A method for predicting and protecting an asynchronous motor from overheating

Abstract

This invention discloses a method for predicting and protecting the overheating risk of asynchronous motors, specifically relating to the field of motor control and protection technology. Based on an intelligent fusion model, it estimates temperature and thermal stress field, calculates the rate of change of thermal stress, non-uniformity, and hotspot trends, and calculates a dynamic health index using historical data. These parameters are then input into a multi-objective reinforcement learning controller to optimize long-term health and short-term performance, generating a thermal shaping control vector to regulate the motor. This invention combines physical mechanisms with data-driven approaches through an intelligent fusion model, utilizing a graph neural network to learn the structure of the heat conduction graph and verify physical laws, thereby improving the accuracy of thermal state estimation. It achieves a multi-dimensional risk characterization combining transient impact and cumulative effects through the rate of change of thermal stress, non-uniformity, hotspot trends, and dynamic health index. By optimizing long-term health and short-term performance losses, a thermal shaping control vector is generated to achieve regulation from passive protection to active prevention, extending the motor's service life.

Description

Technical Field

[0001] This invention relates to the field of motor control and protection technology, and more specifically, to a method for predicting and protecting the overheating risk of asynchronous motors. Background Technology

[0002] As a widely used power device in industrial fields, the operational reliability of asynchronous motors directly affects the stability of production systems. Overheating is one of the main causes of motor insulation aging, efficiency decline, and even sudden failures. Therefore, accurate prediction and effective protection against motor overheating risks are of significant engineering importance.

[0003] With the development of sensing technology and intelligent algorithms, some researchers have attempted to use data-driven methods for motor temperature prediction. For example, they have established mapping relationships between operating parameters such as current and speed and temperature using neural networks, or employed infrared thermal imaging technology to obtain the surface temperature distribution of the motor. Data-driven methods can learn temperature change patterns from historical operating data, but their predictions may suffer from insufficient generalization under conditions outside the coverage of the training data. Embedded temperature sensors can directly measure winding temperature and are mainly used for local monitoring of stator windings; however, monitoring the temperature of rotating components such as the rotor remains challenging.

[0004] At the protection strategy level, existing technologies mainly employ threshold-triggered protection methods, executing power-off protection when the monitored temperature or heat accumulation exceeds a preset threshold. While this protection method can prevent serious motor failures, it is a reactive mechanism and struggles to predict and proactively intervene in potential future thermal risks. Some studies have attempted to use temperature as a feedback quantity to drate the motor, balancing heat generation and dissipation through a pre-set derating curve; however, further research is needed in multi-objective optimization. Therefore, this invention proposes a method for predicting and protecting the overheating risk of asynchronous motors to address the aforementioned problems. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides the following technical solution: A method for predicting and protecting the overheating risk of an asynchronous motor includes the following steps: Real-time operating data of the asynchronous motor is collected and input into the intelligent fusion model to estimate the temperature distribution and thermal stress field distribution of each preset key component inside the motor in real time. The intelligent fusion model includes a simplified physical model and a graph neural network. The graph neural network is used to learn and characterize the heat conduction graph structure between each component inside the motor. The thermal stress change rate and thermal stress field non-uniformity are calculated based on the thermal stress field distribution, and the hot spot temperature change trend in the near future is predicted based on the temperature distribution. Obtain historical thermal stress data of the motor and calculate the dynamic health index of the motor by combining it with a preset aging model; The thermal stress change rate, thermal stress field non-uniformity, hot spot temperature change trend and dynamic health index are used as inputs. The optimization decision is made by a multi-objective reinforcement learning controller. The optimization objective of the multi-objective reinforcement learning controller is to maximize the long-term health of the motor and minimize the short-term performance loss. The thermal shaping control vector is generated based on the optimization decision, and the motor is regulated based on the thermal shaping control vector.

[0006] In a preferred embodiment, the simplified physical model is constructed based on the thermal network method, and its specific application includes the following steps: The internal structure of the motor is divided into multiple temperature nodes based on its geometry and material properties. These temperature nodes include stator winding nodes, stator core nodes, rotor bar nodes, rotor core nodes, housing nodes, and end airflow nodes. Based on heat conduction and convection cooling mechanisms, thermal conductivity parameters are established between each temperature node. These parameters include conductive and convective thermal conductivity. Conductive thermal conductivity is determined based on thermal conductivity, heat dissipation area, and heat transfer path length, while convective thermal conductivity is determined based on surface heat dissipation coefficient and heat dissipation surface area. A heat source matrix is ​​established based on the loss distribution corresponding to each temperature node. This loss distribution includes stator copper losses, rotor aluminum losses, core losses, and mechanical losses. Based on the thermal conductivity parameters and the heat source matrix, the heat balance equations for each temperature node are established. The heat balance equation is expressed as: heat flux of the node is equal to the heat flux of the node multiplied by the heat conductance of the node itself minus the heat conductance of all mutual nodes multiplied by the temperatures of adjacent nodes. The heat balance equations of all temperature nodes are combined to form a set of heat balance equations and solved to obtain the temperature value of each temperature node. Based on the temperature difference between adjacent temperature nodes and the thermal conductivity parameters, the instantaneous heat flux density on each heat transfer path is calculated. The heat flux density is determined by the thermal conductivity multiplied by the node temperature difference. Based on the rate of change of temperature value over time, the temperature rise rate of each temperature node is calculated.

[0007] In a preferred embodiment, the graph neural network works as follows: The temperature values, temperature rise rates, and heat flux densities along each heat transfer path obtained from the simplified physical model calculations are used as the initial node features and edge features of the graph neural network. The graph neural network uses each preset key component as a graph node and the heat conduction path between each preset key component as a graph edge to construct the heat conduction graph structure of the motor. The current, voltage, and speed signals from real-time operating data are used as input features of the graph neural network and fused with node features and edge features. Graph neural networks are trained through supervised learning, using measured temperature data or high-precision simulated temperature data as training labels, to learn the residual mapping relationship between the temperature value output by the simplified physical model and the actual temperature value. The correction value output by the graph neural network is added to the output of the simplified physical model to obtain the temperature distribution after preliminary correction.

[0008] In a preferred embodiment, after the graph neural network completes supervised learning training and obtains a pre-corrected temperature distribution, the physical laws of the pre-corrected temperature distribution are verified through the following methods: By leveraging the thermal coupling relationship between key components over long distances established by multiple stacked graph convolutional layers in a graph neural network, feature propagation results based on heat conduction paths between graph nodes are extracted. Physical constraints are introduced during the training of graph neural networks. These constraints include residual constraints of the heat conduction equation, energy conservation constraints, and constraints of the second law of thermodynamics. The residual constraints of the heat conduction equation are constructed based on the temperature gradient divergence on the graph nodes. The energy conservation constraints are used to ensure that the sum of the heat flux of each graph node matches the total heat generation. The constraints of the second law of thermodynamics are used to ensure that heat spontaneously flows from high-temperature nodes to low-temperature nodes. The temperature distribution after preliminary correction is guided by physical constraints, so that the final temperature distribution output by the graph neural network satisfies both the data-driven correction result and the basic laws of thermodynamics. Based on the final output temperature distribution, the thermal stress field distribution is calculated by combining the thermal expansion coefficient and elastic modulus of the materials of each preset key component. The thermal stress field distribution includes normal stress and shear stress components in each direction.

[0009] In a preferred embodiment, the calculation of the rate of change of thermal stress and the non-uniformity of the thermal stress field based on the thermal stress field distribution is achieved in the following way: The thermal stress field distribution of each preset key component is represented as a tensor field containing multiple stress components. The trajectory curvature of the tensor field in phase space is calculated as the thermal stress change rate. The trajectory curvature is determined by dividing the magnitude of the vector product of the second and first derivatives of the thermal stress tensor with respect to time by the cube of the magnitude of the first derivative. The thermal stress field distribution is mapped to a multi-scale fractal space, the fractal dimension of the thermal stress contour at each scale is calculated, and the fluctuation amplitude of the fractal dimension with scale is used as the first component of the thermal stress field inhomogeneity. Based on the heat conduction diagram structure between each preset key component, the ratio of the number of boundaries where the thermal stress gradient between adjacent diagram nodes exceeds a preset threshold to the total number of boundaries is used as the second component of thermal stress field non-uniformity. The first component and the second component are geometrically weighted and fused to obtain the thermal stress field non-uniformity, wherein the exponent of the geometric weighting is adaptively determined according to the thermal stress field distribution characteristics under the current working condition. Based on the temperature distribution, the future temperature change trajectory of each preset key component is reconstructed in phase space, and the Lyapunov exponent of the reconstructed trajectory is calculated. The number of dimensions with a Lyapunov exponent greater than zero is used as a representation of the future temperature change trend of hotspots.

