete, dedicated operation strategies; A Markov decision process is constructed by taking the coordinate points in the motor health state space as states and the exclusive operation strategy as actions. The cost distribution of state-action pairs is solved from Markov decision processes and policy samples by applying the maximum entropy inverse reinforcement learning algorithm and iterative optimization. The health cost function is constructed by fitting and generalizing the cost distribution of state-action pairs. The health cost function is used to output a scalar health loss cost value after inputting any health state coordinate point and alternative operating strategies.

[0043] In the present embodiment, the policy database is constructed by collecting and organizing a large number of policy samples, each of which is a data tuple that accurately records the decision-making situation of the motor at a specific moment. Specifically, each tuple must contain three core elements: first, the current comprehensive health status of the motor defined by the low-dimensional coordinate point output by the prequel manifold learning model; second, the specific task scenario of the motor, such as high-load climbing, constant-speed cruising, or frequent start-stop working condition description; and third, the motor control parameter set selected by the expert system or experienced operator under the health status and task scenario, such as specific pulse width modulation frequency, current loop gain, and cooling level. These sample data can be obtained from the historical operation log of the motor throughout its life cycle or actively acquired through accelerated aging tests on the test bench. The original motor control parameters are usually multi-dimensional continuous variables, and directly using them as decision actions will make the subsequent inverse reinforcement learning problem extremely complex and difficult to solve. Therefore, the specific control parameters in the policy database need to be abstracted and mapped to a limited number of discrete dedicated running strategies.

[0044] The implementation of this step can use an unsupervised clustering algorithm, such as the K-Means algorithm. Extract the motor control parameter part of all samples in the policy database to form a high-dimensional parameter vector space. Then, apply the K-Means algorithm to find a pre-set number (K) of cluster centers in this space. Each calculated cluster center, i.e., the centroid, represents a class of control parameter combinations with similar characteristics and can be assigned a clear physical meaning, such as high-efficiency strategy, high-torque output strategy, low-temperature running strategy, or vibration suppression strategy. After clustering, traverse the entire policy database and replace the original motor control parameter vector in each sample with the label of its corresponding cluster center, i.e., the corresponding dedicated running strategy. Through this process, a continuous and high-dimensional action space is successfully simplified to a limited action set consisting of K discrete actions, greatly reducing the modeling and solving complexity of the subsequent Markov decision process. In order to apply the inverse reinforcement learning algorithm, the health evolution and decision-making problem of the motor must be formally constructed as a Markov decision process. This is a standard mathematical framework for describing sequential decision-making problems. In this framework, the core components need to be defined first. Specifically, the continuous coordinate points in the low-dimensional motor health state space obtained by the manifold learning model are defined as the state space s. The discrete dedicated running strategy set obtained by abstraction in the previous step is defined as the action space a. The complete definition of the Markov decision process is a tuple where s and a have been defined, P represents the state transition probability, i.e. the probability of transitioning to state s' after performing action a in state s, which is unknown in the initial stage. C is the unknown cost function, representing the immediate health loss incurred by performing action a in state s, which is the core objective of inverse reinforcement learning. is a discount factor between 0 and 1, balancing the importance of immediate cost and future cost. By establishing this Markov Decision Process model, it provides a standardized mathematical problem description for inferring the cost function C from expert policy samples.

[0045] Under the framework of Markov Decision Process, a maximum entropy inverse reinforcement learning algorithm is applied to solve the cost distribution of state-action pairs from the expert samples in the policy database. The core idea of this algorithm is to find a cost function that maximizes the entropy of the expert policy trajectory probability distribution among all cost functions that can explain the expert's demonstration behavior. This method can effectively avoid overfitting to the expert's behavior and exhibit better generalization ability for unseen states. The implementation of the algorithm is an iterative optimization process: first, a cost function is randomly initialized; then, based on the current cost function, the optimal policy under this cost function is obtained by solving the forward reinforcement learning problem through dynamic programming and other methods; next, the expected visit frequency of the states under this optimal policy is calculated and compared with the visit frequency of the states statistically obtained from the expert samples; finally, according to the difference between the two, the gradient descent method is used to update the parameters of the cost function, so that the probability of the expert policy under the current cost function is maximized. A trajectory The probability value under the given cost function parameters where is the total cost of the trajectory. Through repeated iterations, the algorithm eventually converges and outputs a discrete cost distribution, i.e. a quantitative cost value c(s, a) is assigned to each state-action pair (s, a) that appears in the database.

