A graph neural network-based method and system for sensing the thermal state of motors, incorporating physical mechanisms.

Abstract

This invention proposes a graph neural network-based method and system for sensing the thermal state of a motor, integrating physical mechanisms. The method first uses a low-order thermal network model to solve the thermal balance equation to obtain a physical temperature reference. Then, the thermal network topology is mapped to a graph neural network structure to establish a physical correspondence between information transmission and heat conduction. Spatial thermal coupling features are extracted through graph convolutional layers, and a bidirectional gated recurrent network is used to capture the thermal inertia evolution features over time. Finally, a residual learning strategy is employed to construct a composite loss function containing physical constraints for optimization. The data-driven prediction residual is linearly superimposed with the physical reference temperature to achieve high-precision estimation. This invention effectively overcomes the shortcomings of traditional thermal network models, such as difficulty in parameter identification and the lack of physical interpretability in purely data-driven models. It significantly improves the robustness and accuracy of temperature monitoring of key motor components and is suitable for embedded real-time control systems.

Description

Technical Field

[0001] This invention relates to the fields of motor condition monitoring and artificial intelligence, specifically to a graph neural network method and system for sensing the thermal state of a motor that integrates physical mechanisms. Background Technology

[0002] Electric motors, with their high power density, superior efficiency, and excellent dynamic response characteristics, have become the preferred core power components in fields such as electric vehicle traction systems, aerospace actuators, and precision industrial servo drives. As modern industrial applications continuously demand higher performance, electric motors are increasingly being miniaturized and lightweight. This trend significantly increases internal power loss density, leading to increasingly severe thermal loads. Under continuous high-load operation or extreme conditions, excessively high temperatures in critical internal components can cause degradation of stator insulation materials and irreversible demagnetization of permanent magnets, ultimately leading to motor failure or system collapse. Therefore, accurate and real-time temperature monitoring of critical internal components is crucial for ensuring operational safety and extending the service life of motor systems.

[0003] However, directly measuring the real-time temperature of the motor inside key components faces significant challenges in practical engineering applications. First, the compact structure and complex electromagnetic environment of motors make sensor integration difficult. Embedding contact sensors in the stator windings not only increases manufacturing complexity and cost but may also damage the integrity of the stator insulation system, thus affecting motor reliability. Second, traditional contact measurement methods are unsuitable for high-speed rotating magnets, while non-contact technologies such as infrared thermal imaging and wireless telemetry are limited by limited installation space and electromagnetic interference, making them difficult to apply in complex industrial environments. Therefore, temperature estimation methods based on readily available electrical signals such as current, voltage, and speed have gradually become a research hotspot, offering advantages such as non-invasiveness, low cost, and ease of integration.

[0004] Existing methods for estimating motor temperature can be broadly categorized into three types: physics-based, data-driven, and hybrid methods combining physics and data. Physics-based methods utilize thermodynamic principles to establish a thermal behavior model of the motor. Among these, finite element analysis (FEM) can accurately describe the three-dimensional temperature field distribution inside the motor, but its high computational complexity and long solution time limit its application in real-time online estimation. In contrast, low-order thermal network (LRN) methods discretize the continuous temperature field into a finite number of thermal nodes, describing heat generation, transfer, and dissipation through thermal resistance-capacitance networks. LRN models offer advantages such as high computational efficiency and clear physical meaning, and have been widely used in motor thermal analysis. However, due to the simplification of the lumped assumption and the time-varying nonlinear characteristics of thermal parameters, LRN models still face challenges in achieving high-precision prediction of local hot spots under different operating conditions.

[0005] Data-driven approaches directly establish the mapping relationship between operating variables and temperature through machine learning algorithms, without requiring explicit physical modeling. However, data-driven approaches lack physical interpretability, and unconstrained predictions may violate thermodynamic principles, limiting their reliability in safety-critical applications.

[0006] To overcome these limitations, Physical Information Neural Networks (PINs) have attracted widespread attention as a hybrid modeling paradigm. PINs embed physical laws into the neural network through loss function constraints or structural priors, ensuring the physical consistency of the estimation results while maintaining the flexibility of data-driven approaches. However, existing PIN methods are mainly used for solving continuous-field partial differential equations, failing to fully utilize the topological structure of thermal networks, making the deep integration of physical mechanisms and data-driven approaches challenging.

