Temperature field prediction model training method and motor temperature prediction method based on same

Abstract

The application relates to the technical field of intelligent transportation, and discloses a temperature field prediction model training method and a motor temperature prediction method based on the same. The method comprises the following steps: obtaining training data of a motor, wherein the training data comprises training temperature data corresponding to spatial coordinate points of the motor; constructing a high-dimensional physical field mapping neural network; taking the training data as a training sample, combining a first data loss and a physical constraint loss to construct a composite loss function, iteratively updating the weight of the high-dimensional physical field mapping neural network through back propagation, and obtaining a physical benchmark model; wherein the first data loss represents the error between the predicted temperature output by the high-dimensional physical field mapping neural network and the training temperature data; and based on the measured temperature data of the motor, the physical benchmark model is calibrated to obtain a temperature field prediction model, so that the calibrated temperature field is output through the temperature field prediction model. The technical scheme provided by the application can improve the accuracy and generalization ability of the predicted temperature field.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method for training a temperature field prediction model and a method for predicting motor temperature based thereon. Background Technology

[0002] With the rapid development of electric vehicles, robots and industrial automation technologies, motors are constantly evolving towards higher power density, higher torque density and miniaturization, which leads to a significant increase in internal heat load. Excessive temperature can cause irreversible demagnetization of permanent magnets, insulation aging and even breakdown and short circuit, seriously threatening the safety and lifespan of motors. Therefore, achieving high-precision, real-time and global temperature field monitoring and prediction has become the core requirement of motor thermal management and digital twin technology.

[0003] Among relevant temperature prediction technologies, the lumped parameter thermal path method is fast but has limited accuracy and struggles to capture hot spots; the finite element method is accurate but computationally intensive and only suitable for offline design; data-driven proxy models attempt to strike a balance between speed and accuracy. However, these methods are generally limited by three major drawbacks: complex and variable boundary conditions leading to ill-posed solutions, dynamic non-stationarity of source loads causing distortion of mapping relationships, and nonlinear residuals caused by assembly processes and aging. Ultimately, this results in existing digital twin models being unable to simultaneously achieve real-time performance and refined capture, and exhibiting a lack of physical consistency and generalization ability, making it difficult to support high-precision online monitoring and thermal protection of real vehicles.

[0004] Therefore, improving the accuracy and generalization ability of predicted temperature fields during motor temperature prediction has become a technical problem that needs to be solved. Summary of the Invention

[0005] This application provides a method for training a temperature field prediction model and a method for predicting motor temperature based thereon, which solves the technical problem of how to improve the accuracy and generalization ability of the predicted temperature field in the process of predicting motor temperature.

[0006] To achieve the above objectives, the main technical solutions adopted in this application include: In a first aspect, embodiments of this application provide a method for training a temperature field prediction model, the method comprising: Acquire training data for the motor, the training data including training temperature data corresponding to the spatial coordinate points of the motor; A high-dimensional physical field mapping neural network is constructed. The training data is used as training samples. A composite loss function is constructed by combining the first data loss and the physical constraint loss. The weights of the high-dimensional physical field mapping neural network are iteratively updated through backpropagation to obtain a physical benchmark model. The first data loss represents the error between the predicted temperature output by the high-dimensional physical field mapping neural network and the training temperature data. Based on the measured temperature data of the motor, the physical reference model is calibrated to obtain a temperature field prediction model, and the calibrated temperature field is output through the temperature field prediction model.

[0007] This embodiment provides a temperature field prediction model training method. It replaces the traditional discrete convolutional architecture by constructing a high-dimensional physical field mapping neural network (such as a neural operator incorporating geometric coordinate encoding). This achieves a direct mapping from training data to the continuous temperature field function space, making the temperature field prediction model resolution-independent. While maintaining extremely fast inference speed, it accurately captures temperature details at any location, thus resolving the contradiction between computational efficiency and local accuracy. Secondly, physical constraint losses, such as partial differential equations of heat conduction, are introduced during training. A composite loss function integrating the first data loss and the physical constraint loss is constructed, embedding physical laws into the temperature field prediction model. This ensures that the prediction results conform to the principles of energy conservation and thermodynamics, significantly improving the generalization ability and reliability of the temperature field prediction model under different operating conditions and structural parameters. Finally, through a calibration architecture based on measured motor temperature data, a physical benchmark model is obtained using a large amount of high-fidelity training data as a physical benchmark. A nonlinear residual mapping between the training temperature data and reality is learned by combining a small amount of measured temperature data, effectively eliminating deviations caused by manufacturing tolerances and environmental uncertainties. Ultimately, a high-precision, high-reliability temperature field prediction model is obtained, capable of quickly outputting the calibrated temperature field distribution.

[0008] In one implementation, the physical baseline model is trained as follows: Construct a high-dimensional physical field mapping neural network that sequentially includes an input encoding module, a feature iteration module, and a coordinate decoding module; The geometric feature vector, operating condition feature vector, spatial coordinate points, and training temperature data corresponding to the spatial coordinate points are extracted from the training data as training samples. The geometric feature vector and the operating condition feature vector are concatenated into a system state vector through the input encoding module. The geometric feature vector, the operating condition feature vector, and the training temperature data are mapped to the same spatial coordinate point. The feature iteration module performs low-dimensional feature to high-dimensional latent feature enhancement, global and local feature depth extraction and feature dimensionality reduction on the system state vector, and then the coordinate decoding module outputs the predicted temperature field. Using the differential computation unit of the deep learning framework, the first and second derivatives of the predicted temperature field are solved, substituted into the heat conduction control equation, and combined with material property parameters and heat source density to construct a physical residual tensor to obtain the physical constraint loss. A composite loss function is constructed based on the first data loss and the physical constraint loss. The weights of each module in the high-dimensional physical field mapping neural network are then adjusted in reverse iteratively until convergence, resulting in a trained physical benchmark model.

[0009] This embodiment constructs a high-dimensional physics-field mapping neural network. Geometric feature vectors, operating condition feature vectors, spatial coordinates, and corresponding training temperature data are extracted from the training data as training samples. The geometric feature vectors and operating condition feature vectors are then concatenated into a system state vector via an input encoding module. In the high-dimensional physics-field mapping neural network architecture, the feature iteration module sequentially performs dimensionality enhancement of the system state vector, deep extraction of global and local features (global units capture long-distance thermal coupling, and local units simulate gradient propagation), and feature dimensionality reduction and compression, generating implicit feature vectors that do not contain spatial coordinates but contain complete thermal field information. Finally, the coordinate decoding module can output the predicted temperature value by inputting any coordinates. To enhance the physical consistency of the physical benchmark model, the time first derivative and spatial second derivative of the predicted temperature are calculated using automatic differentiation and substituted into the heat conduction control equation (including transient terms, heat conduction terms, and heat source terms) to construct the physical residual. This scheme introduces the heat conduction control equation as a physical constraint loss, enabling the high-dimensional physics-mapping neural network to adhere to thermodynamic laws such as energy conservation during the learning process. It also addresses the problem of prediction errors that easily violate physical principles under different operating conditions by learning the nonlinear deviation between the predicted temperature and the training temperature data. Simultaneously, it effectively compensates for the systematic discrepancies between the prediction and the true value, ultimately forming a physical benchmark model that combines physical interpretability with high accuracy for engineering application. This scheme provides reliable technical support for the precise design and real-time management of motor thermal characteristics.