[0010] In a preferred embodiment, the dynamic health index of the motor is calculated in the following manner: The thermal stress change rate and thermal stress field inhomogeneity in historical thermal stress data are mapped to a pre-defined aging model in the cumulative damage space. The aging model is constructed as follows: Based on the nonlinear cumulative damage model in the material fatigue theory, the thermal stress change rate is used as the damage driving force, and the thermal stress field non-uniformity is used as the damage coupling coefficient. The damage coupling coefficient is determined according to the spatial distribution of the thermal stress gradient between each preset key component. When the thermal stress field non-uniformity increases, the damage coupling coefficient increases monotonically, and the rate of increase increases with the increase of non-uniformity. Historical thermal stress data is decomposed into a continuous sequence of thermal shock events. Each thermal shock event corresponds to a period in which the rate of change of thermal stress exceeds a preset threshold. The basic damage amount of a single thermal shock event is calculated based on the peak value of the rate of change of thermal stress during that period, and the corresponding damage coupling coefficient is determined based on the average thermal stress field non-uniformity during that period. The basic damage values ​​of all thermal shock events prior to the current moment are nonlinearly accumulated after being adjusted by the corresponding damage coupling coefficients. The nonlinear accumulation adopts a damage history-dependent accumulation method, that is, the damage increment of subsequent events increases with the increase of the previous accumulated damage value. The dynamic health index of the motor is calculated based on the ratio between the cumulative damage value and the preset failure threshold. The dynamic health index is expressed as a percentage and decreases as the cumulative damage value increases. When the cumulative damage value reaches the failure threshold, the health index returns to zero.

[0011] In a preferred embodiment, optimization decisions are made using a multi-objective reinforcement learning regulator in the following manner: Construct the state space of a multi-objective reinforcement learning controller. The state space includes the current thermal stress change rate, thermal stress field non-uniformity, hot spot temperature change trend, dynamic health index, and the current operating parameters of the motor, including speed, torque, and current. The action space of a multi-objective reinforcement learning controller is constructed. The action space includes the adjustment amount of the reference torque command, the amplitude and phase of the harmonic current injection, and the modulation amplitude of the switching frequency. Each action component is continuously adjustable within a preset safety range. Construct a reward function for a multi-objective reinforcement learning regulator, which includes a long-term health reward term and a short-term performance loss penalty term. A deep deterministic policy gradient algorithm is used to train a multi-objective reinforcement learning regulator, enabling the regulator to learn a nonlinear mapping relationship from the state space to the action space. During the training process, the goal is to maximize the cumulative discount reward, which is the weighted sum of the reward values ​​at each future time step, with the weights decaying over time. During real-time control, the current state space is input into the trained multi-objective reinforcement learning controller, and the controller outputs the corresponding action space components.

[0012] In a preferred embodiment, the long-term health reward is determined based on the change in the dynamic health index after the action is executed. A positive reward is obtained when the rate of increase or decrease of the dynamic health index slows down, and a negative reward is obtained when the rate of decrease of the dynamic health index accelerates. The short-term performance loss penalty is determined based on the absolute value of the deviation between the actual output torque and the required torque after the action is executed.

[0013] In a preferred embodiment, generating the thermal shaping control vector based on the optimization decision refers to: The reference torque command adjustment in the motion space component is superimposed with the current fundamental torque command to obtain the updated reference torque command; the harmonic current injection amplitude and phase in the motion space component are synthesized with the current fundamental current command to obtain the complete harmonic current injection strategy; the switching frequency modulation amplitude in the motion space component is superimposed with the current reference switching frequency to obtain the updated switching frequency modulation strategy; the updated reference torque command, the complete harmonic current injection strategy, and the updated switching frequency modulation strategy together constitute the thermal shaping control vector.

[0014] The technical effects and advantages of this invention are as follows: This invention constructs an intelligent fusion model that combines a simplified physical model and a graph neural network, integrating physical mechanisms with a data-driven approach. The simplified physical model uses a thermal network method to rapidly calculate the temperature values, temperature rise rates, and heat flux densities at each temperature node, providing benchmark calculation results that conform to fundamental thermodynamic laws. The graph neural network constructs a motor heat conduction graph structure with preset key components as graph nodes and heat conduction paths as graph edges. Through graph convolution operations, it learns the thermal coupling relationships between components and performs error correction and physical law verification on the output of the simplified physical model. During the correction process, residual constraints of the heat conduction equation, energy conservation constraints, and the second law of thermodynamics are introduced. This ensures that the final estimated temperature distribution and thermal stress field distribution closely approximate measured temperature data or high-precision simulated temperature data, while strictly satisfying physical laws such as heat flowing from high-temperature nodes to low-temperature nodes and the sum of heat fluxes at each node matching the total heat generation. This effectively overcomes the problem of fixed parameters in pure physical models being difficult to adapt to changes in operating conditions and avoids the risk of physically unreasonable results that may arise from data-driven methods outside the coverage of training data. Therefore, it significantly improves the accuracy and reliability of thermal state estimation, providing accurate basic data for subsequent overheat risk prediction.

[0015] This invention transforms the static thermal stress field distribution into dynamic characteristic parameters with clear physical meaning by calculating the rate of change of thermal stress and the non-uniformity of the thermal stress field based on the thermal stress field distribution, and predicting the hot spot temperature change trend in the short term based on the temperature distribution. The rate of change of thermal stress is obtained by calculating the trajectory curvature of the thermal stress tensor in phase space, reflecting the severity of thermal shock and the dynamic evolution characteristics of the stress state. The non-uniformity of the thermal stress field is obtained by integrating two components: the fluctuation amplitude of the multi-scale fractal space and the gradient boundary ratio based on the heat conduction diagram structure. This comprehensively characterizes the uniformity of heat distribution in space and the stress synergy effect between components. The hot spot temperature change trend is determined by analyzing the temperature distribution of each preset key component. The temperature time series is reconstructed in phase space, and the number of dimensions with Lyapunov exponents greater than zero is calculated to characterize the degree of chaos and the predictability of future evolution of the temperature dynamic system. At the same time, historical thermal stress data of the motor is obtained and combined with a preset aging model to calculate the dynamic health index. The peak value of the thermal stress change rate of each thermal shock event is used as the damage driving force, and the non-uniformity of the thermal stress field is used as the damage coupling coefficient for nonlinear accumulation. This allows the health index to reflect the cumulative consumption of motor life by the history of thermal stress. It realizes a multi-dimensional quantitative characterization of motor overheating risk from transient impact to steady-state accumulation, and from local hot spots to global distribution, providing rich and physically meaningful input information for optimization decisions.

[0016] This invention uses thermal stress change rate, thermal stress field inhomogeneity, hot spot temperature change trend, and dynamic health index as inputs, and employs a multi-objective reinforcement learning controller for optimization decision-making. It constructs a state space containing operating parameters such as speed, torque, and current, and a continuous action space containing the reference torque command adjustment, harmonic current injection amplitude and phase, and switching frequency modulation amplitude. A reward function is designed that balances long-term health rewards and short-term performance penalty terms. A deep deterministic policy gradient algorithm is used to train the controller to learn the nonlinear mapping relationship from the current thermal state to the optimal control action. During real-time control, a thermally shaped control vector is generated based on the optimization decision. This vector is then used to compare the reference torque command adjustment with the current base torque. The updated reference torque command is obtained by superimposing commands to adjust the output level. The harmonic current injection amplitude and phase are synthesized with the fundamental current command to obtain a complete harmonic current injection strategy to actively change the internal loss distribution of the motor. The switching frequency modulation amplitude is superimposed with the current reference switching frequency to obtain an updated switching frequency modulation strategy to balance the thermal load of the inverter and the motor. This realizes the transformation from a threshold response mode that passively waits for the power-off protection after the temperature exceeds the limit to a preventive control mode that actively adjusts the control command dynamically according to the thermal state prediction results to prevent overheating. It can maintain the performance output as much as possible while ensuring the safe operation of the motor, and reduce the production losses caused by unplanned downtime while extending the overall service life of the motor. Attached Figure Description

[0017] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a schematic diagram of an asynchronous motor overheating risk prediction and protection method according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0019] Reference Figure 1 The following examples were obtained: Example 1: A method for predicting and protecting the overheating risk of an asynchronous motor, comprising the following steps: Real-time operating data of the asynchronous motor is collected and input into the intelligent fusion model to estimate the temperature distribution and thermal stress field distribution of each preset key component inside the motor in real time. The intelligent fusion model includes a simplified physical model and a graph neural network. The graph neural network is used to learn and characterize the heat conduction diagram structure between the components inside the motor. Through this step, a real-time perception capability of the motor's thermal state is established, and the internal temperature field and stress field, which are difficult to measure directly, are transformed into calculable quantitative data, providing a basic input for subsequent risk assessment.

