Motor temperature prediction transfer learning method based on physical constraint graph neural network

Abstract

This invention discloses a transfer learning method for motor temperature prediction based on a physically constrained graph neural network, comprising the following steps: Step S1: Constructing a labeled undirected graph based on the physical structure of the source motor; Step S2: Constructing a physically constrained graph neural network based on the labeled undirected graph; Step S3: Training the physically constrained graph neural network using the operating data and temperature data of the source motor to obtain a pre-trained model; Step S4: Constructing an undirected graph of the target motor based on its physical structure, and reusing the heat capacity parameters and thermal conductivity functions in the pre-trained model to obtain an initialized physically constrained graph neural network for the target motor; Step S5: Obtaining the temperature prediction model of the target motor. This invention, by embedding physical heat transfer laws into a transferable graph neural network structure, enables the rapid construction of a high-precision temperature prediction model with physical interpretability for new motors using only a small batch of data.

Description

Technical Field

[0001] This invention relates to the field of permanent magnet synchronous motor temperature prediction, and specifically to a transfer learning method for motor temperature prediction based on a physical constraint graph neural network. Background Technology

[0002] Currently, electrification and intelligentization have become clear trends in the development of many industries. In fields such as home appliances, electric aircraft, and electric vehicles, the electrification process is accelerating. As the core power component in electrification systems, motors face higher requirements for performance and reliability. Compared to traditional industrial applications, motors in these emerging application scenarios must possess higher power density and operating efficiency. Among the many types of motors, permanent magnet synchronous motors (PMSMs) have become the preferred choice for high-end equipment such as high-power aerospace motors due to their high power density, high efficiency, and excellent speed regulation performance. However, their high power density operation also brings serious temperature rise problems; excessively high temperatures can jeopardize the performance of permanent magnets, the lifespan of insulation materials, and even lead to system failure. Therefore, accurate monitoring and prediction of the internal temperature of the motor are crucial for ensuring the safe, reliable, and efficient operation of the system.

[0003] Currently, the main methods for monitoring motor temperature have the following limitations:

[0004] 1. Direct Measurement Method: This method directly acquires temperature data by installing temperature sensors (such as thermocouples and resistance temperature detectors) in key locations inside the motor. However, this method has significant drawbacks: the wiring required for sensor installation is complex, and the system's resistance to mechanical vibration and reliability in high-temperature and high-humidity environments is poor; under long-term operation, sensors are prone to detachment, loose connections, and insulation aging. Due to the limited space inside the motor and strict electromagnetic compatibility requirements, sensors cannot be installed in many key locations (such as between winding layers and inside permanent magnets), creating temperature observation blind spots. Furthermore, sensor maintenance and calibration usually require equipment shutdown and partial or complete disassembly, which is costly and unsuitable for scenarios requiring large-scale deployment.

[0005] 2. Lumped-parameter thermal network model method: This method divides the motor into a finite number of lumped thermal capacity nodes and constructs a thermal network model through thermal resistance connections to calculate the temperature distribution. Its limitations are: the model has a small number of observable temperature nodes and a fixed topology, making it difficult to accurately characterize the temperature field at fine scales such as between winding layers and within conductors in the slots; key parameters such as thermal capacity and thermal resistance in the model usually rely on repeated iterations of a large amount of steady-state or transient experimental data for identification, a cumbersome and costly process. More importantly, when the motor's operating conditions (such as cooling conditions) or material properties (due to aging or temperature drift) change, the calibrated model parameters will become mismatched, leading to a significant decrease in prediction accuracy and poor model adaptability.

[0006] 3. Purely Data-Driven Machine Learning Approaches: This method directly uses motor operating data and temperature data to train machine learning models (such as neural networks and support vector machines) for temperature prediction. Its main drawbacks are: model training heavily relies on large volumes of high-precision labeled data covering multiple speeds, torques, and cooling conditions, resulting in long data acquisition cycles and high costs; furthermore, the operating data contains many steady-state samples and few transient samples, and this imbalance can lead to insufficient predictive ability for dynamic processes, resulting in bias. More fundamentally, such models are "black boxes," lacking physical interpretability in their internal prediction logic, making it difficult to correlate prediction results with actual heat transfer physics. Therefore, these models are unsuitable for guiding quantitative assessments of safety margins, prediction of thermal failure root causes, and iterative optimization of motor thermal design, limiting their in-depth application in safety-critical engineering fields.