[0010] In one embodiment, the geometric feature vector is obtained by vectorizing the structural parameters of the motor, and the operating condition feature vector is obtained by nonlinearly normalizing the operating data of the motor.

[0011] This embodiment extracts structural parameters from different regions of the motor, such as the stator, rotor, and windings, by extracting features from the motor's geometry and exporting them from the physical simulation software interface. These parameters collectively constitute the geometric features describing the motor's physical structure, comprehensively characterizing the geometric attributes of each component. The embodiment also acquires the motor's operating data, including dynamic operating information such as speed, torque, and current. Since the structural parameters and operating data differ significantly in dimensions and numerical ranges, a nonlinear normalization method is used to map them to a unified dimensionless interval (e.g., [-1,1]), ultimately generating standardized geometric feature vectors and operating condition feature vectors. This lays the foundation for the subsequent training and feature fusion of the high-dimensional physical field mapping neural network model.

[0012] In one embodiment, the temperature field prediction model is obtained through a residual network calibration method, which includes: The weight parameters of the physical benchmark model are fixed, and the physical benchmark model with fixed weight parameters is used as the benchmark output module; A lightweight neural network is constructed, wherein the lightweight neural network uses the same working condition feature vector and geometric feature vector as input as the benchmark output module; The difference between the measured temperature data and the reference temperature output by the reference output module is used as the residual label. The lightweight neural network is trained until the difference between the predicted residual output by the lightweight neural network and the residual label converges, thus obtaining the trained lightweight neural network. The trained lightweight neural network is then used as the residual correction module. The temperature field prediction model is obtained by combining the reference output module and the residual correction module.

[0013] This embodiment achieves an efficient fusion of physical constraints and data-driven approaches by constructing a baseline output module and a residual correction module, significantly improving the accuracy and robustness of motor temperature prediction. The weight parameters of the physical baseline model are completely fixed before training, used to output a baseline temperature that conforms to physical laws, ensuring that the temperature field prediction model maintains reasonable physical boundaries even under complex operating conditions. The residual correction module employs a lightweight neural network, using motor operating condition feature vectors (such as current, speed, and ambient temperature) and geometric feature vectors as inputs to obtain the prediction residuals, until the difference between the predicted residuals output by the lightweight neural network and the residual labels converges, thus serving as the residual correction module. Finally, the baseline output module and the residual correction module are combined to obtain the final temperature field prediction model. This divide-and-conquer structure retains the interpretability and stability of the physical baseline model while leveraging the powerful fitting ability of the lightweight neural network to compensate for the error between the baseline temperature and the measured temperature data, effectively solving the problems of insufficient accuracy and weak generalization ability of purely data-driven models.

[0014] In one embodiment, outputting the calibrated temperature field through the temperature field prediction model includes: The temperature value after calibration at the same spatial coordinate point is obtained by adding the reference temperature output by the reference output module at the same spatial coordinate point in the temperature field prediction model with the prediction residual output by the residual correction module. Iterate through all spatial coordinate points to obtain the calibrated temperature values ​​for all spatial coordinate points, and output the calibrated temperature field.

[0015] This embodiment obtains the calibrated temperature value of a point by superimposing the reference temperature at the same spatial coordinate point with the predicted residual. By traversing all spatial coordinate points within the spatial region and superimposing the reference temperature with the corresponding predicted residual, a complete and continuous high-precision calibrated temperature field is synthesized. This method effectively combines the stability of physical constraints with the flexibility of the residual correction module, significantly improving the prediction accuracy and local detail capture capability of the calibrated temperature field. It can also compensate for deviations caused by operating condition disturbances or modeling errors in real time. The final output calibrated temperature field has stronger realism and reliability, providing solid data support for the precise monitoring, analysis, and optimization of the motor's operation.

[0016] In one embodiment, the temperature field prediction model is obtained through transfer learning fine-tuning, the transfer learning fine-tuning method including: Configure the weights of the physical benchmark model as initial weights, fix the weights of the input encoding module of the physical benchmark model, and configure the weights of the decoding layer of the physical benchmark model as trainable. Using the measured temperature data as supervision, a total loss function is constructed that integrates the second data loss and the physical constraint loss, wherein the second data loss characterizes the error between the reference temperature output by the physical reference model and the measured temperature data; The weights of the decoding layer are iteratively updated through backpropagation until the total loss function converges, thus obtaining the corrected model weights. The corrected model weights are used to replace the initial weights to obtain the temperature field prediction model.

[0017] This embodiment constructs a total loss function that integrates the second data loss and the physical constraint loss, uses the backpropagation algorithm to iteratively optimize the weights of the decoding layer, and loads the corrected model weights back into the physical baseline model to finally obtain the calibrated temperature field prediction model. This method, relying only on a small amount of measured data, can achieve accurate approximation of predicted values ​​at observation points while ensuring that the global prediction results strictly follow physical laws such as heat conduction. It effectively improves the model's prediction accuracy, physical consistency, robustness, and generalization ability in sparse observation scenarios, providing high-fidelity data support for engineering thermal analysis.

[0018] Secondly, embodiments of this application provide a method for predicting motor temperature, the prediction method comprising: Obtain the structural parameters, operating data, and spatial coordinates of the motor under test; Based on the structural parameters and operating data of the motor under test, the target geometric feature vector and the target operating condition feature vector are obtained; The target geometric feature vector, the target operating condition feature vector, and the spatial coordinate point to be queried are input into the temperature field prediction model obtained by any of the above training methods to obtain the temperature value corresponding to the spatial coordinate point to be queried in the motor under test.

[0019] Thirdly, embodiments of this application provide a temperature field prediction model training device, the training device comprising: A data acquisition unit is used to acquire training data of the motor, the training data including training temperature data corresponding to the spatial coordinate points of the motor; The model training unit is used to construct a high-dimensional physical field mapping neural network. Using the training data as training samples, it combines a composite loss function constructed by the first data loss and the physical constraint loss, and iteratively updates the weights of the high-dimensional physical field mapping neural network through backpropagation to obtain a physical benchmark model. The first data loss represents the error between the predicted temperature output by the high-dimensional physical field mapping neural network and the training temperature data. The model calibration unit is used to calibrate the physical reference model based on the measured temperature data of the motor to obtain a temperature field prediction model, and output the calibrated temperature field through the temperature field prediction model.

[0020] Fourthly, embodiments of this application provide a computer device, including: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform either the temperature field prediction model training method or the motor temperature prediction method described above.