[0020] The thermal stress change rate and thermal stress field non-uniformity are calculated based on the thermal stress field distribution. The hot spot temperature change trend is predicted based on the temperature distribution in the short term. This step transforms the static thermal stress field distribution into dynamic characteristic parameters. The thermal stress change rate reflects the severity of thermal shock, the thermal stress field non-uniformity reflects the balance of heat distribution, and the hot spot temperature change trend reveals the direction of temperature evolution in the short term. These parameters together constitute a multi-dimensional characterization of the thermal risk of the motor.

[0021] Historical thermal stress data of the motor is acquired, and the dynamic health index of the motor is calculated by combining it with a preset aging model. This step maps instantaneous thermal stress events to the cumulative damage space, so that the health index can reflect the cumulative consumption of motor life by historical thermal stress. The dynamic health index decreases with the accumulation of thermal shock events, providing a long-term health dimension reference for optimization decisions.

[0022] Using thermal stress change rate, thermal stress field non-uniformity, hot spot temperature change trend, and dynamic health index as inputs, a multi-objective reinforcement learning controller makes optimization decisions. The optimization objective of the multi-objective reinforcement learning controller is to maximize the long-term health of the motor and minimize the weighted sum of short-term performance loss. This step seeks a balance among multiple conflicting objectives, enabling the controller to autonomously decide how to delay the thermal aging process while ensuring performance based on the current thermal and health status.

[0023] The thermal shaping control vector is generated based on the optimization decision, and the motor is regulated based on the thermal shaping control vector. This step transforms the optimization decision into specific execution instructions. The thermal shaping control vector actively changes the heat distribution and heat flow path inside the motor by adjusting the torque command, injecting harmonic current, and modulating the switching frequency, so that the thermal stress field is always maintained within the preset range, realizing closed-loop control from passive protection to active prevention.

[0024] The simplified physical model is constructed based on the thermal network method, and its specific application includes the following steps: The internal structure of the motor is divided into multiple temperature nodes based on its geometry and material properties. These temperature nodes include stator winding nodes, stator core nodes, rotor bar nodes, rotor core nodes, housing nodes, and end airflow nodes. This step discretizes a continuous physical entity into a finite number of representative points. Each temperature node corresponds to the thermal state of a specific region inside the motor. For example, a stator winding node can represent the average temperature of a portion of the winding conductors. These nodes correspond to predefined key components, but a predefined key component may be divided into multiple temperature nodes to achieve a more refined description of the temperature distribution. For instance, the stator winding can be divided into multiple nodes along the axial or radial direction.

[0025] Based on the mechanisms of heat conduction and convection, thermal conductivity parameters are established between each temperature node. These parameters include conductive and convective thermal conductance. Conductive thermal conductance is determined by thermal conductivity, heat dissipation area, and heat transfer path length. For example, for a heat conduction path between two adjacent solid nodes, conductive thermal conductance equals thermal conductivity multiplied by the cross-sectional area perpendicular to the heat flow direction, divided by the heat transfer path length. Here, thermal conductivity depends on material properties, while the cross-sectional area and path length are determined by the geometry. Convective thermal conductance is determined by the surface heat dissipation coefficient and heat dissipation surface area. For example, for a convective heat transfer path between a solid node and a fluid node, convective thermal conductance equals surface heat dissipation coefficient multiplied by the surface area participating in convection. The surface heat dissipation coefficient is affected by factors such as fluid velocity and physical properties.

[0026] A heat source matrix is ​​established based on the loss distribution corresponding to each temperature node. The loss distribution includes stator copper loss, rotor aluminum loss, core loss, and mechanical loss. For example, the loss corresponding to the stator winding node is mainly the copper loss of that part of the winding, calculated based on the effective value of the current flowing through the winding and the winding resistance; the loss corresponding to the rotor bar node is the rotor aluminum loss, determined by the rotor current and rotor resistance; the loss corresponding to the core node is the iron loss, calculated based on magnetic flux density, frequency, and material loss curves; mechanical loss is usually applied to the rotor core node or bearing node, determined based on the rotational speed and friction coefficient. The loss value of each node constitutes the corresponding element in the heat source matrix.

[0027] Based on the thermal conductivity parameters and the heat source matrix, the heat balance equations for each temperature node are established. For any temperature node, its heat balance equation is expressed as the sum of the thermal conductances connecting this node and its adjacent nodes multiplied by the temperature difference between this node and its corresponding adjacent node, equaling the heat flux value in the heat source matrix corresponding to this node. The thermal conductance connecting this node and its adjacent node is determined as either conductive or convective thermal conductance based on the nature of the heat transfer path. Conductive thermal conductance is used when both nodes are solid nodes, and convective thermal conductance is used when one of the nodes is a fluid node. The heat balance equations for all temperature nodes are combined to form a linear system of equations about the temperature of each node, which is then solved to obtain the temperature value of each temperature node. Based on the temperature difference between adjacent temperature nodes and the thermal conductances connecting the two nodes, the instantaneous heat flux density on each heat transfer path is calculated. For conductive paths, the heat flux density is determined by multiplying the conductive thermal conductance by the node temperature difference and then dividing by the heat transfer area. For convective paths, the heat flux density is determined by multiplying the convective thermal conductance by the node temperature difference and then dividing by the convective surface area. Based on the rate of change of temperature value over time, the temperature rise rate of each temperature node is calculated.

[0028] The thermal balance equations for all temperature nodes are combined to form a system of thermal balance equations, which is then solved to obtain the temperature values ​​for each node. This system of linear equations can be solved using either a direct method or an iterative method. The resulting temperature values ​​are the steady-state temperatures of each node under the current operating conditions. For example, the temperatures at locations such as stator winding nodes and rotor bar nodes can be obtained by solving this system.

[0029] Based on the temperature difference and thermal conductivity parameters between adjacent temperature nodes, the instantaneous heat flux density on each heat transfer path is calculated. The heat flux density is determined by multiplying the thermal conductivity by the node temperature difference. For conduction paths, the heat flow rate on that path is obtained by multiplying the conduction thermal conductivity by the node temperature difference, and this heat flow rate divided by the heat transfer area is the conduction heat flux density. For convection paths, the convective heat transfer rate is obtained by multiplying the convective thermal conductivity by the node temperature difference, and this heat transfer rate divided by the convective surface area is the convective heat flux density. For example, the conduction thermal conductivity between the stator winding node and the stator core node is multiplied by the temperature difference between the two to obtain the heat flow rate from the winding to the core, which is then divided by the contact area to obtain the heat flux density.

[0030] The temperature rise rate at each temperature node is calculated based on the rate of change of temperature over time. By continuously collecting temperature values ​​at different times, the rate of change of temperature at each node is obtained by differentiating with respect to time. For example, the rate of temperature rise at the stator winding node reflects how fast the temperature rises over time at that location. This parameter is subsequently used to calculate the rate of change of thermal stress.

[0031] The working principle of graph neural networks is as follows: The temperature values, temperature rise rates, and heat flux densities along each heat transfer path obtained from the simplified physical model calculations are used as the initial node features and edge features of the graph neural network. The node features are the temperature values ​​and temperature rise rates of the set of temperature nodes corresponding to each preset key component. For example, the initial feature vector of a stator winding node is 85 degrees Celsius and the temperature rise rate is 0.5 degrees Celsius per second. The edge features are the heat flux densities along the heat transfer path connecting two temperature nodes. For example, the initial feature vector of an edge is 1200 watts per square meter from the stator winding node to the stator core node. These physical quantities output by the simplified physical model serve as the basic inputs to the graph neural network, providing prior physical information for the subsequent learning process.