[0046] The cost distribution solved by the maximum entropy inverse reinforcement learning algorithm is discrete and only effective for state-action pairs that have appeared in the policy database. In order to obtain a general health cost function that can handle any health state and evaluate all candidate strategies, the discrete cost distribution needs to be fitted and generalized. This step can be achieved through function approximation methods, among which using a deep neural network as a function approximator is an efficient implementation. Specifically, a neural network model is constructed, whose input is the concatenation vector [s, a] of state and action, and whose output is a scalar cost value. Then, the discrete cost distribution As a training dataset, the neural network is supervised learning trained. The goal of the training is to minimize the error between the cost value predicted by the neural network and the cost value calculated by the inverse reinforcement learning algorithm, such as mean square error. When the neural network training converges, the internal weight parameters are fixed, forming the final health cost function, which can be expressed as wherein is a nonlinear function represented by the neural network, and W is the trained network weight. This function has strong generalization ability and can receive any health state s defined by coordinate points and an alternative operation strategy a, and output a scalar health loss cost value at once, providing a key quantitative evaluation basis for subsequent optimal control decisions.

[0047] In one embodiment, a dynamic economic system model is constructed by virtualizing different health-related physical components of the motor into a plurality of interrelated virtual agents, and a decision model is trained by an optimal control algorithm combined with the health cost function, including the following steps: Define corresponding virtual agents for a plurality of key physical components of the motor, and set a utility function for each virtual agent to maximize its own health capital, wherein the key physical components at least include the motor rotor, motor winding and motor bearing; Establish a dynamic economic system model, which describes the interaction between the virtual agents and the impact on the total health capital of the motor system under the influence of the operation strategy through a stochastic dynamic equation; Construct a reinforcement learning training framework in the environment of the dynamic economic system model, and define a reward function of the reinforcement learning training framework combined with the health cost function and the task completion reward; Use the actor-critic reinforcement learning algorithm to perform simulated interaction training in the reinforcement learning training framework until the actor network in the actor-critic reinforcement learning algorithm converges, wherein the actor network in the actor-critic reinforcement learning algorithm is used to output the operation strategy, and the critic network is used to evaluate the state value; Solidify the trained actor network as a decision model for outputting the optimal operation strategy.

[0048] In this embodiment, since static and dynamic eccentricity can be detected by fluctuations in the current value of electrical stress, winding faults can be calculated by calculating the life value through prolonged temperature exceedance, bearing wear can be analyzed by combining bearing vibration, bearing temperature, and current fluctuations, component fatigue and corrosion can be calculated by considering the internal winding temperature and online time of the motor, and selecting the failure time of the anti-corrosion coating, and cage damage can also be judged by combining vibration, current waveform, and noise, the core physical components closely related to health degradation in the motor system can be identified and abstracted, such as the motor rotor, stator windings, and support bearings, and a corresponding virtual agent can be defined for each component. Each virtual agent is assigned a single objective, namely, maximizing its own long-term health capital. To this end, a utility function needs to be set for each agent i. This function uses its health capital This is mapped to a scalar utility value. Health capital is a quantitative indicator that can be extracted or transformed from coordinate points in the motor's health state space to characterize the health level of that component. A simple implementation is to directly define the utility function as health capital itself, i.e. This setup means that each agent is committed to taking actions to maximize the expected cumulative value of their own health capital.

[0049] Based on the virtual agent setting, the next step is to establish an economic model that can describe the dynamic evolution of the entire system. The core of this model is a set of stochastic dynamic equations used to characterize how the health capital of each virtual agent changes over time under the influence of different operating strategies, and how they interact with each other. Specifically, for each agent i, its health capital at the next time step... health capital Modeled as its current moment's health capital Health capital of other agents System-selected operating strategy and a random perturbation term The function. This dynamic relationship can be represented as ,in It is a nonlinear function that can be constructed through system identification or based on physical knowledge, while the random term is used to describe the uncertainty of the model. This modeling approach can capture key physical coupling relationships. For example, an operating strategy that leads to winding overheating will not only directly deplete the winding's health capital but may also indirectly deplete the bearing's health capital by accelerating the aging of bearing grease through heat conduction. Ultimately, the total health capital of the entire motor system... It can be defined as the weighted sum of the health capital of all agents.