[0021] Fifthly, embodiments of this application provide a computer-readable storage medium storing computer instructions, which are used to cause a computer to execute the temperature field prediction model training method or the motor temperature prediction method described above. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 A flowchart illustrating a temperature field prediction model training method provided in this application embodiment; Figure 2A flowchart of step S31 provided in an embodiment of this application; Figure 3 A diagram illustrating the architecture of a high-dimensional physical field mapping neural network provided in an embodiment of this application; Figure 4 A block diagram illustrating the principle of obtaining the composite loss function provided in this application embodiment; Figure 5 A flowchart of step S511 provided in an embodiment of this application; Figure 6 A flowchart of step S5171 provided in the embodiments of this application; Figure 7 A flowchart of step S531 provided in an embodiment of this application; Figure 8 A temperature field prediction model training device is provided for embodiments of this application; Figure 9 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] With the rapid development of electric vehicles, robotics, and industrial automation technologies, motors are continuously evolving towards higher power density, higher torque density, and miniaturization. This trend has led to a sharp increase in the internal heat load density of motors. Excessive temperature rise can directly induce irreversible demagnetization of permanent magnets, aging of stator winding insulation, and even breakdown and short circuits, seriously threatening the operational safety and service life of motors. Therefore, building a motor temperature field monitoring and prediction system with high precision, real-time performance, and full-coverage capabilities has become a core requirement for the development of motor thermal management and digital twin technologies, ensuring their safe and reliable operation under high-load conditions.

[0026] In the construction of digital twins for motor thermal field monitoring, related technologies suffer from three major, insurmountable defects: First, their core architecture often employs convolutional neural networks that rely on regular spatial discretization. This necessitates a trade-off between computational efficiency and local accuracy. When increasing mesh resolution to capture tiny hot spots such as winding ends, the computational load increases cubically, and the inference process suffers from interpolation errors due to the inability to obtain continuous temperature information between meshes. Second, purely data-driven training methods aim only at fitting the data distribution, lacking guidance from prior physical knowledge such as partial differential equations of heat conduction and energy conservation. This leads to models prone to making predictions that violate physical laws under uncovered operating conditions, and exhibits extremely weak generalization ability for different motor geometries and operating data. Finally, due to the difficulty in obtaining full-domain measured data, models often rely on finite element simulations under ideal boundary conditions for training. However, the simple convolutional neural network architecture cannot effectively learn and compensate for the nonlinear residuals between the simulation and the actual entity caused by factors such as manufacturing tolerances, assembly processes, and nonlinear aging. In summary, these three major shortcomings make it difficult for existing digital twin models to meet both real-time and fine-grained capture requirements. Furthermore, they suffer from significant deficiencies in physical consistency and generalization ability in engineering applications, making them unable to directly support high-precision online monitoring and thermal protection of real vehicles.

[0027] In summary, improving the accuracy and generalization ability of predicted temperature fields during motor temperature prediction is a technical problem that needs to be solved.

[0028] To address the aforementioned technical problems, this application provides a method for training a temperature field prediction model. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system containing a set of computer-executable instructions. Furthermore, although the logical order of each step is shown in the flowchart, in some cases, these steps can be executed in a different order than those shown or described herein.

[0029] This embodiment provides a method for training a temperature field prediction model. Figure 1 A flowchart of a temperature field prediction model training method provided in this application embodiment is shown below. Figure 1 As shown, the process includes the following steps: Step S1: Obtain the training data of the motor, which includes the training temperature data corresponding to the spatial coordinate points of the motor.

[0030] Specifically, training data under various operating conditions is generated in batches through simulation. This training data is defined on a continuous space or a fine grid (e.g., containing 100,000 grid points), which can accurately characterize the temperature distribution of key areas inside the motor, including the stator winding ends, rotor bars, and heat dissipation fins in the casing. At the same time, the training data also includes the training temperature data corresponding to the spatial coordinate points.

[0031] Step S3: Construct a high-dimensional physical field mapping neural network. Using training data as training samples, combine the first data loss and the physical constraint loss to construct a composite loss function. Iteratively update the weights of the high-dimensional physical field mapping neural network through backpropagation to obtain the physical benchmark model. The first data loss represents the error between the predicted temperature output by the high-dimensional physical field mapping neural network and the training temperature data.

[0032] Specifically, a high-dimensional physical field mapping neural network is constructed, using training data as training samples, and a composite loss function is built for optimization. The composite loss function consists of two parts: first, a data loss, which addresses the difference between the predicted temperature and the training temperature data, ensuring consistency between the two; second, a physical constraint loss, such as residuals from partial differential equations based on the heat conduction equation, boundary condition constraints, or the law of conservation of energy, forcing the high-dimensional physical field mapping neural network to satisfy physical laws. During training, the weights of the high-dimensional physical field mapping neural network are iteratively updated using a backpropagation algorithm, gradually reducing the composite loss function, ultimately obtaining a physical benchmark model that combines data fitting accuracy with physical consistency. The physical benchmark model can efficiently predict the temperature field under different operating conditions and provide a reliable benchmark for subsequent analysis.

[0033] Step S5: Based on the measured temperature data of the motor, the physical reference model is calibrated to obtain the temperature field prediction model, and the calibrated temperature field is output through the temperature field prediction model.

[0034] Specifically, based on measured temperature data of the motor, a physical baseline model is calibrated to construct a temperature field prediction model. The calibration process can be achieved using methods such as transfer learning or parameter optimization. The physical baseline model can be fine-tuned using measured data, or residual correction terms can be added to compensate for deviations between training and actual measurements. The calibrated temperature field prediction model can integrate physical laws and real-world operating data, more accurately reflecting the thermal behavior of the motor under actual operating conditions. By inputting new parameters into the temperature field prediction model, the calibrated temperature field distribution can be quickly output, providing a reliable basis for motor thermal management design and condition monitoring.

[0035] This embodiment provides a temperature field prediction model training method. It replaces the traditional discrete convolutional architecture by constructing a high-dimensional physical field mapping neural network (such as a neural operator incorporating geometric coordinate encoding). This achieves a direct mapping from training data to the continuous temperature field function space, making the temperature field prediction model resolution-independent. While maintaining extremely fast inference speed, it accurately captures temperature details at any location, thus resolving the contradiction between computational efficiency and local accuracy. Secondly, physical constraint losses, such as partial differential equations of heat conduction, are introduced during training. A composite loss function integrating the first data loss and the physical constraint loss is constructed, embedding physical laws into the temperature field prediction model. This ensures that the prediction results conform to the principles of energy conservation and thermodynamics, significantly improving the generalization ability and reliability of the temperature field prediction model under different operating conditions and structural parameters. Finally, through a calibration architecture based on measured motor temperature data, a physical benchmark model is obtained using a large amount of high-fidelity training data as a physical benchmark. A nonlinear residual mapping between the training temperature data and reality is learned by combining a small amount of measured temperature data, effectively eliminating deviations caused by manufacturing tolerances and environmental uncertainties. Ultimately, a high-precision, high-reliability temperature field prediction model is obtained, capable of quickly outputting the calibrated temperature field distribution.

[0036] Figure 2 The flowchart for step S3 provided in the embodiments of this application may include the following steps: Step S31: Construct a high-dimensional physical field mapping neural network that sequentially includes an input encoding module, a feature iteration module, and a coordinate decoding module.