[0007] In summary, existing motor temperature monitoring and prediction methods have shortcomings in terms of deployability, model accuracy, parameter adaptability, data dependence, and physical interpretability. Therefore, there is an urgent need for a temperature prediction method that can integrate the advantages of physical mechanisms and data-driven approaches, and can be efficiently transferred between different motor models. Summary of the Invention

[0008] To overcome the shortcomings of existing technologies, the present invention aims to provide a transfer learning method for motor temperature prediction based on physical constraint graph neural networks. By embedding physical heat transfer laws into a transferable graph neural network structure, it enables the rapid construction of a high-precision temperature prediction model with physical interpretability for new motors using only a small batch of data.

[0009] To achieve the objectives of this invention, the following solution is adopted:

[0010] The transfer learning method for motor temperature prediction based on physical constraint graph neural networks includes the following steps:

[0011] Step S1: Based on the physical structure of the source domain motor, construct a labeled undirected graph. The nodes of the undirected graph represent the temperature measurement points of the motor and are labeled with type. The edges represent thermal coupling relationships.

[0012] Step S2: Based on the undirected graph of the label, construct a physical constraint graph neural network, which includes a node type embedding module, an edge-level heat conduction network module, a node heat loss network module, and a heat capacity parameter module.

[0013] Step S3: Use the operating data and temperature data of the source domain motor to train the physical constraint graph neural network to obtain a pre-trained model. The pre-trained model includes the heat capacity parameters of various nodes, the thermal conductivity function between nodes, and the node heat loss function.

[0014] Step S4: Based on the physical structure of the target motor, construct an undirected graph of the target motor, and reuse the heat capacity parameters and thermal conductivity function in the pre-trained model to obtain an initialized physical constraint graph neural network for the target motor.

[0015] Step S5: Using small-batch operating data and temperature data of the target motor, train the initialized physical constraint graph neural network for the target motor. By adjusting the node heat loss function and the thermal conductivity parameter of the newly added edges, obtain the temperature prediction model of the target motor.

[0016] Furthermore, the node types indicated by the type labels include winding type, stator core type, rotor permanent magnet type, and boundary temperature type.

[0017] Furthermore, the edge-level heat conduction network module defines the thermal conductivity between node i and node j in the following manner:

[0018]

[0019] in, This represents the thermal conductivity between node i and node j. and These represent the type labels for nodes i and j, respectively. This is the global operating condition vector. It is a learnable neural network.

[0020] Furthermore, the node heat loss network module defines the heat power injection of node i at time t in the following manner:

[0021]

[0022] in, This represents the heat power injection at node i at time t. It is a neural network. Indicates the label type of node i. This is the global operating condition vector.

[0023] Furthermore, the heat capacity parameter module learns a scalar heat capacity for each type of node. Numerical stability is ensured by truncating the lower bound during training.

[0024] Furthermore, the physical constraint graph neural network predicts temperature using the energy conservation equation:

[0025]

[0026] in, Let the edge heat flow between nodes i and j at time t be denoted as . Injecting thermal power into node i at time t. Let i be the heat capacity of node i. Let be the temperature of node i at time t.

[0027] Furthermore, the edge heat flow Defined as And it uses the graph Laplacian matrix form for efficient computation. Among them, Let the edge heat flow between nodes i and j at time t be denoted as . This represents the thermal conductivity between node i and node j. and Let J and I be the temperatures at nodes j and i respectively at time t.

[0028] Furthermore, in the process of constructing and training the physical constraint graph neural network, the loss function used includes the basic temperature prediction loss and the dynamic temporal consistency loss, and the total loss function is:

[0029]

[0030] Where L is the total loss function, Based on the loss function, For dynamic loss function, The weights are for dynamic loss.