[0050] To train an optimal decision model using reinforcement learning algorithm, a complete training framework must be constructed, the core of which is to define a reward function that can accurately guide the learning direction of the agent. The training framework takes the previously established dynamic economic system model as the simulation environment. In this environment, the reinforcement learning agent (i.e., the future decision model) observes the current state of the system (defined by the motor health state space coordinate point) at each time step and selects an operating strategy as the action. The environment updates the health capital of each virtual agent according to the action and returns a scalar reward signal. The reward function must be able to balance the short-term task performance and long-term health maintenance goals. The reward function can be constructed as a linear combination of task completion benefits and health loss costs: where and are the state and action at time t, respectively, is the positive reward brought by completing the production task, such as being positively correlated with output torque or efficiency. c(s, a) is the health cost function obtained previously through inverse reinforcement learning, representing the irreversible health loss caused by executing action a. is a weight coefficient used to adjust the punishment level for health loss. Through the setting of this reward function, the optimization goal of the reinforcement learning agent becomes clear: find an operating strategy that maximizes the long-term cumulative reward while completing the task, which is equivalent to minimizing the total health cost throughout the life cycle.

[0051] After constructing the reinforcement learning training framework and reward function, the actor-critic algorithm is used to conduct extensive interaction training in the simulation environment to solve the optimal operating strategy. The actor-critic algorithm includes two deep neural networks: the actor network and the critic network. The function of the actor network is the policy function, which receives the current state s as input and outputs an optimal action a (or a probability distribution of actions), with the goal of learning how to act. The function of the critic network is the value function, which also receives the state s as input but outputs the long-term value assessment of the state , i.e., the expected cumulative reward that can be obtained by following the current actor policy from this state, with the goal of learning how to judge the goodness of the state. The training process is a mutually reinforcing cycle: the actor performs an action in the simulation environment according to the current policy, moves the environment to a new state s' and gives an immediate reward R. The critic network calculates a time difference error based on the value assessment of the new state and the immediate reward, which is (where The discount factor measures the accuracy of the value assessment. This error signal is used to update both networks simultaneously: on the one hand, it guides the critic network to adjust its parameters to make more accurate value predictions; on the other hand, it guides the actor network to adjust its parameters, increasing the probability of selecting actions that lead to positive errors (i.e., results better than expected) and decreasing the probability of selecting actions that lead to negative errors.

[0052] Once the actor-critic algorithm training is complete and the actor network converges, the final step is to extract this fully optimized actor network and solidify it into an independent, directly deployable decision model. During training, the critic network guides the actor network's learning by providing accurate value assessments. However, once training is complete, the actor network's strategy has internalized considerations of long-term value, and the critic network has fulfilled its purpose and is no longer needed in the decision-making phase. Therefore, all parameters of the converged actor network (including weights and biases) are frozen to prevent accidental modification in subsequent applications. The solidified actor network itself is a highly nonlinear function, establishing a direct mapping between the motor's current overall health state and future task requirements to the optimal operating strategy. This model can be viewed as an expert system, but its decision logic is not written by human rules; rather, it is autonomously formed through massive trial-and-error learning in a simulated environment. Finally, this lightweight and computationally efficient decision model is deployed into the motor's control system. In actual operation, it can receive real-time status input, perform a fast forward propagation calculation, and instantly output the optimal operating strategy under the current condition, thereby realizing closed-loop intelligent management of the motor's health status.

[0053] In one implementation, establishing a dynamic economic system model and constructing a reinforcement learning training framework based on the dynamic economic system model includes the following steps: Based on the utility function of the virtual agent and the changing pattern of the total health capital of the motor system, a stochastic differential equation describing the dynamic economic system model is established. Design a physical information neural network that takes the current time, current health status coordinates, and operating strategy as inputs, and outputs the predicted health status of each virtual agent at future time. The objective loss function of the physical information neural network is constructed. The objective loss function includes data residual terms and equation residual terms. The equation residual terms are calculated by automatic differentiation techniques to calculate the derivative of the physical information neural network output with respect to the input, and the degree of non-satisfaction of the equation obtained by substituting the derivative into the stochastic differential equation is quantified. In the simulated interaction process of the reinforcement learning training framework, the collected state transition trajectory data is used to update the data residual term, and the target loss function is continuously minimized to train the physical information neural network. The environment for constructing a reinforcement learning training framework is built using the converged physical information neural network as a differentiable surrogate model, which is used to provide state value evaluation for the critic network in the actor-critic reinforcement learning algorithm.

[0054] In this embodiment, a set of stochastic differential equations is established to characterize the continuous-time evolution of a dynamic economic system. This set of equations aims to capture the deterministic degradation trend and stochastic fluctuations of the health capital of each virtual agent. For the i-th virtual agent, its health capital... The infinitesimal change in time t It can be modeled. This change consists of two parts: a drift term, describing the average rate of change of health capital under the current system state and operating strategy; and a diffusion term, describing the stochastic fluctuations of health capital caused by unmodeled factors or inherent randomness. Specifically, the stochastic differential equation can be written as: In this equation, The representative at time t is composed of all agents' health capital. The vector state that constitutes the state; This is the currently applied operational strategy; It is the drift coefficient function that determines the deterministic evolutionary path of health capital; It is the diffusion coefficient function that determines the magnitude of the random fluctuation; dt is the infinitesimal time increment; and This represents the increment of a standard Wiener process, used to introduce Gaussian white noise.