[0032] The graph neural network constructs a heat transfer graph structure for the motor by using each preset key component as a graph node and the heat conduction path between each preset key component as a graph edge. Each preset key component corresponds to a graph node. For example, the stator core is one graph node and the rotor bars are another graph node. If there is a physical contact or heat radiation path between two preset key components, a graph edge is constructed between the corresponding graph nodes. For example, if there is a conduction heat transfer path between the stator core and the casing, a connection is established between the graph nodes corresponding to the two. This graph structure completely describes the network topology of heat transfer inside the motor.

[0033] The current, voltage, and speed signals from real-time operating data are used as input features of the graph neural network and fused with node and edge features. The current, voltage, and speed signals reflect the current external operating conditions and electrical state of the motor. For example, the collected three-phase current RMS value of 30 amps, line voltage of 380 volts, and speed of 1450 rpm are used as additional feature vectors and concatenated with existing node and edge features or merged through a specific fusion layer. This allows the graph neural network to simultaneously utilize the physical model output and real-time operating condition information for comprehensive analysis.

[0034] Graph neural networks are trained through supervised learning, using measured temperature data or high-precision simulated temperature data as training labels. They learn the residual mapping relationship between the temperature value output by the simplified physical model and the actual temperature value. During training, a large number of simplified physical model outputs under historical operating conditions are combined with corresponding measured or high-precision simulated temperatures to form training sample pairs. The graph neural network continuously adjusts its internal weight parameters through the backpropagation algorithm, so that the residual prediction value output by the network gradually approaches the actual residual, that is, the difference between the actual temperature and the temperature output by the simplified physical model. For example, if the simplified physical model outputs a stator winding temperature of 82 degrees Celsius for a certain operating condition, while the measured temperature is 85 degrees Celsius, the residual is 3 degrees Celsius. The graph neural network learns to output this 3-degree Celsius residual value under the same input conditions.

[0035] The correction value output by the graph neural network is added to the output of the simplified physical model to obtain the preliminary corrected temperature distribution. For example, the simplified physical model outputs 82 degrees Celsius for the stator winding node mentioned above, and the residual correction value output by the graph neural network is 3 degrees Celsius. The two are added together to get 85 degrees Celsius as the preliminary corrected temperature of the node. This process is repeated for all nodes to obtain the preliminary corrected temperature distribution of the entire motor. This temperature distribution has integrated the fast calculation capability of the physical model and the error correction capability of the data-driven model.

[0036] Based on the supervised learning training of the graph neural network and the acquisition of the initially corrected temperature distribution, the physical laws of the initially corrected temperature distribution are verified through the following methods: By leveraging the thermal coupling relationships established between key components over long distances through multiple stacked graph convolutional layers in a graph neural network, the feature propagation results between graph nodes based on heat conduction paths are extracted. The graph neural network, by stacking multiple graph convolutional layers, enables each graph node to aggregate feature information from multi-hop neighbor nodes. For example, the first graph convolutional layer aggregates temperature information from adjacent stator core nodes for stator winding nodes, and the second graph convolutional layer further aggregates temperature information from casing nodes adjacent to the stator core. After multiple stacking layers, an indirect thermal coupling relationship is established between stator winding nodes and end airflow nodes that originally had no direct heat conduction path. This long-distance thermal coupling relationship established through multi-layer feature propagation reflects the comprehensive effect of heat transfer along multiple paths within the motor. The feature propagation results between graph nodes based on heat conduction paths are extracted as the input basis for subsequent physical constraints.

[0037] Physical constraints are introduced during the training of graph neural networks. These constraints include residual constraints on the heat conduction equation, energy conservation constraints, and constraints based on the second law of thermodynamics. The residual constraints on the heat conduction equation are constructed based on the temperature gradient divergence at the graph nodes. For each node, the second spatial derivative (Laplace operator) of the surrounding temperature field is calculated. The heat conduction equation requires that the sum of the Laplace operator of the temperature field and the heat source term equals zero. The residual constraint is the difference between the left and right sides of this equation. For example, for a stator core node, the temperature gradient divergence is calculated based on the temperature distribution of its neighboring nodes. If the calculated value does not match the actual heat generation of that node, a residual constraint loss occurs. The energy conservation constraint ensures that the sum of the heat fluxes at each graph node matches the total heat generation. Specifically, it is calculated by summing all the heat fluxes flowing into and out of the graph nodes. This sum should equal the total motor losses. For example, the conductive and convective heat fluxes between all preset key components are added together and then compared with the sum of the heat source matrices of each node. The difference between the two is the loss term for the energy conservation constraint. The second law of thermodynamics is used to ensure that heat flows spontaneously from high-temperature nodes to low-temperature nodes. For each heat conduction path, the consistency between the temperature gradient direction and the heat flow direction is checked. If an abnormal situation occurs where heat flows from a low-temperature node to a high-temperature node, a constraint loss occurs. For example, when the temperature of the stator winding node is higher than that of the rotor bar node, the heat flow direction should be from the stator to the rotor. If the temperature distribution output by the graph neural network shows the opposite temperature gradient, a constraint penalty is triggered.

[0038] The physical constraints guide the initially corrected temperature distribution, ensuring that the final temperature distribution output by the graph neural network simultaneously satisfies both the data-driven correction result and fundamental thermodynamic laws. During training, these physical constraints are added as regularization terms to the total loss function, working in conjunction with the residual loss from supervised learning for optimization. For example, the total loss function can be represented as a weighted sum of the supervised learning loss plus the residual constraint loss from the heat conduction equation, the energy conservation constraint loss, and the second law of thermodynamics constraint loss. During backpropagation, the network minimizes these loss terms, ensuring that the final output temperature distribution approximates the measured data without violating fundamental physical laws. For instance, the physical constraints for operating conditions not covered by the training data prevent the network from outputting thermodynamically unreasonable temperature distributions.

[0039] Based on the final output temperature distribution, the thermal stress field distribution is calculated by combining the thermal expansion coefficients and elastic moduli of the materials of each preset key component. The thermal stress field distribution includes normal stress and shear stress components in each direction. The thermal expansion coefficient of a material is defined as the elongation per unit length of the material per degree Celsius increase in temperature. Different materials in the motor, such as copper, silicon steel, cast aluminum, and insulating materials, have their own thermal expansion coefficients, which can be found in material handbooks or obtained through experimental testing with a thermal expansion meter. For example, the thermal expansion coefficient of copper conductors in the stator winding is approximately 17 × 10⁻⁶. -6Each degree Celsius. The elastic modulus is defined as the ratio of stress to strain within the elastic deformation range of a material. It can also be obtained through material handbooks or tensile tests. For example, the elastic modulus of silicon steel sheets is approximately 200 gigapascals. The thermal stress field calculation is based on the thermoelasticity theory. The temperature change of each node relative to the reference temperature is calculated based on the final output temperature distribution. The thermal strain is obtained by multiplying it by the coefficient of thermal expansion. Then, the thermal stress is calculated from the thermal strain based on the elastic modulus and Poisson's ratio. The specific calculation method includes the normal stress and shear stress components in each direction. For example, in a rectangular coordinate system, the normal stress includes σxx, σyy, and σzz, which represent the normal stress along the x, y, and z directions, respectively. The shear stress includes τxy, τyz, and τzx, which represent the shear stress in the xy plane, yz plane, and zx plane, respectively. The complete thermal stress field distribution inside each preset key component and at the interface is obtained by solving the finite element method or constitutive equations. For example, the radial normal stress and axial shear stress at the mating surface between the stator core and the housing.