[0031] Furthermore, during the construction and training of the physical constraint graph neural network, the input data is normalized and the predicted temperature is limited to prevent numerical divergence.

[0032] Furthermore, during the training of the initialized physical constraint graph neural network for the target motor, only the thermal conductivity of the newly added edges and the thermal loss of each node are fully trained, while the heat capacity and thermal conductivity reused from the pre-trained model are fine-tuned.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] 1. This invention integrates physical mechanisms with data-driven approaches, ensuring high prediction accuracy while endowing the model with strong physical interpretability. By discretizing the motor structure into an undirected graph labeled with physical types and constructing a graph neural network incorporating the law of energy conservation, the model's learning process is strictly constrained by physical laws. The heat capacity parameters, thermal conductivity functions, and nodal heat loss functions learned by the model all have clear physical meanings (thermal characteristic parameters and loss sources), rather than being "black box" parameters. This not only improves the temperature prediction accuracy under complex and transient operating conditions, but more importantly, the model output can be directly used for physical-level safety margin assessment, thermal failure root cause analysis, and design iteration optimization, overcoming the fundamental deficiency of purely data-driven methods in lacking physical interpretability.

[0035] 2. This invention enables the effective transfer of core physical parameters of the model between motors with different topologies, significantly reducing the modeling cost and data requirements for new models. Since the heat capacity parameters and thermal conductivity functions in the model are only related to the node material type and global operating conditions, and not to the specific node connection topology, the physical knowledge learned from the source domain motor can be directly reused in the target motor. When modeling the target motor, it is only necessary to construct a graph network using its physical structure and reuse these parameters for initialization, and then use a small batch of running data to train some parameters (such as the thermal conductivity of newly added edges and node losses). This avoids the expensive cost of conducting numerous experiments to calibrate the thermal network parameters for each new motor (such as the lumped parameter method) or collecting large amounts of labeled data covering all operating conditions (such as pure machine learning), enabling rapid deployment and application of the model.

[0036] 3. This invention enables precise modeling of the internal temperature field of a motor, effectively solving the "observation blind zone" problem in areas inaccessible to sensors. This method discretizes the motor into multiple nodes, theoretically allowing for the definition and modeling of any fine-scale area (such as between winding layers or specific points within the slots). Its precision far surpasses that of lumped-parameter thermal networks with fixed numbers of nodes. By uniformly calculating and predicting all nodes, a complete and continuous temperature field distribution inside the motor can be obtained, thereby enabling reliable temperature prediction for critical areas where sensors cannot be installed. This overcomes the observation blind zone problem caused by spatial and electromagnetic compatibility limitations in direct measurement methods.

[0037] 4. The model of this invention possesses stronger adaptability and robustness to different operating conditions, mitigating parameter mismatch issues. The thermal conductivity function and nodal heat loss function in the model take global operating conditions (such as speed, current, and voltage) as input, enabling dynamic response to changes in motor operating status. When cooling conditions or load change, the model parameters can be automatically adjusted through the input operating condition vector, rather than using fixed values ​​as in lumped parameter thermal network models. This enhances the model's adaptability under different operating conditions and reduces model parameter mismatch and prediction failure problems caused by operating condition drift.

[0038] 5. This invention improves long-term operational reliability and maintenance convenience. This invention is a prediction method based on software algorithms, completely avoiding a series of reliability and maintainability problems (the drawbacks of direct measurement methods) caused by installing physical sensors inside the motor, such as complex wiring, poor vibration resistance, easy aging and detachment, and the need for downtime disassembly for maintenance. It provides a stable and low-cost solution for non-invasive, online, large-scale monitoring of motor temperature status. Attached Figure Description

[0039] Fig. 1This is a flowchart of the motor temperature prediction transfer learning method based on a physical constraint graph neural network in an embodiment of the present invention;

[0040] Fig. 2 This is a flowchart illustrating one embodiment of the present invention;