[0055] The design goal of the physical information neural network is to learn the complex dynamics described by the aforementioned stochastic differential equations. Specifically, a deep feedforward neural network is constructed, whose input layer receives the current time t and the initial health capital vector derived from the current health state coordinate point s. The network's output layer then provides the health capital of each virtual agent at input time t, along with the selected constant operating strategy a. Predicted value The hidden layers of the network consist of multiple layers of neurons with nonlinear activation functions, enabling it to approximate highly nonlinear dynamic evolution functions. The entire physical information neural network can be viewed as a function approximator. Here, N represents the mapping function of the neural network, while W consists of all the weights and bias parameters in the network that need to be optimized through training. The design of this network structure enables it not only to make predictions based on data, but more importantly, due to its analytical mathematical expression, it allows for the calculation of the derivative of its output with respect to any order of any input using automatic differentiation techniques—a crucial prerequisite for imposing constraints on physical equations.

[0056] The training of the physical information neural network is not guided by a traditional supervised learning loss, but by a combined objective loss function that contains double information. The objective loss function is weighted by a data residual term and an equation residual term . The data residual term works similarly to standard supervised learning, using the state transition trajectory data collected from the simulation or real environment to measure the difference between the network's predicted value and the observed true value. For example, the mean squared error can be used to quantify: where is the collected data point. More crucially, the equation residual term encodes the known physical law (stochastic differential equation) directly into the loss function. In implementation, the derivative of the physical information neural network output with respect to the time input t is calculated using automatic differentiation techniques. Subsequently, this derivative and the network's own output are substituted into the drift part of the pre-established stochastic differential equation to obtain the equation residual . The equation residual term is then the squared norm of this residual f over the solution domain. The final combined loss function is where and are weight coefficients used to balance data fitting and physical constraints.

[0057] The training process of the physical information neural network is tightly coupled with the simulation interaction process of the reinforcement learning agent, forming a closed loop of online learning. In the reinforcement learning training framework, the actor network selects an action according to the current state and executes it in the dynamic economic system model (simulation environment). After the environment responds, a state transition trajectory is generated, i.e., from the initial state s to the next state s' after a period of time . This complete state transition tuple (s, a, s') is not only used to update the actor and critic networks, but also stored in an experience replay pool as training data for the physical information neural network. At each step of training, a batch of state transition trajectory data is randomly sampled from the replay pool. For this batch of data, on the one hand, the data residual term is calculated, i.e., the error between the network's prediction of the next state and the true observed value. On the other hand, the equation residual term is calculated at these trajectory points and other randomly sampled configuration points within the solution domain. Subsequently, the two residual terms are weighted and summed to obtain the total objective loss function, and a gradient-based optimization algorithm is used to calculate the gradient of the loss function with respect to the weights of the physical information neural network, and the weights are updated accordingly.

[0058] When the physical information neural network training converges, it becomes a fully differentiable dynamic economic system agent model. This agent model can replace the original high-cost simulation environment in the reinforcement learning training framework, thereby greatly improving the training efficiency, especially for the critic network learning. In the traditional actor-critic algorithm, the critic network can only learn the state value function slowly by observing the one-step time difference error. However, when there is a differentiable agent model, the learning method of the critic network can be fundamentally changed. The goal of the critic network is to learn a value function V(s) that satisfies the Bellman equation. Using the trained physical information neural network as the state transition function s'=N(s,a), the Bellman equation residual can be directly represented as a differentiable function with respect to the critic network parameters. This means that the gradient of the Bellman residual with respect to the critic network parameters can be directly calculated by the backpropagation algorithm, thereby more directly and stably minimizing the residual. This model-based method can "penetrate" the gradient information from the future multi-step state value evaluation through the environment model to the current state, providing more abundant and long-term supervision signals for the critic network update, thereby accelerating its convergence and ultimately helping the actor network to learn a better long-term strategy.