[0037] Specifically, please refer to Figure 3 The input encoding module is responsible for acquiring motor operating condition feature vectors (such as speed, torque, coolant flow rate, etc.) and geometric feature vectors, and concatenating them to obtain the system state vector. The feature iteration module processes the system state vector to map low-dimensional features to high-dimensional latent features, and then deeply extracts global and local features. Subsequently, the iterated high-dimensional features are gradually reduced in dimensionality to obtain implicit feature vectors. Finally, the coordinate decoding module decodes the implicit feature vectors and outputs the predicted temperature value at the corresponding sampling point. This high-dimensional physical field mapping neural network realizes an end-to-end mapping from input to continuous temperature field output, which can quickly reproduce the temperature distribution pattern in the training data and provide technical support for real-time temperature prediction of motors.

[0038] Step S33: Extract the geometric feature vector, operating condition feature vector, spatial coordinate points, and corresponding training temperature data from the training data as training samples. Then, use the input encoding module to concatenate the geometric feature vector and operating condition feature vector into a system state vector. The geometric feature vector, operating condition feature vector, and training temperature data are mapped to the same spatial coordinate points.

[0039] Specifically, in the training process of the high-dimensional physical field mapping neural network, the following steps are first taken from the training data: geometric feature vectors describing the physical structure of the motor, operating condition feature vectors reflecting the real-time operating state of the motor, spatial coordinates of the temperature to be queried, and the corresponding training temperature data of the spatial coordinates, which are then used as training samples. The geometric feature vectors and operating condition feature vectors are concatenated along the channel dimension to generate a system state vector, which serves as the input to the feature iteration module. Simultaneously, a data indexing mechanism is established to create a precise mapping relationship between the system state vector, spatial coordinates, and the corresponding training temperature data, ensuring that each set of input combinations has a unique set of corresponding training temperature data for both the geometric feature vector and the operating condition feature vector. The final output of the input encoding module is a standardized, integrated low-dimensional system state vector and a coordinate-encoded spatial coordinate point feature vector. The low-dimensional input system state vector is passed to the feature iteration module for high-dimensional feature extraction, while the spatial coordinates are input to the coordinate decoding module.

[0040] Step S35: The feature iteration module improves the system state vector from low-dimensional features to high-dimensional potential features, performs global and local feature depth extraction and feature dimensionality reduction, and then outputs the predicted temperature field through the coordinate decoding module.

[0041] Specifically, a learnable fully connected lifting layer maps the low-dimensional features (e.g., 32-dimensional) of the system state vector to high-dimensional latent features (e.g., 128-dimensional or 256-dimensional), achieving dimensional expansion and spatial projection, providing rich representation capabilities for subsequent complex feature extraction. Subsequently, the feature iteration module performs global and local feature depth extraction on the high-dimensional latent features. Its core consists of a global feature extraction unit and a local feature extraction unit connected in parallel: the global unit captures long-distance thermally coupled global features through an integral transform mechanism, such as the thermal impact of winding heating on the distant casing; the local unit approximates local features of heat conduction through differential approximation, such as the temperature gradient between adjacent teeth. The outputs of the global and local feature extraction units are element-wise added and then fused and updated using a nonlinear activation function. The high-dimensional features, after multiple rounds of iterative optimization, then enter the projection layer, where they are progressively reduced in dimension and compressed through a multilayer perceptron with a shrinking structure, ultimately forming an implicit feature vector. This feature vector is a deep encoding of the complete temperature field shape of the motor under the current operating conditions. Although it does not correspond to specific spatial coordinates, it includes all thermal field information determined by the geometric feature vector and the operating condition feature vector. It is worth noting that the implicit feature vector differs from the discrete voxel mesh representation; its essence is an implicit neural representation defined on a continuous function space. Therefore, it can obtain the predicted temperature field, thereby achieving end-to-end high-precision mapping from the operating condition to the global temperature field. Subsequently, only spatial coordinates of any point need to be input through the coordinate decoding module. These can be high-resolution mesh coordinates or specific points. The decoder fuses the implicit feature vector with the spatial coordinates of the point to be retrieved, maps it back to real number space, and outputs a scalar predicted temperature value.

[0042] Step S37: Using the differential computation unit of the deep learning framework, solve the first and second derivatives of the predicted temperature field, substitute them into the heat conduction control equation, and construct the physical residual tensor by combining the material property parameters and heat source density to obtain the physical constraint loss.

[0043] Specifically, using a differential calculation unit, the first derivative of the predicted temperature field with respect to time is obtained by calculating backwards from the output to the input using the chain rule. and the second derivative The first derivative characterizes the dynamic rate of change of the temperature field over time, providing a core parameter for transient temperature field prediction; the second derivative characterizes the distribution gradient of the temperature field in three-dimensional space, providing a basis for the quantitative analysis of the heat conduction process. In the mapping unit between material properties and equations, the first and second derivatives of the predicted temperature field are substituted into the heat conduction control equations, material properties, and heat source density to calculate the physical constraint loss, where the transient term is: in, For material density, , where T represents the specific heat capacity of the material, and are all pre-input material property parameters. For the steady-state temperature field prediction scenario, the temperature field does not change dynamically over time; the transient term T... tran The value is zero, retaining only spatial heat conduction characteristics; for transient temperature field prediction scenarios, the transient term T is zero. tran Non-zero, fully characterizing the spatiotemporal variation of the temperature field.

[0044] The heat conduction term is: in, This represents the thermal conductivity of the material. The thermal conductivity of a material can be a scalar, used for isotropic materials, or a tensor matrix, used for anisotropic materials such as silicon steel sheets.

[0045] The heat generation power term is mapped to the heat source density Q at various locations in the motor's three-dimensional space based on the input motor speed, torque, current, and other operating data through an electromagnetic loss calculation model. This term represents the energy input, including stator copper losses, iron losses, and magnet eddy current losses.

[0046] Based on the general form of the heat conduction governing equation: Substituting the transient, heat conduction, and heat generation power terms obtained from the above calculations into the equation, a residual tensor is constructed, which represents the physical constraint loss. In addition, for oil-cooled or water-cooled motors, a convection term T may also be included. water .

[0047] The magnitude of the residual tensor characterizes the degree of deviation between the predicted temperature field and the law of conservation of heat conduction: the closer the residual tensor value is to 0, the more the predicted temperature field conforms to the thermodynamic law of energy conservation; the greater the residual tensor value deviates from 0, the more non-physical deviations exist in the predicted temperature field. In this embodiment, the physical residual tensor is directly used as the physical constraint loss, realizing the quantitative transformation of thermodynamic laws into network training loss.

[0048] Step S39: Construct a composite loss function based on the first data loss and the physical constraint loss, and iteratively adjust the weights of each module in the high-dimensional physical field mapping neural network until convergence, thereby obtaining the trained physical benchmark model.