[0040] The calculation of the rate of change of thermal stress and the non-uniformity of the thermal stress field based on the thermal stress field distribution is achieved as follows: The thermal stress field distribution of each preset key component is represented as a tensor field containing multiple stress components. The curvature of the trajectory of the tensor field in phase space is calculated as the rate of change of thermal stress. The trajectory curvature is determined by dividing the magnitude of the vector product of the second and first derivatives of the thermal stress tensor with respect to time by the cube of the magnitude of the first derivative. The thermal stress field distribution itself is a second-order tensor field with multiple stress components at each spatial location and at each time. For example, for a stator core node, it contains normal stress components σxx, σyy, σzz and shear stress components τxy, τyz, τzx. These components constitute a point in a six-dimensional space. As time progresses, this point moves in the six-dimensional space, forming a trajectory. The trajectory curvature characterizes the degree of bending of the trajectory in phase space, reflecting the severity of the change in thermal stress state. In curvature calculations, the denominator is the cube of the magnitude of the first derivative of the thermal stress tensor with respect to time. When the magnitude of the first derivative approaches zero, the denominator may become zero. To avoid this, in numerical calculations, when the magnitude of the first derivative is less than a preset velocity magnitude threshold, the curvature is forcibly set to zero. This preset velocity magnitude threshold is set, for example, to 10, based on the accuracy of the numerical calculation and the actual physical meaning of the thermal stress change. -6 Megapascals per second.

[0041] The thermal stress field distribution is mapped to a multi-scale fractal space, and the fractal dimension of the thermal stress contour at each scale is calculated. The fluctuation amplitude of the fractal dimension with scale is used as the first component of the thermal stress field inhomogeneity. Multi-scale fractal space refers to observing the thermal stress field distribution at different observation scales. For example, thermal stress contours on the motor cross-section are sampled and their contours extracted at different scales such as 1 mm, 2 mm, and 5 mm. At each scale, the box counting method is used to calculate the fractal dimension of the thermal stress contour at that scale. The fractal dimension reflects the complexity and space-filling ability of the contour. As the scale increases, the fractal dimension fluctuates. The larger the fluctuation amplitude, the worse the self-similarity of the thermal stress field at different scales, i.e., the more inhomogeneous the thermal stress distribution.

[0042] Based on the heat conduction graph structure between each preset key component, the ratio of the number of boundaries where the thermal stress gradient between adjacent graph nodes exceeds a preset gradient threshold to the total number of boundaries is used as the second component of the thermal stress field non-uniformity. According to the motor heat conduction graph structure constructed by the graph neural network, each preset key component corresponds to a graph node. Graph edges connecting nodes indicate the existence of heat conduction paths. For each graph edge, the thermal stress gradient between the two graph nodes it connects is calculated. The thermal stress gradient is determined by dividing the difference in thermal stress tensors between the two nodes by the physical distance. When this gradient exceeds the preset gradient threshold, the graph edge is marked as a high gradient boundary. The ratio of the number of graph edges marked as high gradient boundaries to the total number of graph edges is counted. This ratio reflects the degree of discontinuity in the spatial distribution of thermal stress. The preset gradient threshold is set based on the material's thermal stress bearing capacity and historical fault data. For example, 50 MPa per millimeter is used as the gradient threshold for the interface between the cast iron casing and the silicon steel sheet core.

[0043] The thermal stress field nonuniformity is obtained by geometrically weighting and fusing the first and second components. The geometric weighting exponents are adaptively determined based on the thermal stress field distribution characteristics under the current operating conditions. The geometric weighting fusion uses a product form rather than an additive form, reflecting the coupling effect between the two components rather than a simple superposition. Specifically, the thermal stress field nonuniformity is equal to the first component raised to the power of α multiplied by the second component raised to the power of β, where α and β are weighting exponents and satisfy α + β = 1. The values ​​of α and β are dynamically adjusted according to the current operating conditions. For example, when the motor is operating at high speed, the thermal stress field is mainly affected by centrifugal force, and the weighting exponent α of the first component takes a larger value, such as 0.7. When the motor is under a sudden load change condition, the thermal stress field is mainly affected by thermal shock, and the weighting exponent β of the second component takes a larger value, such as 0.7.

[0044] Based on the temperature distribution, the future temperature change trajectories of each preset key component are reconstructed in phase space. The Lyapunov exponent of the reconstructed trajectory is calculated, and the number of dimensions with a Lyapunov exponent greater than zero is used as a representation of the future temperature change trend of hotspots. Specifically, for each preset key component's temperature time series T(t), a delay time parameter L and an embedding dimension M are selected to construct a state vector [T(t), T(tL), T(t-2L), ..., T(t-(M-1)L)] in an M-dimensional phase space. The delay time parameter L is determined by the time corresponding to the first decrease of the autocorrelation function to 1 / 10^-e of its initial value. The embedding dimension M is determined using the spurious nearest neighbor method, i.e., gradually increasing M until the proportion of spurious nearest neighbors decreases to near zero. The reconstructed phase space trajectory reflects the complete evolutionary information of the temperature dynamic system. The Lyapunov exponent is an existing technique in chaos theory used to quantify the average exponential divergence rate of adjacent trajectories in phase space. A positive exponent indicates that the trajectory diverges in that direction and the system is sensitive to initial conditions; a negative exponent indicates that the trajectory converges; and a zero exponent indicates that the trajectory maintains a constant distance. For the M-dimensional reconstructed phase space, a total of M Lyapunov exponents are calculated to form an exponent spectrum. During the calculation, the QR decomposition method is used to perform orthogonal triangular decomposition of the Jacobian matrix and accumulate the logarithmic factors of the diagonal elements. When the stretching factor in some directions is too small, numerical instability may occur. Therefore, a singular value truncation method is introduced. That is, during the QR decomposition process, when the absolute value of the diagonal element is less than a preset singular value threshold, such as 10... -8 When necessary, the diagonal element can be replaced with the threshold to maintain decomposition stability, or the Tikhonov regularization method can be introduced to add a regularization term to the Jacobian matrix before decomposition. The number of exponents greater than zero in the Lyapunov exponent spectrum is counted. This number represents the number of independent physical factors causing unpredictable temperature changes. A dimension greater than zero of zero indicates that the temperature change is completely deterministic and predictable. A dimension greater than zero of one indicates that there is a dominant factor, such as periodic load fluctuations causing chaotic temperature changes that can still be predicted in the short term. A dimension greater than or equal to two indicates that multiple independent factors, such as load fluctuations, deteriorating heat dissipation, and material aging, simultaneously drive the temperature towards hyperchaos, and the system is about to experience thermal runaway. This indicator can provide early warning when the temperature has not yet exceeded the limit but the dynamics have already become unstable.

[0045] The geometrically weighted index is adaptively determined based on the thermal stress field distribution characteristics under the current working conditions. Specifically, the motor speed change rate and torque change rate are collected in real time as working condition characteristic parameters. The first component weight index α and the second component weight index β are dynamically allocated through preset fuzzy inference rules, and the sum of α and β is always one. When the rate of change of rotational speed exceeds a preset speed threshold and the rate of change of torque is lower than a preset torque threshold, the motor is judged to be in a high-speed stable operating condition. At this time, the thermal stress field is mainly affected by centrifugal force, and the global fluctuation amplitude is dominant. Therefore, α is set to a larger value in the range of 0.8 to 0.9, and β is correspondingly taken as 0.2 to 0.1. When the rate of change of torque exceeds a preset torque threshold, the motor is judged to be in a sudden load change condition. At this time, the thermal shock effect is significant, and the local gradient change is dominant. Therefore, β is set to a larger value in the range of 0.8 to 0.9, and α is correspondingly taken as 0.2 to 0.1. For other intermediate operating conditions, the specific values ​​of α and β are determined by linear interpolation based on the normalized values ​​of the rate of change of rotational speed and the rate of change of torque. For example, define the normalization factor of the rate of change of rotational speed γ_n and the normalization factor of the rate of change of torque γ_T, let α = 0.5 + 0.4 × (γ_n - γ_T), β = 1 - α, to ensure that α and β are always between 0.1 and 0.9 and keep their sum to 1. This mapping relationship can also be obtained through offline simulation data calibration or online adaptive learning.

[0046] The dynamic health index of the motor is calculated by mapping the rate of change of thermal stress and the non-uniformity of the thermal stress field from historical thermal stress data to a pre-defined aging model in the cumulative damage space. Historical thermal stress data refers to the time series of all thermal stress change rates and thermal stress field non-uniformities recorded since the motor was put into operation. The cumulative damage space is an abstract space with the damage amount as the coordinate axis. The thermal stress state at each moment is mapped to this space to correspond to a damage increment. The aging model is the mathematical rule that defines this mapping relationship.