[0059] In one embodiment, the motor health management method based on full life cycle monitoring further comprises the following steps: During the operation of the motor, the dynamic economic system model is used to predict the predicted health state coordinate point at the next time according to the current health state coordinate point and the selected optimal operation strategy; At the next time, the real health state coordinate point actually mapped by the manifold learning model is obtained; The geometric distance between the predicted health state coordinate point and the real health state coordinate point in the motor health state space is calculated as the model prediction error; If the model prediction error exceeds the preset error threshold continuously for multiple time steps, it is determined that a model mismatch event occurs; When the model mismatch event occurs, the data collected in the near future, including multi-source sensor data, executed optimal operation strategy and real health state coordinate point, are stored in the preset online learning database, and the online learning database is used to fine-tune the decision model online.

[0060] In this embodiment, at the beginning of each control cycle, the coordinate point representing the current overall health state of the motor is first obtained, and the optimal operation strategy selected at this moment is obtained from the decision model. Subsequently, the previously trained dynamic economic system model is called, taking these two pieces of information as input, and a forward deduction is performed. According to the random dynamic equation describing the evolution of health capital embedded in the model, the predicted value of the health state coordinate point at the next moment is calculated and output. The mathematical expression of this prediction process can be abstracted as where is the real health state coordinate point observed at the current moment t, a(t) is the optimal operation strategy selected and about to be executed at this state, g represents the evolution function of the entire dynamic economic system model, and is the predicted result of the health state coordinate point at the next moment t+1. After the prediction step is completed, the motor driver will drive the motor to operate according to the specified operation strategy for one control cycle. At the end of this cycle, i.e. at the next moment, the real observation value of the motor health state must be obtained in order to compare with the previous prediction. This process is completely consistent with the daily real-time health state monitoring. The multi-source sensor network deployed on the motor collects real-time raw physical signals such as vibration, current and temperature at this moment. These signals are immediately sent to the data preprocessing module and converted into a standard high-dimensional feature vector through the same feature extraction process as in offline training.

[0061] Next, the encoder part of the manifold learning model that has been trained and converged is called, taking this freshly baked high-dimensional feature vector as input. The encoder network performs an efficient forward propagation calculation, nonlinearly mapping it to a low-dimensional health state space, thereby obtaining the real health state coordinate point at this moment. After obtaining the predicted coordinate point and the real coordinate point of the health state at the next moment, the deviation between the two needs to be accurately quantified in order to evaluate the current prediction performance of the dynamic economic system model. The quantification operation is achieved by calculating the geometric distance between the two coordinate points in the motor health state space. Since the health state space is a multidimensional Euclidean space, the most intuitive and effective distance measurement method is to calculate the Euclidean distance between the two points, also known as the L2 norm. The specific calculation formula is: In this formula, is the predicted health state coordinate point vector, is the real health state coordinate point vector, represents the L2 norm of the vector difference. This calculation process will result in a single non-negative scalar value i.e. the model prediction error. This error value intuitively reflects the deviation between the model-predicted health state evolution trajectory and the actual physical process. A smaller error value indicates that the model is currently accurately describing the system dynamics, while a significantly increased error value alerts that the model might have lost its ability to accurately capture the real degradation behavior of the motor.

[0062] The prediction error at a single time step can be caused by random noise or transient disturbances. Triggering model adjustment directly based on a single large error can lead to system instability. Therefore, a more robust criterion is needed to determine whether a fundamental model mismatch event has occurred. The implementation of this criterion is a continuous over-threshold detection mechanism based on a sliding time window. First, a reasonable error threshold and a window size representing the number of consecutive time steps are pre-set. After the model prediction error is calculated at each time step, the system checks whether all prediction errors within the window of time steps back from the current time step are continuously greater than the pre-set error threshold. The judgment logic condition is IF for all THEN Mismatch=True. Only when this condition is met, a model mismatch event is determined to have occurred.

[0063] Once a model mismatch event is determined to have occurred, indicating that the existing decision model can no longer accurately reflect the current condition of the motor, the online fine-tuning process must be initiated for adaptive update. At this time, the system automatically collects the recent historical data that led to the model mismatch and stores it in a pre-set database dedicated to online learning. The stored data samples are complete tuples containing rich information, each including at least: the original multi-source sensor data during the period of prediction inaccuracy, the optimal operation strategy executed at that time, and the real health state coordinate point trajectory mapped out by the manifold learning model. When a sufficient number of new samples have accumulated in the online learning database, the fine-tuning training of the decision model is triggered. The specific implementation is to use these new data to perform gradient update of the existing actor-critic network with an additional small learning rate. This means that on the basis of the original knowledge, the new patterns that led to high prediction errors are focused on learning. This process is not a complete retraining from scratch, but a fine-tuning based on the existing model parameters, which not only preserves the achievements of historical learning, but also efficiently absorbs new knowledge, ultimately enabling the decision model to quickly adapt to new changes in the motor and restore its optimality in decision-making.