[0049] For specific details, please refer to Figure 4 The flowchart for constructing the composite loss function is shown below. The composite loss function is constructed by combining the first data loss and the physical constraint loss. in, For physical constraint loss, The physical constraint loss weighting coefficient, For the first data loss, The weighting coefficients for the first data loss. First data loss. L data This represents the difference between the temperature prediction value and the temperature data obtained from the training temperature data in the high-dimensional physical field mapping neural network. In each iteration, the gradient of the total loss relative to the weights of each module in the high-dimensional physical field mapping neural network is calculated, and this gradient is backpropagated to the entire high-dimensional physical field mapping neural network, thereby updating the weight parameters layer by layer. This process is repeated until the loss function converges to a preset threshold, ultimately yielding a trained physical baseline model that accurately describes the mapping relationship of the high-dimensional physical field while satisfying the constraints of physical laws and the requirements of data fitting.

[0050] This embodiment constructs a high-dimensional physics-field mapping neural network. Geometric feature vectors, operating condition feature vectors, spatial coordinates, and corresponding training temperature data are extracted from the training data as training samples. The geometric feature vectors and operating condition feature vectors are then concatenated into a system state vector via an input encoding module. In the high-dimensional physics-field mapping neural network architecture, the feature iteration module sequentially performs dimensionality enhancement of the system state vector, deep extraction of global and local features (global units capture long-distance thermal coupling, and local units simulate gradient propagation), and feature dimensionality reduction and compression, generating implicit feature vectors that do not contain spatial coordinates but contain complete thermal field information. Finally, the coordinate decoding module can output the predicted temperature value by inputting any coordinates. To enhance the physical consistency of the physical benchmark model, the time first derivative and spatial second derivative of the predicted temperature are calculated using automatic differentiation and substituted into the heat conduction control equation (including transient terms, heat conduction terms, and heat source terms) to construct the physical residual. This scheme introduces the heat conduction control equation as a physical constraint loss, enabling the high-dimensional physics-mapping neural network to adhere to thermodynamic laws such as energy conservation during the learning process. It also addresses the problem of prediction errors that easily violate physical laws under different operating conditions by learning the nonlinear deviation between the predicted temperature and the training data. Simultaneously, it effectively compensates for the systematic differences between the prediction and the true value, ultimately forming a physical benchmark model that combines physical interpretability with high accuracy for engineering application. This scheme provides reliable technical support for the precise design and real-time management of motor thermal characteristics.

[0051] In one alternative embodiment, the geometric feature vector is obtained by vectorizing the structural parameters of the motor, and the operating condition feature vector is obtained by nonlinearly normalizing the operating data of the motor.

[0052] This embodiment extracts structural parameters from different regions of the motor, such as the stator, rotor, and windings, by extracting features from the motor's geometry and exporting them from the physical simulation software interface. These parameters collectively constitute the geometric features describing the motor's physical structure, comprehensively characterizing the geometric attributes of each component. The embodiment also acquires the motor's operating data, including dynamic operating information such as speed, torque, and current. Since the structural parameters and operating data differ significantly in dimensions and numerical ranges, a nonlinear normalization method is used to map them to a unified dimensionless interval (e.g., [-1,1]), ultimately generating standardized geometric feature vectors and operating condition feature vectors. This lays the foundation for the subsequent training and feature fusion of the high-dimensional physical field mapping neural network model.

[0053] Figure 5 A flowchart of a method for obtaining a temperature field prediction model provided in an embodiment of this application is shown. The process may include the following steps: Step S511: Fix the weight parameters of the physical benchmark model and use the physical benchmark model with fixed weight parameters as the benchmark output module.

[0054] Specifically, to prevent subsequent training from violating the inherent physical laws, all weight parameters of the physical baseline model are fixed and kept unchanged throughout end-to-end training. This physical baseline model serves only as a stable physical feature extractor, ensuring that the output baseline temperature always conforms to fundamental physical constraints. Subsequently, this parameter-fixed physical baseline model is used as the baseline output module to output the baseline temperature. This design ensures that the prediction results do not violate physical laws and provides a reliable physical basis for subsequent temperature field predictions.

[0055] Step S513: Construct a lightweight neural network. The lightweight neural network uses the same working condition feature vector and geometric feature vector as input as the benchmark output module.

[0056] Specifically, construct a lightweight neural network, such as a fully connected neural network (MLP), a small-scale Fourier neural operator (FNO), or other lightweight architectures such as a small-scale convolutional neural network (CNN), a graph neural network (GNN), etc.

[0057] Step S515: Using the difference between the measured temperature data and the reference temperature output by the reference output module as the residual label, train a lightweight neural network until the difference between the predicted residual output by the lightweight neural network and the residual label converges, thus obtaining a trained lightweight neural network, and use the trained lightweight neural network as a residual correction module.

[0058] Specifically, for each coordinate point Calculate the measured temperature data Reference temperature generated by the reference output module Deviation between: And The residual labels are used for training a lightweight neural network. The lightweight neural network takes the condition feature vector and geometric feature vector as input to generate coordinates. Predicted residuals at the location The optimization objective is to minimize the difference between the predicted residual and the residual label. Based on this objective, an iterative optimization algorithm is used to continuously update the parameters of the lightweight neural network until the loss function converges. At this point, the lightweight neural network is used as a residual correction module, which, together with the baseline output module, constitutes a complete temperature field prediction model, ultimately outputting a high-precision motor temperature prediction value. This method retains the interpretability and stability of the physical baseline model while using the residual correction module to compensate for the difference between the baseline temperature of the physical baseline model and the measured temperature data, significantly improving the accuracy and generalization ability of motor temperature prediction.

[0059] Step S517: Based on the combination of the benchmark output module and the residual correction module, the temperature field prediction model is obtained.

[0060] Specifically, a complete temperature field prediction model is obtained by combining the benchmark output module with the residual correction module. This method achieves complementarity between the physical benchmark model and the residual correction module. The physical benchmark model ensures that the prediction results strictly follow basic physical laws, avoiding errors that violate physical laws; the residual correction module, on the other hand, learns from the numerical differences in measured temperature data and flexibly corrects temperature distribution deviations caused by complex operating conditions, geometric details, and other factors. The combination of these two approaches not only guarantees the temperature field prediction model's adherence to physical laws but also endows it with the ability to capture subtle differences in the data, thereby significantly improving the accuracy and generalization performance of temperature field prediction while maintaining physical consistency.

[0061] This embodiment achieves an efficient fusion of physical constraints and data-driven approaches by constructing a baseline output module and a residual correction module, significantly improving the accuracy and robustness of motor temperature prediction. The weight parameters of the physical baseline model are completely fixed before training, used to output a baseline temperature that conforms to physical laws, ensuring that the temperature field prediction model maintains reasonable physical boundaries even under complex operating conditions. The residual correction module employs a lightweight neural network, using motor operating condition feature vectors (such as current, speed, and ambient temperature) and geometric feature vectors as inputs to obtain the prediction residuals, until the difference between the predicted residuals output by the lightweight neural network and the residual labels converges, thus serving as the residual correction module. Finally, the baseline output module and the residual correction module are combined to obtain the final temperature field prediction model. This divide-and-conquer structure retains the interpretability and stability of the physical baseline model while leveraging the powerful fitting ability of the lightweight neural network to compensate for the error between the baseline temperature and the measured temperature data, effectively solving the problems of insufficient accuracy and weak generalization ability of purely data-driven models.