[0047] Based on the nonlinear cumulative damage model in materials fatigue theory, this study uses the rate of change of thermal stress as the damage driving force and the non-uniformity of the thermal stress field as the damage coupling coefficient. The damage coupling coefficient is determined according to the spatial distribution of the thermal stress gradient between pre-defined key components. As the non-uniformity of the thermal stress field increases, the damage coupling coefficient monotonically increases, and the rate of increase increases with increasing non-uniformity. Nonlinear cumulative damage models in materials fatigue theory, such as the Corten-Dolan model or the Manson-Halford model, assume that damage accumulation is not a simple linear superposition; subsequent damage is influenced by the degree of accumulated damage. The damage driving force refers to the fundamental cause of material performance degradation; a larger rate of change of thermal stress indicates more intense stress cycling per unit time, resulting in greater damage. The damage coupling coefficient reflects the synergistic effect of accelerated damage when multiple components are simultaneously subjected to thermal stress. For example, when the stator core and rotor bars are both under high stress, their interaction will cause the overall damage to be greater than the simple sum of their individual damages. As the non-uniformity of the thermal stress field increases, the damage coupling coefficient monotonically increases. Its functional form can be an exponential function, such as the coupling coefficient being equal to e raised to the power of k times the non-uniformity, or a power function, such as the coupling coefficient being equal to the power of n times the non-uniformity, where k and n are material constants. The increasing rate of increase with the increase of non-uniformity means that the function is a convex function. For example, when the non-uniformity increases from 0.2 to 0.3, the coupling coefficient increases from 1.2 to 1.5, and when the non-uniformity increases from 0.5 to 0.6, the coupling coefficient increases from 2.5 to 3.5, with the increment gradually increasing.

[0048] Historical thermal stress data is decomposed into a continuous sequence of thermal shock events. Each thermal shock event corresponds to a period in which the rate of change of thermal stress exceeds a preset event threshold. The basic damage amount of a single thermal shock event is calculated based on the peak value of the rate of change of thermal stress within that period, and the corresponding damage coupling coefficient is determined based on the average thermal stress field inhomogeneity within that period. A thermal shock event is defined as the complete process of the rate of change of thermal stress starting below the event threshold, exceeding the event threshold, and then falling back below the event threshold. For example, if the event threshold is set to 5 MPa / s, when the rate of change of thermal stress rises from 3 MPa / s to 8 MPa / s and then falls back to 4 MPa / s, then the period from the first time exceeding 5 MPa / s to the last time falling below 5 MPa / s constitutes a thermal shock event. The basic damage amount is calculated based on the maximum value of the rate of change of thermal stress within that period, which can be expressed using the Basquin equation. For example, the basic damage amount equals the peak value of the rate of change of thermal stress multiplied by the event duration, where P is a material constant. The average thermal stress field inhomogeneity is taken as the arithmetic mean of the thermal stress field inhomogeneity at all moments within that period, and is used to determine the damage coupling coefficient corresponding to that event.

[0049] The cumulative damage value is obtained by nonlinearly accumulating the basic damage values ​​of all thermal shock events prior to the current moment after adjusting for the corresponding damage coupling coefficients. This nonlinear accumulation uses a damage history-dependent accumulation method, meaning the damage increment of subsequent events increases with the increase of the previous accumulated damage value. The specific implementation of nonlinear accumulation can be achieved using the damage curve method or the damage interaction factor method. For example, when using the damage curve method, the effective damage increment is defined as the basic damage value multiplied by the damage coupling coefficient and then multiplied by the current accumulated damage value raised to the power of γ, where γ is a material constant greater than zero. Assuming the current accumulated damage value is 0.2, γ is 0.5, the basic damage value is 0.01, and the damage coupling coefficient is 1.2, then the effective damage increment is 0.01 multiplied by 1.2 multiplied by 0.2 raised to the power of 0.5, which is 0.012 multiplied by 0.4. 47 equals 0.00536, the new cumulative damage value is 0.20536. When the same event occurs again, the base damage is still 0.01, but the cumulative damage value has increased to 0.20536. The effective damage increment becomes 0.01 multiplied by 1.2 multiplied by 0.20536 to the power of 0.5, which is 0.012 multiplied by 0.453 equals 0.00544. The new cumulative damage value is 0.2108, which reflects the nonlinear characteristic that the increment caused by the same event gradually increases as the damage accumulates.

[0050] The dynamic health index of the motor is calculated based on the ratio between the cumulative damage value and the preset failure threshold. The dynamic health index is expressed as a percentage and decreases as the cumulative damage value increases. When the cumulative damage value reaches the failure threshold, the health index returns to zero. The preset failure threshold is determined based on the heat resistance grade and fatigue life curve of the motor's insulation material. In existing technologies, the lifespan of insulation materials typically follows a 10-degree rule, meaning that the lifespan halves for every 10-degree Celsius increase in temperature. Accelerated aging tests can obtain data on the number of cycles to failure of the insulation material under a specific thermal stress level. Combining this with the SN curve from materials mechanics, the relationship between stress amplitude and cycle number is extrapolated to the actual operating conditions of the motor, determining that the motor's lifespan ends when the cumulative damage value reaches 1. Therefore, the failure threshold is usually set to 1. The dynamic health index is calculated as 100% multiplied by the number in parentheses minus the ratio of the current cumulative damage value to the failure threshold. When the cumulative damage value is 0.2 and the failure threshold is 1, the health index is 80%. When the cumulative damage value reaches 1, the health index is zero, indicating that the motor has reached its design life and needs replacement or major overhaul.

[0051] Optimization decisions are achieved through a multi-objective reinforcement learning regulator in the following ways: A state space for a multi-objective reinforcement learning controller is constructed. This state space includes the current rate of change of thermal stress, the non-uniformity of the thermal stress field, the hot spot temperature change trend, the dynamic health index, and the current operating parameters of the motor, including speed, torque, and current. The state space serves as a window for the reinforcement learning agent to perceive environmental information. The agent determines its next action based on the values ​​in the state space. For example, at a given moment, the rate of change of thermal stress is 2.5 MPa per second, the non-uniformity of the thermal stress field is 0.35, the hot spot temperature change trend is represented by two Lyapunov exponents greater than zero, the dynamic health index is 78%, the speed is 1450 rpm, the torque is 35 N·m, and the current is 28 amperes. These values ​​constitute a seven-dimensional state vector, which is input to the controller. The controller uses this state vector to determine the current thermal risk level and health status of the motor.

[0052] An action space is constructed for a multi-objective reinforcement learning controller. This action space includes the adjustment amount of the reference torque command, the amplitude and phase of the harmonic current injection, and the modulation amplitude of the switching frequency. Each action component is continuously adjustable within a preset safety range. The action space is the set of commands that the agent can execute. The adjustment amount of the reference torque command represents the increase or decrease in torque based on the current demand torque, for example, continuously adjustable from -3 Nm to +3 Nm. The amplitude of the harmonic current injection represents the magnitude of the injected harmonic current, for example, continuously adjustable from 0 to 5 amperes. The phase of the harmonic current injection represents the phase angle of the injected harmonic relative to the fundamental frequency, for example, continuously adjustable from 0 to 360 degrees. The modulation amplitude of the switching frequency represents the increase or decrease in the inverter's switching frequency, for example, continuously adjustable from -2000 Hz to +2000 Hz. These action components together constitute a four-dimensional continuous action space. The execution of each action changes the heat distribution and heat dissipation conditions of the motor.

[0053] A reward function for a multi-objective reinforcement learning regulator is constructed, comprising a long-term health reward and a short-term performance penalty. The reward function is the core guiding principle of reinforcement learning, telling the agent what actions are good and what actions are bad. During training, the agent continuously adjusts its strategy to obtain higher cumulative rewards. The long-term health reward guides the agent to focus on the long-term lifespan of the engine, while the short-term performance penalty constrains the agent from sacrificing performance indefinitely for health. The two are weighted and summed to form the comprehensive reward value; for example, the total reward equals 0.7 times the long-term health reward minus 0.3 times the short-term performance penalty.