[0064] In one embodiment, reference is made to Figure 3receive a production task from an upper control system of the motor, determine an optimal operation strategy at a current time by combining a current coordinate point and the production task and using a decision model, and parse the optimal operation strategy into motor control parameters and send the motor control parameters to a motor driver to execute, including the following steps: S301. A mapping table between operation strategies and motor control parameter sets is established in advance, each of the operation strategies corresponding to a set of motor control parameters including pulse width modulation frequency, current loop controller parameters, torque change rate limit, and cooling system control level. S302. The current coordinate point and the production task are input into the decision model, and when the decision model outputs an optimal operation strategy, the optimal motor control parameter set corresponding to the optimal operation strategy is retrieved through the mapping table. S303. The optimal motor control parameter set is packaged into a data frame conforming to a preset industrial bus protocol. S304. The data frame is sent to the motor driver through a high-speed real-time industrial bus. S305. The data frame is received and parsed by the motor driver, and the optimal motor control parameters in the data frame are updated in real time to the corresponding control registers in the motor driver.

[0065] In this embodiment, the construction of the mapping table is an offline engineering task. First, a series of discrete operation strategies with clear physical meaning are defined in advance, such as high efficiency strategy, high dynamic response strategy, low vibration strategy, or durability priority strategy. For each strategy, a set of specific motor control parameters that can optimally achieve the strategy goal is set through a large number of simulation analysis and experimental bench tests. This parameter set is a multi-dimensional vector, and its elements cover all key adjustable parameters inside the driver, including at least: pulse width modulation frequency that determines inverter switching loss and current harmonic; proportional and integral controller parameters that affect current loop response speed and stability; torque change rate limit value that limits mechanical impact; and cooling system control level that adjusts motor temperature rise. Finally, these strategies and corresponding optimal parameter sets are stored one by one to form a data structure that can be quickly queried.

[0066] In the real-time operation control loop of the motor, once the production task instruction issued by the upper system is received, the decision-making process is triggered. First, the current motor health status and the current production task demand are fed into the trained and converged decision-making model as dual inputs. The decision-making model performs a forward propagation calculation, and according to the complex decision-making logic learned inside, outputs the optimal operation strategy under the current comprehensive situation. After obtaining this abstract strategy instruction, the optimal operation strategy is used as the key to search in the pre-established mapping table at high speed. Through this query operation, the optimal motor control parameter set corresponding to the optimal operation strategy is accurately found. After obtaining the optimal motor control parameter set, these numerical parameters cannot be directly transmitted on the communication network, and must be packaged into standardized data units, i.e. data frames, which conform to the specification of the specific industrial bus protocol. In specific implementation, first, the high-speed real-time industrial bus protocol supported by the target motor driver needs to be determined, such as EtherCAT, Profinet or CANopen. Each protocol has a strict definition of the structure of the data frame. The packaging process is to convert each parameter value in the optimal motor control parameter set (such as pulse width modulation frequency, controller gain, etc.) into binary format according to the protocol specified data type (such as 16-bit unsigned integer, 32-bit floating point) and byte sequence (such as little-endian or big-endian). Then, according to the pre-set offset address in the device description file or object dictionary, the binary data is accurately filled into the payload area of the data frame. In addition to parameter data, a complete data frame also includes a frame header for addressing and routing, a cyclic redundancy check code for ensuring data integrity, and other parts.

[0067] After the data frame is packed, it is physically sent from the host controller to the target motor driver over a high-speed real-time industrial bus network. Specifically, the packed data frame is delivered to a dedicated network interface controller, such as an EtherCAT master interface card, by the central processing unit of the upper controller. The interface controller is responsible for handling all tasks of the data link layer and the physical layer, converting the digitized data frame into electrical or optical signals that can be transmitted over a cable. These signals are then sent out through a physical medium, such as a shielded twisted pair Ethernet cable. Unlike general office networks, the transmission process of industrial buses has a high degree of determinism and real-time performance. For example, in an EtherCAT network, the sending of data frames follows a cyclic communication schedule that is strictly controlled by the master and is accurate to the microsecond level. The data frame carrying the optimal motor control parameters will be sent within a predetermined time window and arrive at the motor driver with extremely low delay and jitter. When the data frame carrying the optimal motor control parameters arrives at the network interface of the motor driver, the motor driver first receives the physical signal through its network interface hardware and decodes it to reconstruct the original digital data frame. The firmware program inside the driver immediately checks the data frame by calculating the cyclic redundancy check code and comparing it with the value at the end of the frame to ensure that the data has not been corrupted during transmission. After the check passes, the firmware begins to parse the payload section of the data frame. This is a reverse process of the packing process, or deserialization: the firmware extracts binary data from different offset addresses in the payload according to the pre-set protocol specification and converts it back to the original values of the control parameters. Finally, the microcontroller or digital signal processor of the driver writes these newly parsed, up-to-date optimal control parameters into the corresponding control register addresses in its internal memory in real time.