[0062] Figure 6 A flowchart illustrating the output of the calibrated temperature field by the temperature field prediction model provided in this application embodiment, the flowchart may include the following steps: Step S5171: The reference temperature output by the reference output module at the same spatial coordinate point in the temperature field prediction model is added to the prediction residual output by the residual correction module to obtain the calibrated temperature value at the same spatial coordinate point.

[0063] Specifically, for coordinates The reference temperature output by the reference output module The predicted residuals output by the residual correction module Add together, final point The temperature value is: Step S5173: Traverse all spatial coordinate points to obtain the calibrated temperature values ​​of all spatial coordinate points, and output the calibrated temperature field.

[0064] Specifically, due to the reference temperature generated by the reference output module The predicted residual field generated by the residual correction module All are spatial coordinates It is a continuous function, therefore the calibrated temperature value can be directly calculated at any spatial coordinate point. To obtain the calibrated temperature field for the entire region, it is necessary to traverse all spatial coordinate points within the region, i.e., perform the above summation operation for each spatial coordinate point. This process ensures the spatial continuity and integrity of the calibrated temperature field, enabling the temperature field prediction model to output high-fidelity temperature estimates for any spatial coordinate point, providing a reliable data foundation for subsequent thermal analysis, control, or optimization.

[0065] This embodiment obtains the calibrated temperature value of a point by superimposing the reference temperature at the same spatial coordinate point with the predicted residual. By traversing all spatial coordinate points within the spatial region and superimposing the reference temperature with the corresponding predicted residual, a complete and continuous high-precision calibrated temperature field is synthesized. This method effectively combines the stability of physical constraints with the flexibility of the residual correction module, significantly improving the prediction accuracy and local detail capture capability of the calibrated temperature field. It can also compensate for deviations caused by operating condition disturbances or modeling errors in real time. The final output calibrated temperature field has stronger realism and reliability, providing solid data support for the precise monitoring, analysis, and optimization of the motor's operation.

[0066] Figure 7 A flowchart illustrating the transfer learning fine-tuning method provided in this application embodiment is included, and the process may include the following steps: Step S531: Configure the weights of the physical benchmark model as initial weights, fix the weights of the input encoding module of the physical benchmark model, and configure the weights of the decoding layer of the physical benchmark model as trainable.

[0067] Specifically, the physical baseline model trained on massive simulation data is first loaded, and its weight parameter β will be used as the initial value for this training. In order to prevent overfitting due to insufficient measured data, the weight of the front-end input encoding module of the physical baseline model is fixed to maintain its function of extracting geometric and feature parameters. Only the back-end decoding layer is changed, and the weight of the back-end is set to be adjustable so that it can be adapted according to the measured data.

[0068] Step S533: Using the measured temperature data as supervision, construct a total loss function that integrates the second data loss and the physical constraint loss, where the second data loss characterizes the error between the reference temperature output by the physical reference model and the measured temperature data.

[0069] Specifically, the weights of the decoding layer are updated using a small amount of measured data, and a total loss function that combines the second data loss and the physical constraint loss is defined: in, The predicted temperature values ​​given by the physical baseline model. Predicted temperature values ​​based on measured temperature data. Compared with measured temperature data The difference between them is the second data loss. For correction factor, The loss parameter is based on physical laws (consistent with the physical constraint loss in step S37; the specific acquisition process will not be repeated). It quantifies the degree to which the physical baseline model violates physical laws by calculating the residuals of the predicted values ​​on the governing equations (such as the partial differential equation of heat conduction). By introducing this term, when updating the decoding layer weights using a small amount of measured temperature data, the physical baseline model output can be guided to both approximate the measured temperature data and strictly follow the laws of physical evolution, thereby improving the generalization ability and physical consistency of the physical baseline model.

[0070] Step S535: Iteratively update the weights of the decoding layer through backpropagation until the total loss function converges, and obtain the corrected model weights.

[0071] Specifically, based on the total loss function The backpropagation algorithm is used to calculate the gradient of the loss with respect to the decoding layer weights, and the initial weights are updated accordingly. This process continues until the loss function value stabilizes or reaches the preset number of iterations, ultimately obtaining the corrected physical baseline model weights, thus achieving model calibration and correction.

[0072] Step S537: Replace the initial weights with the corrected model weights to obtain the temperature field prediction model.

[0073] Specifically, the corrected physical baseline model weights replace the initial weights to obtain the calibrated temperature field prediction model. Specifically, when the total loss function converges, the decoder layer weights β0, corrected by measured data, are obtained and loaded back into the physical baseline model, replacing the original initial weights. At this point, the temperature field prediction model has completed its transition from purely physics-driven to a data- and physics-driven model, directly outputting high-fidelity temperature field prediction results. This process not only ensures that the predicted values ​​accurately approximate the actual measured values ​​at the measured points, but also strictly adheres to physical laws such as heat conduction globally, significantly improving the generalization ability and robustness of the temperature field prediction model under sparse sample conditions, providing reliable data support for subsequent engineering analysis.

[0074] This embodiment constructs a total loss function that integrates the second data loss and the physical constraint loss, uses the backpropagation algorithm to iteratively optimize the weights of the decoding layer, and loads the corrected weights back into the physical baseline model to finally obtain a calibrated temperature field prediction model. This method, relying only on a small amount of measured data, can achieve accurate approximation of predicted values ​​at observation points while ensuring that the global prediction results strictly follow physical laws such as heat conduction. It effectively improves the model's prediction accuracy, physical consistency, robustness, and generalization ability in sparse sample scenarios, providing high-fidelity data support for engineering thermal analysis.

[0075] In one optional embodiment, the structural parameters, operating data, and spatial coordinates of the motor under test are obtained; based on the structural parameters and operating data, a target geometric feature vector and a target operating condition feature vector are obtained; the target geometric feature vector, the target operating condition feature vector, and the spatial coordinates are input into the temperature field prediction model obtained by the training methods S1 to S5 to obtain the temperature value corresponding to the coordinates in the motor under test. The use case is as follows: In one optional implementation scenario, this temperature field prediction model can be successfully applied to the intelligent thermal management system of electric vehicles, providing core algorithmic support for real-time thermal safety monitoring of the motor. In actual deployment, real-time operating data such as vehicle voltage, phase current, motor speed, and torque, as well as structural parameters, are read via the Controller Area Network (CAN). These parameters serve as inputs to the temperature field prediction model. Forward inference is performed based on a pre-trained high-dimensional physical field mapping neural network (using a neural operator architecture that integrates geometric coordinate encoding). Using training data as training samples, a composite loss function integrating the first data loss and physical constraint loss is constructed. The network weights are iteratively optimized through backpropagation to establish a baseline model. Subsequently, the physical baseline model is calibrated using measured temperature data, ultimately obtaining the temperature field prediction model, which outputs the calibrated temperature field distribution. Thanks to the model's resolution independence and extremely low computational complexity, the processor can quickly calculate the continuous temperature distribution of key areas such as the stator winding ends and permanent magnets within a 1-second cycle, ensuring both efficient calculation of the global thermal field and high-precision capture of local hotspots. When the model predicts that the temperature in a certain area exceeds a preset threshold (e.g., 180℃), the processor immediately sends a thermal alarm to the vehicle controller and actively triggers a torque limiting strategy to reduce the motor's output power, effectively preventing burn-out faults such as insulation damage or permanent magnet demagnetization caused by overheating. Simultaneously, the system records the operating conditions at the time of the alarm, the predicted temperature, and actual feedback (if any, from sensors) in a log, providing valuable data for subsequent model iterations and fault analysis. This closed-loop thermal management scheme fully utilizes physical constraints to enhance the model's generalization ability and reliability, and eliminates simulation bias through calibration with measured data. This enables accurate and robust temperature prediction under complex and variable driving conditions, significantly improving the safety and reliability of electric vehicles.