[0054] A deep deterministic policy gradient algorithm is used to train a multi-objective reinforcement learning regulator, enabling the regulator to learn a nonlinear mapping from the state space to the action space. The training objective is to maximize the cumulative discount reward, which is a weighted sum of reward values ​​at future time steps, with weights decaying over time. The deep deterministic policy gradient algorithm is a deep reinforcement learning algorithm suitable for continuous action spaces. Based on an actor-critic framework, it comprises four neural networks: an online actor network, a target actor network, an online critic network, and a target critic network. The online actor network outputs a deterministic action based on the current state. The online critic network evaluates the value (Q-value) of the action based on the current state and the action. The target network is used to calculate a stable target Q-value to improve training convergence. During training, historical transformation samples (current state, action, reward, and next state) are randomly sampled from the experience replay buffer. The target Q-value is calculated as the current reward plus a discount factor multiplied by the next state input to the target critic network. The target Q-value obtained by the online critic network is discounted by a factor typically between 0.95 and 0.99, indicating that the importance of future rewards decays over time. The critic network is updated by minimizing the mean squared error between the output of the online critic network and the target Q-value, while the actor network is updated by maximizing the Q-value of the online critic network's output action to the online actor network. At regular intervals, the parameters of the online network are soft-updated to the target network. For example, the learning rate is set to 0.001, the discount factor to 0.98, the soft update coefficient to 0.005, the experience replay cache capacity to 10,000 records, and the batch size to 64 samples. After hundreds of thousands of iterations of training, the actor network learns the optimal mapping relationship from state to action.

[0055] During real-time control, the current state space is input into the trained multi-objective reinforcement learning controller, which outputs the corresponding action space components. No network parameter updates are performed during real-time control; only forward propagation calculations are executed. For example, the current seven-dimensional state vector is input into the actor network, and after several fully connected layers, the network outputs a four-dimensional vector corresponding to the reference torque command adjustment, harmonic current injection amplitude, harmonic current injection phase, and switching frequency modulation amplitude. These values ​​are directly used for subsequent thermal shaping control vector generation.

[0056] Long-term health rewards are determined based on the change in the dynamic health index after an action is performed. Positive rewards are given when the rate of increase or decrease in the dynamic health index slows down, and negative rewards are given when the rate of decrease accelerates. The dynamic health index decreases over time, but the rate of decrease can be slowed down through appropriate control strategies. Specifically, the dynamic health index values ​​before and after the action are collected. The change in health index is calculated as the difference between the post-action health index and the pre-action health index. This change is usually negative, indicating a decrease in health. The rate of change of the health index at the current moment is compared with the rate of change of the health index at the previous moment. If the absolute value of the rate of change decreases, meaning the rate of decrease in health is slower, the long-term health reward is positive, for example, by adding 0.1 multiplied by the decrease in the rate of change. If the absolute value of the rate of change increases, meaning the rate of decrease accelerates, the long-term health reward is negative, for example, by subtracting 0.2 multiplied by the increase in the rate of change. If the health index experiences a temporary increase, such as due to improved heat dissipation causing a rebound in the calculated health index, a larger positive reward is given, for example, by adding 0.5.

[0057] The short-term performance loss penalty is determined based on the absolute value of the deviation between the actual output torque and the required torque after the action is executed. The actual output torque is the torque value actually generated by the motor under the action command, and the required torque is the expected torque value required by the upper-level controller. The absolute value of the deviation between the two reflects the loss of control accuracy. The penalty value is proportional to the absolute value of the deviation. For example, when the absolute value of the deviation is 0.5 N·m, the penalty value is -0.1; when the absolute value of the deviation is 1.5 N·m, the penalty value is -0.3; when the deviation is zero, the penalty value is zero; and when the deviation exceeds a preset tolerance threshold, such as 3 N·m, the penalty value is a larger negative value, such as -1.0. This penalty mechanism constrains the agent to avoid excessively sacrificing instantaneous performance while pursuing long-term health.

[0058] The process of generating a thermally shaped control vector based on optimization decisions involves superimposing the adjustment amount of the reference torque command in the motion space component with the current basic torque command to obtain an updated reference torque command. The adjustment amount of the reference torque command in the motion space component is a value output by the multi-objective reinforcement learning controller, such as -2.5 Nm, indicating that the torque output needs to be appropriately reduced under the current operating conditions to mitigate thermal stress accumulation. The current basic torque command is the original torque command issued by the upper-level controller based on load requirements, such as 50 Nm. The superposition of the two yields an updated reference torque command of 47.5 Nm, which is sent to the motor driver as a new torque control target. This reduces heat generation by moderately reducing output while meeting basic load requirements.

[0059] The amplitude and phase of the harmonic current injection in the action space component are combined with the current fundamental current command to obtain a complete harmonic current injection strategy. The harmonic current injection amplitude is a value output by the multi-objective reinforcement learning controller, such as 1.2 amps, indicating the magnitude of a specific harmonic current to be injected. The harmonic current injection phase is an angle value output by the multi-objective reinforcement learning controller, such as 45 degrees, indicating the phase relationship between the injected harmonic and the fundamental current. The current fundamental current command is a basic current component calculated by a vector control algorithm based on the updated reference torque command, such as a fundamental sinusoidal current with an amplitude of 30 amps and a frequency of 50 Hz. The fundamental current command is synthesized with the harmonic current injection amplitude and phase by superimposing a harmonic component with a specified amplitude and phase onto the fundamental current waveform. For example, injecting a 5th harmonic current with an amplitude of 1.2 amps and a phase of 45 degrees results in a synthesized current command containing both the fundamental and 5th harmonic components. This harmonic current will generate additional harmonic losses inside the motor. By reasonably selecting the injected harmonic order, amplitude, and phase, the loss distribution of parts such as rotor bars and stator windings can be actively changed, achieving the effect of dissipating heat from hot spots.

[0060] The switching frequency modulation amplitude in the action space component is superimposed with the current reference switching frequency to obtain the updated switching frequency modulation strategy. The switching frequency modulation amplitude is a value output by the multi-objective reinforcement learning controller, such as +800 Hz, indicating the value by which the inverter's switching frequency needs to be increased or decreased. The current reference switching frequency is the inverter's default switching frequency setting, such as 8000 Hz. The superposition of the two yields an updated switching frequency modulation strategy of 8800 Hz. This instruction is sent to the inverter driver to adjust the switching frequency of the power devices. Changes in the switching frequency affect the inverter's own heat generation and the harmonic content of the output current. For example, increasing the switching frequency can reduce current harmonics but increase inverter switching losses, while decreasing the switching frequency can reduce inverter heat generation but increase motor harmonic losses. By dynamically modulating the switching frequency, a trade-off and balance can be achieved between inverter heat generation and motor heat generation.

[0061] The thermal shaping control vector is composed of an updated reference torque command, a complete harmonic current injection strategy, and an updated switching frequency modulation strategy. The thermal shaping control vector is a comprehensive set of control commands that actively regulates the motor's thermal state from three dimensions: torque output, current waveform, and switching frequency. For example, at a given moment, the thermal shaping control vector includes an updated reference torque command of 47.5 Nm, a complete harmonic current injection strategy of 1.2 amperes with a 45-degree phase fifth harmonic, and an updated switching frequency modulation strategy of 8800 Hz. These three commands are sent in parallel to the motor control system, working together to actively shape the internal heat distribution and heat flow path of the motor, maintaining the thermal stress field within a preset range.

[0062] The above algorithms or formulas are all dimensionless and numerical calculations, and the results are obtained by software simulation based on a large amount of collected data to obtain the most recent real-world results. The preset parameters are set by those skilled in the art according to the actual situation.

[0063] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0064] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0065] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0066] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting and protecting the overheating risk of an asynchronous motor, characterized in that, Includes the following steps: Real-time operating data of the asynchronous motor is collected and input into the intelligent fusion model to estimate the temperature distribution and thermal stress field distribution of each preset key component inside the motor in real time. The intelligent fusion model includes a simplified physical model and a graph neural network. The graph neural network is used to learn and characterize the heat conduction graph structure between preset key components inside the motor. The thermal stress change rate and thermal stress field non-uniformity are calculated based on the thermal stress field distribution, and the hot spot temperature change trend in the near future is predicted based on the temperature distribution. Obtain historical thermal stress data of the motor and calculate the dynamic health index of the motor by combining it with a preset aging model; The thermal stress change rate, thermal stress field non-uniformity, hot spot temperature change trend and dynamic health index are used as inputs. The optimization decision is made by a multi-objective reinforcement learning controller. The optimization objective of the multi-objective reinforcement learning controller is to maximize the long-term health of the motor and minimize the short-term performance loss. The thermal shaping control vector is generated based on the optimization decision, and the motor is regulated based on the thermal shaping control vector.