[0068] The application further discloses a motor health management system based on whole-life-cycle monitoring, which comprises a memory, a processor, and a computer program stored in the memory and capable of running on the processor, and the processor implements the motor health management method based on whole-life-cycle monitoring according to any one of the preceding embodiments when the computer program is executed.

[0069] The processor can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, and the application does not limit this.

[0070] The memory can be an internal storage unit of the computer device, for example, a hard disk or a memory of the computer device, or an external storage device of the computer device, for example, a plug-in hard disk, a smart memory card (SMC), a secure digital card (SD), or a flash card (FC) equipped on the computer device, or a combination of the internal storage unit and the external storage device of the computer device. The memory is used to store a computer program and other programs and data required by the computer device, and can also be used to temporarily store data that has been output or will be output, which is not limited in the application.

[0071] The application further discloses a computer readable storage medium, which stores instructions, and the instructions enable a processor to be configured to perform the motor health management method based on whole life cycle monitoring when the instructions are executed by the processor.

[0072] The computer program can be stored in a machine readable medium, and the computer program includes computer program code in the form of source code, object code, an executable file, or some intermediate form, etc. The machine readable medium includes any entity or device, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal, and software distribution medium, etc. that can carry the computer program code. It should be noted that the machine readable medium includes but is not limited to the above-mentioned elements.

[0073] The computer readable storage medium stores the motor health management method based on whole life cycle monitoring in the computer readable storage medium, and is loaded and executed on the processor to facilitate storage and application of the method.

[0074] Those skilled in the art should understand that the above discussion of any embodiment is only exemplary and is not intended to limit the protection scope of the application; the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes of different aspects of one or more embodiments of the application as described above. In order to be brief, they are not provided in details.

[0075] One or more embodiments of the application are intended to cover all such alternatives, modifications and variations falling within the broad scope of the application. Therefore, any omission, modification, equivalent replacement, improvement, etc. made in the spirit and principle of one or more embodiments of the application should be included in the protection scope of the application.

Claims

1. A method for motor health management based on full life cycle monitoring, characterized in that, The method comprises the following steps: Collecting multi-source sensor data containing electrical, mechanical and thermodynamic characteristics of the motor during operation, and mapping the multi-source sensor data to a motor health state space in real time through a manifold learning model that maintains the intrinsic physical constraints of the motor system, wherein the current comprehensive health state of the motor is represented by a unique coordinate point in the motor health state space; Based on a strategy database containing operation strategies under different health states and task scenarios, a health cost function associated with the coordinate point in the motor health state space is constructed by reverse inference using an inverse reinforcement learning algorithm, which is used to quantify the irreversible health loss caused by executing different operation strategies under different health states; A dynamic economic system model is constructed by virtually associating different health-related physical components of the motor with a plurality of interrelated virtual agents, and a decision model is trained using an optimal control algorithm in combination with the health cost function, which is used to output the optimal operation strategy that minimizes the total health cost of the motor throughout its life cycle, given the coordinate point corresponding to the current comprehensive health state and future task requirements; In the actual operation of the motor, real-time multi-source sensor data is obtained and mapped to the current coordinate point in the motor health state space; The production task is received from the upper control system of the motor, the current coordinate point and the production task are combined, and the optimal operation strategy at the current time is determined using the decision model, and the optimal operation strategy is parsed into motor control parameters and sent to the motor driver for execution.

2. The full life cycle monitoring based motor health management method of claim 1, wherein, The multi-source sensor data at least includes three-phase current waveform data reflecting the electrical stress of the motor, multi-axis vibration data of the bearing seat reflecting the mechanical fatigue of the motor, winding end and bearing temperature data reflecting the internal thermal stress state of the motor, and shell temperature data reflecting the external thermal stress state of the motor.