[0041] Fig. 3 This is a schematic diagram of transfer learning in an embodiment of the present invention. Detailed Implementation

[0042] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0043] like Figs. 1 to 3 As shown, this embodiment of the invention provides a motor temperature prediction transfer learning method based on a physical constraint graph neural network. This method facilitates the transfer of temperature prediction models between motors with different topologies, while also improving model interpretability and prediction accuracy. The motor temperature prediction transfer learning method based on a physical constraint graph neural network in this embodiment includes the following steps:

[0044] Step S1: Based on the physical structure of the source domain motor, construct a labeled undirected graph. The nodes of the undirected graph represent the temperature measurement points of the motor and are labeled with type. The edges represent thermal coupling relationships.

[0045] Step S2: Based on the undirected graph of the label, construct a physical constraint graph neural network, which includes a node type embedding module, an edge-level heat conduction network module, a node heat loss network module, and a heat capacity parameter module.

[0046] Step S3: Use the operating data and temperature data of the source domain motor to train the physical constraint graph neural network to obtain a pre-trained model. The pre-trained model includes the thermal capacity parameters of various nodes, the thermal conductivity function between nodes, and the node heat loss function.

[0047] Step S4: Based on the physical structure of the target motor, construct an undirected graph of the target motor, and reuse the heat capacity parameters and thermal conductivity functions in the pre-trained model to obtain an initialized physical constraint graph neural network for the target motor.

[0048] Step S5: Using small-batch operating data and temperature data of the target motor, train the initialized physical constraint graph neural network for the target motor. By adjusting the node heat loss function and the thermal conductivity parameter of the newly added edges, obtain the temperature prediction model of the target motor.

[0049] The following section will further elaborate on the motor temperature prediction transfer learning method based on physical constraint graph neural networks in this embodiment:

[0050] In this embodiment, the graph neural network under physical constraints is constructed first, and the construction process is as follows:

[0051] Graph neural networks mainly consist of four parts: node type embedding, edge-level heat conduction network, node heat loss network, and heat capacity parameter. The motor is discretized into a labeled undirected graph G=(V,E), where the node set V represents temperature measurement points, and the edge set E represents thermal coupling or spatial adjacency relationships. Each node carries a type label. Such as windings Stator core type Rotor permanent magnet type Boundary temperature category The temperature of node i at each time t. Global operating condition vector (Including rotational speed, d-axis voltage and current, q-axis voltage and current, and torque). First, the type of each node is embedded into the graph neural network.

[0052] This represents the vectorized representation of the i-th node after embedding the type label.

[0053] Furthermore, a side-level heat conduction network is constructed. The learnable thermal conductivity function is parameterized, as shown in equation (1).

[0054]

[0055] In the formula, This represents the thermal conductivity between node i and node j. and Let i and j represent the type labels of nodes, respectively. Since thermal conductivity is always positive, the activation function is Softplus. Thermal conductivity is independent of the specific graph topology, depending only on the node type and the global condition, thus achieving transferable weights.

[0056] Furthermore, a node heat loss network is constructed. The heat sources / losses inside the nodes are predicted, as shown in equation (2):

[0057]

[0058] This represents the heat power injection at node i at time t. This is a neural network. The formula means that a conditional neural network is used to map "node type information + current global condition" to the internal heat source (non-negative scalar) of the node at that moment.

[0059] Finally, the heat capacity parameters are trained. Scalar heat capacity is learned according to node type. Node heat capacity The number of parameters is kept to grow linearly with the number of types rather than with the number of nodes.

[0060] Understandably, due to and The input consists only of node type embeddings and the global condition; the weights are independent of the node sequence or total number. (The last sentence appears to be incomplete and unrelated to the preceding text.) and By copying to the new topology, minor adjustments can be made, thereby reducing the annotation cost of the new model.

[0061] Energy is conserved at each node, as shown in equation (3):

[0062]

[0063] In the formula, The edge heat flow between nodes i and j is specifically defined as shown in equation (4).