2. The method for predicting and protecting the overheating risk of an asynchronous motor according to claim 1, characterized in that, The simplified physical model is constructed based on the thermal network method, and its specific application includes the following steps: The interior of the motor is divided into multiple temperature nodes according to its geometric structure and material properties. These temperature nodes include stator winding nodes, stator core nodes, rotor bar nodes, rotor core nodes, housing nodes, and end airflow nodes. Based on the heat conduction and convection cooling mechanisms, thermal conductivity parameters between each temperature node are established. These parameters include conductive thermal conductivity and convective thermal conductivity. Then, a heat source matrix is ​​established based on the loss distribution corresponding to each temperature node. The loss distribution includes stator copper loss, rotor aluminum loss, core loss, and mechanical loss. Based on the thermal conductivity parameters and the heat source matrix, the heat balance equations for each temperature node are established and solved to obtain the temperature values ​​of each temperature node; based on the temperature difference between adjacent temperature nodes and the thermal conductivity parameters, the instantaneous heat flux density on each heat transfer path is calculated; based on the rate of change of temperature values ​​over time, the temperature rise rate of each temperature node is calculated.

3. The method for predicting and protecting the overheating risk of an asynchronous motor according to claim 2, characterized in that, The working principle of graph neural networks is as follows: The temperature values, temperature rise rates, and heat flux densities along each heat transfer path obtained from the simplified physical model calculations are used as the initial node features and edge features of the graph neural network. The graph neural network uses each preset key component as a graph node and the heat conduction path between each preset key component as a graph edge to construct the heat conduction graph structure of the motor. The current, voltage, and speed signals from real-time operating data are used as input features of the graph neural network and fused with node features and edge features. Graph neural networks are trained through supervised learning, using measured or simulated temperature data as the true temperature values ​​and labeling them with training data. They learn to simplify the residual mapping relationship between the temperature values ​​output by the physical model and the true temperature values. The correction value output by the graph neural network is added to the output of the simplified physical model to obtain the temperature distribution after preliminary correction.

4. The method for predicting and protecting the overheating risk of an asynchronous motor according to claim 3, characterized in that, Based on the supervised learning training of the graph neural network and the acquisition of the initially corrected temperature distribution, the physical laws of the initially corrected temperature distribution are verified through the following methods: By leveraging the thermal coupling relationship between key components over long distances established by multiple stacked graph convolutional layers in a graph neural network, feature propagation results based on heat conduction paths between graph nodes are extracted. Physical constraints are introduced during the training of graph neural networks. These physical constraints include residual constraints of the heat conduction equation, energy conservation constraints, and constraints of the second law of thermodynamics. The temperature distribution after preliminary correction is guided by physical constraints, so that the final temperature distribution output by the graph neural network satisfies both the data-driven correction result and the basic laws of thermodynamics. Based on the final output temperature distribution, the thermal stress field distribution is calculated by combining the thermal expansion coefficient and elastic modulus of the materials of each preset key component. The thermal stress field distribution includes normal stress and shear stress components in each direction.

5. The method for predicting and protecting the overheating risk of an asynchronous motor according to claim 4, characterized in that, The calculation of the rate of change of thermal stress and the non-uniformity of the thermal stress field based on the thermal stress field distribution is achieved in the following way: The thermal stress field distribution of each preset key component is represented as a tensor field containing multiple stress components. The trajectory curvature of the tensor field in phase space is calculated as the thermal stress change rate. The trajectory curvature is determined by dividing the magnitude of the vector product of the second and first derivatives of the thermal stress tensor with respect to time by the cube of the magnitude of the first derivative. The thermal stress field distribution is mapped to a multi-scale fractal space, the fractal dimension of the thermal stress contour at each scale is calculated, and the fluctuation amplitude of the fractal dimension with scale is used as the first component of the thermal stress field inhomogeneity. Based on the heat conduction diagram structure between each preset key component, the ratio of the number of boundaries where the thermal stress gradient between adjacent diagram nodes exceeds a preset threshold to the total number of boundaries is used as the second component of thermal stress field non-uniformity. The thermal stress field non-uniformity is obtained by geometrically weighting and fusing the first component and the second component. Based on the temperature distribution, the future temperature change trajectory of each preset key component is reconstructed in phase space, and the Lyapunov exponent of the reconstructed trajectory is calculated. The number of dimensions with a Lyapunov exponent greater than zero is used as a representation of the future temperature change trend of hotspots.

6. The method for predicting and protecting the overheating risk of an asynchronous motor according to claim 5, characterized in that, The dynamic health index of the motor is calculated in the following way: The thermal stress change rate and thermal stress field inhomogeneity in historical thermal stress data are mapped to a pre-defined aging model in the cumulative damage space. The aging model is constructed as follows: Based on the nonlinear cumulative damage model, the rate of change of thermal stress is used as the damage driving force, and the non-uniformity of thermal stress field is used as the damage coupling coefficient. The damage coupling coefficient is determined according to the spatial distribution of thermal stress gradient between each preset key component. When the non-uniformity of thermal stress field increases, the damage coupling coefficient increases monotonically, and the rate of increase increases with the increase of non-uniformity. Historical thermal stress data is decomposed into a continuous sequence of thermal shock events. Each thermal shock event corresponds to a period in which the rate of change of thermal stress exceeds a preset event threshold. The basic damage amount of a single thermal shock event is calculated based on the peak value of the rate of change of thermal stress during that period, and the corresponding damage coupling coefficient is determined based on the average thermal stress field non-uniformity during that period. The basic damage values ​​of all thermal shock events prior to the current moment are adjusted by the corresponding damage coupling coefficient and then nonlinearly accumulated to obtain the cumulative damage value. The nonlinear accumulation adopts a damage history-dependent accumulation method. The dynamic health index of the motor is calculated based on the ratio between the cumulative damage value and the preset failure threshold. The dynamic health index is expressed as a percentage and decreases as the cumulative damage value increases. When the cumulative damage value reaches the failure threshold, the health index returns to zero.

7. The method for predicting and protecting the overheating risk of an asynchronous motor according to claim 1, characterized in that, Optimization decisions are achieved through multi-objective reinforcement learning regulators in the following ways: Construct the state space of a multi-objective reinforcement learning controller. The state space includes the current thermal stress change rate, thermal stress field non-uniformity, hot spot temperature change trend, dynamic health index, and the current operating parameters of the motor, including speed, torque, and current. The action space of a multi-objective reinforcement learning controller is constructed. The action space includes the adjustment amount of the reference torque command, the amplitude and phase of the harmonic current injection, and the modulation amplitude of the switching frequency. Each action component is continuously adjustable within a preset safety range. Construct a reward function for a multi-objective reinforcement learning regulator, which includes a long-term health reward term and a short-term performance loss penalty term. A deep deterministic policy gradient algorithm is used to train a multi-objective reinforcement learning regulator, enabling the regulator to learn a nonlinear mapping relationship from the state space to the action space. During the training process, the goal is to maximize the cumulative discount reward, which is the weighted sum of the reward values ​​at each future time step, with the weights decaying over time. During real-time control, the current state space is input into the trained multi-objective reinforcement learning controller, and the controller outputs the corresponding action space components.

8. The method for predicting and protecting the overheating risk of an asynchronous motor according to claim 7, characterized in that, The long-term health reward is determined based on the change in the dynamic health index after the action is executed. A positive reward is given when the rate of increase or decrease of the dynamic health index slows down, and a negative reward is given when the rate of decrease of the dynamic health index accelerates. The short-term performance loss penalty is determined based on the absolute value of the deviation between the actual output torque and the required torque after the action is executed.

9. The method for predicting and protecting the overheating risk of an asynchronous motor according to claim 7, characterized in that, The generation of thermal shaping control vectors based on optimization decisions refers to: The reference torque command adjustment in the motion space component is superimposed with the current fundamental torque command to obtain the updated reference torque command; the harmonic current injection amplitude and phase in the motion space component are synthesized with the current fundamental current command to obtain the complete harmonic current injection strategy. The switching frequency modulation amplitude in the action space component is superimposed with the current reference switching frequency to obtain the updated switching frequency modulation strategy. The thermal shaping control vector is composed of the updated reference torque command, the complete harmonic current injection strategy, and the updated switching frequency modulation strategy.

Images