3. The full life cycle monitoring based motor health management method of claim 1, wherein, The manifold learning model that maintains the intrinsic physical constraints of the motor system to map the multi-source sensor data to the motor health state space in real time comprises the following steps: Obtain historical multi-source sensor data of the motor under different health stages and operating conditions, and pre-process the historical multi-source sensor data into high-dimensional feature vectors; Design an autoencoder neural network comprising an encoder and a decoder as a manifold learning model, wherein the encoder is used to compress the high-dimensional feature vectors into a pre-set motor health state space, and the decoder is used to reconstruct the coordinate point in the motor health state space into a high-dimensional feature vector; Construct a combined loss function comprising a reconstruction error term and a symplectic structure penalty term, wherein the reconstruction error term is used to measure the difference between the original high-dimensional feature vector and the reconstructed high-dimensional feature vector, and the symplectic structure penalty term is used to measure the deviation of the motor health state space from the pre-set intrinsic physical constraints; Train the manifold learning model using the historical multi-source sensor data and the combined loss function until the manifold learning model converges; Map the high-dimensional feature vector formed by the real-time collected multi-source sensor data to the coordinate point in the motor health state space using the trained manifold learning model.

4. The full life cycle monitoring based motor health management method of claim 3, wherein, The construction and calculation of the symplectic structure penalty term comprises the following steps: In the process of mapping high-dimensional feature vectors to the motor health state space by the encoder in the manifold learning model, a Jacobian matrix of the mapping transformation is calculated; A standard symplectic form matrix corresponding to the dimension of the motor health state space is defined; A metric matrix is obtained by matrix operation combining the Jacobian matrix and the standard symplectic form matrix; The difference between the metric matrix and the standard symplectic form matrix is taken as the quantization result of the symplectic structure penalty term; The quantization result of the symplectic structure penalty term is multiplied by a preset weight coefficient and added to the reconstruction error term to form a combined loss function.

5. The full life cycle monitoring based motor health management method of claim 3, wherein, The method further comprises the following steps: The strategy database containing multiple strategy samples is established, and each strategy sample contains a health state defined by a coordinate point in the motor health state space, a task scenario, and a corresponding selected motor control parameter; The motor control parameters in the strategy database are abstracted into discrete exclusive operation strategies; The coordinate points in the motor health state space are taken as states, and the exclusive operation strategies are taken as actions to construct a Markov decision process; The maximum entropy inverse reinforcement learning algorithm is applied to solve the cost distribution of state-action pairs from the Markov decision process and the strategy samples through iterative optimization; The cost distribution of state-action pairs is fitted and generalized to construct a health cost function, which is used to input any health state coordinate point and a candidate operation strategy to output a scalar health loss cost value.

6. The full life cycle monitoring based motor health management method of claim 5, wherein, The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps:

7. The full life cycle monitoring based motor health management method of claim 6, wherein, The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: The method further comprises the following steps: acquire the real health status coordinate point actually mapped by the manifold learning model in the next moment; calculate the geometric distance between the predicted health status coordinate point and the real health status coordinate point in the motor health status space as the model prediction error; if the model prediction error exceeds the preset error threshold continuously for multiple time steps, it is determined that a model mismatch event occurs; when the model mismatch event occurs, the data collected in the recent period including the multi-source sensor data, the optimal operation strategy executed and the real health status coordinate point are stored in the preset online learning database, and the online learning database is used to fine-tune the decision model online.

8. The full life cycle monitoring based motor health management method of claim 1, wherein, The production task is received from the upper control system of the motor, the current coordinate point and the production task are combined, and the optimal operation strategy at the current moment is determined by using the decision model, and the optimal operation strategy is parsed into motor control parameters and sent to the motor driver for execution, including the following steps: a mapping table between the operation strategy and the motor control parameter set is established in advance, each operation strategy corresponds to a set of motor control parameter set including pulse width modulation frequency, current loop controller parameter, torque change rate limit and cooling system control level; the current coordinate point and the production task are input into the decision model, and when the decision model outputs the optimal operation strategy, the optimal motor control parameter set corresponding to the optimal operation strategy is retrieved through the mapping table; the optimal motor control parameter set is packaged into a data frame conforming to the preset industrial bus protocol; the data frame is sent to the motor driver through the high-speed real-time industrial bus; the data frame is received and parsed by the motor driver, and the optimal motor control parameters in the data frame are updated to the corresponding internal control register in real time. 9.A motor health management system based on full life cycle monitoring, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the computer program to realize the motor health management method based on full life cycle monitoring as claimed in any one of claims 1 to 8.

10. A computer-readable storage medium having stored thereon instructions, the computer-readable storage medium comprising: The instructions, when executed by the processor, cause the processor to be configured to perform the motor health management method based on full life cycle monitoring according to any one of claims 1 to 8. The instructions, when executed by the processor, cause the processor to be configured to perform the motor health management method based on full life cycle monitoring according to any one of claims 1 to 8.

Images