[0076] In one optional implementation scenario, engineers can flexibly adjust input parameters according to design requirements, including key structural parameters such as stator outer diameter, slot geometry, and winding arrangement, while pre-setting multiple extreme operating condition curves (such as peak torque continuous operating condition and high-speed field weakening operating condition). For example, engineers can quickly compare the impact of adjusting the stator outer diameter from 200mm to 220mm on heat dissipation performance, or evaluate the temperature distribution of different slot structures (such as trapezoidal slots and circular slots) under extreme loads. This completely replaces the traditional finite element simulation process. Thanks to the continuous field mapping capability of neural operators, the model does not need to repeat tedious and time-consuming steps such as geometric modeling, mesh generation, and pre-processing settings, greatly reducing the calculation cycle. Traditional finite element simulation may take several hours or even days for a single calculation, while this model can output complete three-dimensional temperature field results in just seconds, covering the temperature distribution of key parts such as stator windings, core, and permanent magnets. This allows designers to complete complex thermal design optimization work that would take weeks using traditional methods in a very short time, thereby significantly improving R&D efficiency and shortening the development cycle of motor products.

[0077] Accordingly, please refer to Figure 8 A block diagram of a temperature field prediction model training device provided in this application embodiment, the device comprising: The data acquisition unit 101 is used to acquire training data of the motor, including training temperature data corresponding to the spatial coordinate points of the motor.

[0078] The model training unit 103 is used to construct a high-dimensional physical field mapping neural network. Using training data as training samples, a composite loss function is constructed by combining the first data loss and the physical constraint loss. The weights of the high-dimensional physical field mapping neural network are iteratively updated through backpropagation to obtain a physical benchmark model. The first data loss represents the error between the predicted temperature output by the high-dimensional physical field mapping neural network and the training temperature data.

[0079] The model calibration unit 105 is used to calibrate the physical reference model based on the measured temperature data of the motor to obtain a temperature field prediction model, and output the calibrated temperature field through the temperature field prediction model.

[0080] In some alternative implementations, the model training unit 103 is as follows: A high-dimensional physical field mapping neural network is constructed, which sequentially includes an input encoding module, a feature iteration module, and a coordinate decoding module.

[0081] Geometric feature vectors, operating condition feature vectors, spatial coordinate points, and corresponding training temperature data are extracted from the training data as training samples. The geometric feature vectors and operating condition feature vectors are concatenated into a system state vector through the input encoding module. The geometric feature vectors, operating condition feature vectors, and training temperature data are mapped to the same spatial coordinate point.

[0082] The feature iteration module improves the system state vector from low-dimensional features to high-dimensional latent features, performs global and local feature depth extraction and feature dimensionality reduction, and then outputs the predicted temperature field through the coordinate decoding module.

[0083] By utilizing the differential computation unit of a deep learning framework, the first and second derivatives of the predicted temperature field are solved. These derivatives are then substituted into the heat conduction control equations, and a physical residual tensor is constructed by combining material properties and heat source density to obtain the physical constraint loss.

[0084] A composite loss function is constructed based on the first data loss and the physical constraint loss. The weights of each module in the high-dimensional physical field mapping neural network are adjusted in reverse iteratively until convergence, thus obtaining a trained physical benchmark model.

[0085] In some alternative implementations, the model training unit 103 is as follows: The geometric feature vector is obtained by vectorizing the structural parameters of the motor, while the operating condition feature vector is obtained by nonlinearly normalizing the operating data of the motor.

[0086] In some optional implementations, the model calibration unit 105 is as follows: The weight parameters of the physical baseline model are fixed, and the physical baseline model with fixed weight parameters is used as the baseline output module.

[0087] A lightweight neural network is constructed, which uses the same operating condition feature vector and geometric feature vector as input as the benchmark output module.

[0088] The difference between the measured temperature data and the reference temperature output by the reference output module is used as the residual label. A lightweight neural network is trained until the difference between the predicted residual output by the lightweight neural network and the residual label converges, thus obtaining a trained lightweight neural network. The trained lightweight neural network is then used as a residual correction module.

[0089] The temperature field prediction model is obtained by combining the benchmark output module and the residual correction module.

[0090] In some optional implementations, the model calibration unit 105 is as follows: The calibrated temperature value at the same spatial coordinate point is obtained by adding the reference temperature output by the reference output module and the prediction residual output by the residual correction module at the same spatial coordinate point in the temperature field prediction model.

[0091] Iterate through all spatial coordinate points to obtain the calibrated temperature values ​​for all spatial coordinate points, and output the calibrated temperature field.

[0092] In some optional implementations, the model calibration unit 105 is as follows: Configure the weights of the physical baseline model as initial weights, fix the weights of the input encoding module of the physical baseline model, and configure the weights of the decoding layer of the physical baseline model as trainable. Using measured temperature data as supervision, a total loss function is constructed that integrates the second data loss and the physical constraint loss.

[0093] The weights of the decoding layer are iteratively updated through backpropagation until the total loss function converges, resulting in the corrected model weights.

[0094] The corrected model weights are used to replace the initial weights to obtain the temperature field prediction model.

[0095] In this embodiment, a temperature field prediction model training device is presented in the form of a functional unit. Here, a unit refers to an ASIC (Application Specific Integrated Circuit) circuit, a processor and memory that execute one or more software or fixed programs, and / or other devices that can provide the above functions.

[0096] Please see Figure 9 , Figure 9 This application provides a schematic diagram of the structure of a computer device, as shown in the embodiment of the present application. Figure 9 As shown, the computer device includes one or more processors 10, memory 20, and interfaces for connecting the components, including high-speed interfaces and low-speed interfaces. The components communicate with each other via different buses and can be mounted on a common motherboard or otherwise installed as needed. The processors can process instructions executed within the computer device, including instructions stored in or on memory to display graphical information of a GUI on external input / output devices (such as display devices coupled to the interfaces). In some alternative implementations, multiple processors and / or multiple buses can be used with multiple memories and multiple memory modules, if desired. Similarly, multiple computer devices can be connected, each providing some of the necessary operations (e.g., as a server array, a group of blade servers, or a multiprocessor system). Figure 9 Take a processor 10 as an example.