[0064]

[0065] To concentrate the "heat conduction contribution of each edge" into "net heat inflow to each node", we introduce the graph Laplacian operator L. We define the graph Laplacian matrix L(W).

[0066]

[0067] In the formula, Let W be a column vector with all elements equal to 1, and let the weight matrix W (symmetric) satisfy the following condition:

[0068]

[0069] The value is 0 if there is no edge connecting the two nodes.

[0070] Therefore, for node i, combining equations (4) and (5), the net heat inflow into all nodes can be expressed as follows:

[0071]

[0072] Substituting equation (7) into equation (3) yields the matrix form.

[0073]

[0074] Where C is a diagonal matrix composed of the heat capacities of node i, i.e.

[0075]

[0076] Furthermore, the discrete-time explicit Euler integral is expressed as follows:

[0077]

[0078] In the formula, For explicit Euler step size.

[0079] In implementation,

[0080]

[0081] The output from the neural network is non-negative thanks to the Softplus activation function. The heat capacity C is learned by type and truncated with a lower bound to ensure numerical stability.

[0082] In this embodiment, the pre-trained model is transferred to the target domain motor, and the implementation is as follows:

[0083] During model construction, for each node, the inputs for thermal conductivity and power loss are only the node type embedding and global operating conditions; the weights are independent of the node sequence or total number. Copying the same thermal conductivity and power loss to the new topology allows for minor fine-tuning, thus reducing the annotation cost for new models. Specifically, for the motor to be trained, heat transfer nodes are first divided into the three categories mentioned above: winding type... Stator core type Rotor permanent magnet type For each type of node, the heat capacity parameter C of the pre-trained model is reused. Specifically, for the experimental motor, the winding end... With the inside of the winding For winding type, slot , tank wall , bottom of the trough For stator core type, rotor For rotor permanent magnets, as shown in equation (12).

[0084]

[0085] Furthermore, a heat transfer node graph network for the motor to be trained is constructed. For parts with the same heat transfer structure as the pre-trained model, the thermal conductivity of the corresponding edges can be reused. For newly added parts, only the thermal conductivity of the new heat transfer edges needs to be trained separately. Specifically, for the experimental motor, only the winding interior and winding ends need to be retrained. Tank bottom and tank wall Tank walls and tank openings The thermal conductivity of the connecting edges between these nodes.

[0086] The process of constructing the loss function is as follows:

[0087] Basic loss function As in equation (13)

[0088]

[0089] In the formula, W is the total number of sample points included in the loss, B represents the batch size, S is the sequence length, and I is the target node index set. The predicted temperature for that node, and This represents the actual temperature of the node.

[0090] To encourage time dynamism, a dynamic loss is added. The time difference is defined as follows:

[0091]

[0092] Then there is a dynamic loss function. ,

[0093]

[0094] In the formula, This represents the total number of sample points included in the dynamic loss; the meanings of the other symbols are the same as above.

[0095] Therefore, the total loss of the model is as follows:

[0096]

[0097] In the formula, The weights are for dynamic loss.

[0098] Taking the motor data used in the experiment as an example, the complete operation is as follows:

[0099] First, the source domain data is normalized by dividing all temperature values ​​by 200 and all non-temperature values ​​by their respective maximum values. Simultaneously, the geometric mean of voltage and current along the d-axis and q-axis is calculated, and the voltage and current amplitudes are used for feature engineering. The pre-processed source domain data is then used for full-parameter training of the pre-trained model. To avoid numerical divergence, the predicted temperature is limited to the normalized range of -5 to +5. Gradient clipping, learning rate scheduling, and early stopping mechanisms are also added. The same preprocessing is applied to the target motor data, and a graph network for the target motor is constructed. Based on the heat capacity of the three types of nodes and the thermal conductivity of the edges connecting nodes obtained from the pre-trained model, the heat capacity and the thermal conductivity of the edges connecting the same nodes are fine-tuned. Only the losses of each node and the thermal conductivity of the edges connecting newly added nodes are fully trained.

[0100] This embodiment of the motor temperature prediction transfer learning method based on a physically constrained graph neural network first constructs the motor as a labeled undirected graph and embeds the physical heat transfer equations into the graph neural network structure. Through training, the thermal capacity parameters of various nodes, the thermal conductivity of the edges connecting nodes, and the power loss of nodes are obtained, thereby improving the prediction accuracy and physical interpretability of the model. Based on this, the method can reuse the node thermal capacity parameters and edge thermal conductivity learned in the pre-trained model to quickly construct a graph network suitable for new motors. Furthermore, by using only a small batch of operating temperature rise data from the target motor for training, a multi-node temperature prediction model for the motor can be obtained, achieving efficient knowledge transfer and rapid modeling between motors with different topologies.

[0101] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A transfer learning method for motor temperature prediction based on a physical constraint graph neural network, characterized in that, Includes the following steps: Step S1: Based on the physical structure of the source domain motor, construct a labeled undirected graph. The nodes of the undirected graph represent the temperature measurement points of the motor and are labeled with type. The edges represent thermal coupling relationships. Step S2: Based on the undirected graph of the label, construct a physical constraint graph neural network, which includes a node type embedding module, an edge-level heat conduction network module, a node heat loss network module, and a heat capacity parameter module. Step S3: Use the operating data and temperature data of the source domain motor to train the physical constraint graph neural network to obtain a pre-trained model. The pre-trained model includes the heat capacity parameters of various nodes, the thermal conductivity function between nodes, and the node heat loss function. Step S4: Based on the physical structure of the target motor, construct an undirected graph of the target motor, and reuse the heat capacity parameters and thermal conductivity function in the pre-trained model to obtain an initialized physical constraint graph neural network for the target motor. Step S5: Using small-batch operating data and temperature data of the target motor, train the initialized physical constraint graph neural network for the target motor. By adjusting the node heat loss function and the thermal conductivity parameter of the newly added edges, obtain the temperature prediction model of the target motor.

2. The motor temperature prediction transfer learning method based on physical constraint graph neural network according to claim 1, characterized in that, The node types indicated by the type labels include winding type, stator core type, rotor permanent magnet type, and boundary temperature type.

3. The motor temperature prediction transfer learning method based on physical constraint graph neural network according to claim 1, characterized in that, The heat capacity parameter module learns a scalar heat capacity for each type of node. Numerical stability is ensured by truncating the lower bound during training.

4. The motor temperature prediction transfer learning method based on physical constraint graph neural network according to claim 1, characterized in that, The physical constraint graph neural network predicts temperature using the energy conservation equation: in, Let the edge heat flow between nodes i and j at time t be denoted as . Injecting thermal power into node i at time t. Let i be the heat capacity of node i. Let be the temperature of node i at time t.

5. The motor temperature prediction transfer learning method based on physical constraint graph neural network according to claim 4, characterized in that, The edge heat flow Defined as And it uses the Graph Laplacian matrix form for efficient computation, where, Let the edge heat flow between nodes i and j at time t be denoted as . This represents the thermal conductivity between node i and node j. and Let J and I be the temperatures at nodes j and i respectively at time t.

6. The motor temperature prediction transfer learning method based on physical constraint graph neural network according to claim 1, characterized in that, In constructing and training the physical constraint graph neural network, the loss functions used include basic temperature prediction loss and dynamic time consistency loss, and the total loss function is: Where L is the total loss function, Based on the loss function, For dynamic loss function, The weights are for dynamic loss.

7. The motor temperature prediction transfer learning method based on physical constraint graph neural network according to claim 1, characterized in that, During the construction and training of the physical constraint graph neural network, the input data is normalized and the predicted temperature is limited to prevent numerical divergence.

8. The motor temperature prediction transfer learning method based on physical constraint graph neural network according to claim 1, characterized in that, During the training of the initialized physical constraint graph neural network for the target motor, only the thermal conductivity of the newly added edges and the thermal loss of each node are fully trained, while the heat capacity and thermal conductivity reused from the pre-trained model are fine-tuned.

Images