[0097] Processor 10 may be a central processing unit, a network processor, or a combination thereof. Processor 10 may further include a hardware chip. The hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (PLD), or a combination thereof. The programmable logic device may be a complex programmable logic device (CAMP), a field-programmable gate array (FPGA), a general-purpose array logic (GPA), or any combination thereof.

[0098] The memory 20 stores instructions executable by at least one processor 10 to cause the at least one processor 10 to perform the method shown in the above embodiments.

[0099] The memory 20 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the computer device. Furthermore, the memory 20 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some alternative embodiments, the memory 20 may optionally include memory remotely located relative to the processor 10, and these remote memories may be connected to the computer device via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0100] The memory 20 may include volatile memory, such as random access memory; the memory may also include non-volatile memory, such as flash memory, hard disk or solid-state drive; the memory 20 may also include a combination of the above types of memory.

[0101] The computer device also includes a communication interface 30 for communicating with other devices or communication networks.

[0102] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and subsequently stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It can be understood that the computer, processor, microprocessor controller, or programmable hardware includes a storage component capable of storing or receiving software or computer code, which, when accessed and executed by the computer, processor, or hardware, implements the methods shown in the above embodiments.

[0103] The apparatus and units described in the above embodiments can be implemented by a computer chip or physical entity, or by a product with a certain function. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or any combination of these devices.

[0104] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0105] Those skilled in the art will understand that the embodiments of this application can be provided as methods or apparatus. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0106] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatuses, and devices according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0107] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0108] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0109] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0110] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the apparatus embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0111] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

[0112] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for training a temperature field prediction model, characterized in that, The training method includes: Acquire training data for the motor, the training data including training temperature data corresponding to the spatial coordinate points of the motor; A high-dimensional physical field mapping neural network is constructed. The training data is used as training samples. A composite loss function is constructed by combining the first data loss and the physical constraint loss. The weights of the high-dimensional physical field mapping neural network are iteratively updated through backpropagation to obtain a physical benchmark model. The first data loss represents the error between the predicted temperature output by the high-dimensional physical field mapping neural network and the training temperature data. Based on the measured temperature data of the motor, the physical reference model is calibrated to obtain a temperature field prediction model, and the calibrated temperature field is output through the temperature field prediction model.

2. The training method according to claim 1, characterized in that, The training method for the physical benchmark model is as follows: Construct a high-dimensional physical field mapping neural network that sequentially includes an input encoding module, a feature iteration module, and a coordinate decoding module; The geometric feature vector, operating condition feature vector, spatial coordinate points, and training temperature data corresponding to the spatial coordinate points are extracted from the training data as training samples. The geometric feature vector and the operating condition feature vector are concatenated into a system state vector through the input encoding module. The geometric feature vector, the operating condition feature vector, and the training temperature data are mapped to the same spatial coordinate point. The feature iteration module performs low-dimensional feature to high-dimensional latent feature enhancement, global and local feature depth extraction and feature dimensionality reduction on the system state vector, and then the coordinate decoding module outputs the predicted temperature field. Using the differential computation unit of the deep learning framework, the first and second derivatives of the predicted temperature field are solved, substituted into the heat conduction control equation, and combined with material property parameters and heat source density to construct a physical residual tensor to obtain the physical constraint loss. A composite loss function is constructed based on the first data loss and the physical constraint loss. The weights of each module in the high-dimensional physical field mapping neural network are then adjusted in reverse iteratively until convergence, resulting in a trained physical benchmark model.

3. The training method according to claim 2, characterized in that, The geometric feature vector is obtained by vectorizing the structural parameters of the motor, and the operating condition feature vector is obtained by nonlinearly normalizing the operating data of the motor.

4. The training method according to claim 1, characterized in that, The temperature field prediction model is obtained through residual network calibration, which includes: The weight parameters of the physical benchmark model are fixed, and the physical benchmark model with fixed weight parameters is used as the benchmark output module; A lightweight neural network is constructed, wherein the lightweight neural network uses the same working condition feature vector and geometric feature vector as input as the benchmark output module; The difference between the measured temperature data and the reference temperature output by the reference output module is used as the residual label. The lightweight neural network is trained until the difference between the predicted residual output by the lightweight neural network and the residual label converges, thus obtaining the trained lightweight neural network. The trained lightweight neural network is then used as the residual correction module. The temperature field prediction model is obtained by combining the reference output module and the residual correction module.

5. The training method according to claim 4, characterized in that, The step of outputting the calibrated temperature field through the temperature field prediction model includes: The temperature value after calibration at the same spatial coordinate point is obtained by adding the reference temperature output by the reference output module at the same spatial coordinate point in the temperature field prediction model with the prediction residual output by the residual correction module. Iterate through all spatial coordinate points to obtain the calibrated temperature values ​​for all spatial coordinate points, and output the calibrated temperature field.

6. The training method according to claim 1, characterized in that, The temperature field prediction model is obtained through transfer learning fine-tuning, which includes: Configure the weights of the physical benchmark model as initial weights, fix the weights of the input encoding module of the physical benchmark model, and configure the weights of the decoding layer of the physical benchmark model as trainable. Using the measured temperature data as supervision, a total loss function is constructed that integrates the second data loss and the physical constraint loss, wherein the second data loss characterizes the error between the reference temperature output by the physical reference model and the measured temperature data; The weights of the decoding layer are iteratively updated through backpropagation until the total loss function converges, thus obtaining the corrected model weights. The corrected model weights are used to replace the initial weights to obtain the temperature field prediction model.

7. A method for predicting motor temperature, characterized in that, The prediction method includes: Obtain the structural parameters, operating data, and spatial coordinates of the motor under test; Based on the structural parameters and operating data of the motor under test, the target geometric feature vector and the target operating condition feature vector are obtained; The target geometric feature vector, the target operating condition feature vector, and the spatial coordinate point to be queried are input into the temperature field prediction model obtained by the training method of any one of claims 1-6 to obtain the temperature value corresponding to the spatial coordinate point to be queried in the motor under test.

8. A temperature field prediction model training device, characterized in that, The training device includes: A data acquisition unit is used to acquire training data of the motor, the training data including training temperature data corresponding to the spatial coordinate points of the motor; The model training unit is used to construct a high-dimensional physical field mapping neural network. Using the training data as training samples, it combines a composite loss function constructed by the first data loss and the physical constraint loss, and iteratively updates the weights of the high-dimensional physical field mapping neural network through backpropagation to obtain a physical benchmark model. The first data loss represents the error between the predicted temperature output by the high-dimensional physical field mapping neural network and the training temperature data. The model calibration unit is used to calibrate the physical reference model based on the measured temperature data of the motor to obtain a temperature field prediction model, and output the calibrated temperature field through the temperature field prediction model.

9. A computer device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the temperature field prediction model training method of any one of claims 1 to 6 or the motor temperature prediction method of claim 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the temperature field prediction model training method of any one of claims 1 to 6 or the motor temperature prediction method of claim 7.

Images