[0007] Graph neural networks (GNNs) propagate and aggregate information between nodes through information transfer mechanisms, naturally modeling relational structures and providing a promising approach to addressing the aforementioned challenges. The information aggregation process in GNNs has a natural structural correspondence with the heat transfer mechanism in thermal networks, enabling the explicit capture of nonlinear coupling relationships between hot nodes.

[0008] In summary, accurate temperature estimation and real-time monitoring of critical internal components of motors in high-performance electromechanical systems are crucial. Therefore, this invention proposes a graph neural network-based method and system for sensing the thermal state of motors, incorporating physical mechanisms, to achieve high-precision estimation of the motor's internal temperature. Summary of the Invention

[0009] The purpose of this invention is to provide a graph neural network method and system for sensing the thermal state of a motor that integrates physical mechanisms, so as to achieve high-precision estimation of the internal temperature of the motor.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] In a first aspect, the present invention provides a graph neural network method for sensing the thermal state of a motor that integrates physical mechanisms, the method comprising the following steps: S1. Use the thermal network model to identify the key component nodes that characterize the thermal dynamics of the motor and the heat transfer paths between them, and solve the thermal balance equation of the thermal network model to obtain the physical reference temperature prediction of each thermal node. S2. Map the physical topology of the thermal network to the graph structure of the graph neural network, configure input features for the nodes and edges of the graph neural network, and establish a physical correspondence between the neural network information transmission mechanism and the heat transfer process. S3. Based on the input features, spatial thermal coupling features are extracted through the message passing and node update mechanism of the graph convolutional layer of the graph neural network to simulate the dynamic distribution process of heat inside the motor on the spatial topology. S4. The spatial thermal coupling features are sequentially fed into a bidirectional gated recurrent network to capture the dynamic evolution features of motor thermal inertia and time dimension, and the prediction residual is obtained through the output layer. S5. Introduce a residual learning strategy to construct a composite loss function containing physical mechanism constraints for optimization; and linearly superimpose the predicted residual with the physical reference temperature for compensation to achieve accurate estimation of the internal temperature of the motor.

[0012] Further, step S1 includes: Based on the thermal network topology, the principle of energy conservation is applied to any hot node. i The rate of temperature change is described by the heat balance equation, expressed as:

[0013] in, For nodes i The equivalent heat capacity, For nodes i temperature, For time variables, For nodes j temperature, For nodes i Heating power at the point, For nodes i With nodes j The equivalent thermal resistance between them Represents nodes i The set of adjacent nodes that are thermally connected; Based on a four-node low-order thermal network model of coolant, stator core, stator windings, and permanent magnet, using coolant temperature... As boundary conditions, the heat balance equation is expressed in state space as follows:

[0014] Where C is the heat capacity matrix and T is the state vector. Let P be the thermal conductivity matrix, and P be the heat source vector. This represents the boundary input matrix.

[0015] Furthermore, in step S2, establishing the physical correspondence between the neural network information transmission mechanism and the heat transfer process specifically includes: Based on the connection relationships of the thermal network model, an adjacency matrix of a graph neural network is constructed to establish a one-to-one mapping isomorphic relationship between graph nodes and motor components, and between graph edges and heat transfer paths; the thermal network is represented as a graph with attributes. ,in It is a set of nodes. It is an edge set. It is the node feature matrix. is the edge feature matrix.

[0016] Furthermore, step S3 includes: Combining the source node features and the edge features , and integrating the physical properties of the heat transfer path, expressed as:

[0017] where represents the message vector passed from the source node to the target node j in the i -th layer; is the ReLU activation function; , are learnable parameters, represents the feature dimension of the node, represents the dimension of the edge feature; represents vector concatenation; Aggregating the information of all neighbor nodes of node and updating by combining this information with its own features, expressed as:

[0018] where represents the feature vector of node 1 in the updated i -th layer, representing the extracted spatial heat coupling feature; represents the current feature of the central node in the i -th layer; and are learnable parameters; Performing global average pooling on all node features to obtain the graph-level representation.

[0019] Furthermore, step S4 includes: Introducing a bidirectional gated recurrent unit after the graph convolutional layer to capture the temporal dependence of the motor temperature, and concatenating the bidirectional hidden states of the final time step to obtain the temporal feature vector; Directly passing the initial features of the stator winding and permanent magnet nodes to the output layer, and fusing them with the temporal features after linear transformation; Mapping the fused features to the residual prediction through a multi-layer perceptron to obtain the residual correction value, that is, the prediction residual.

[0020] Furthermore, step S5 includes: The measured temperature t representing time and the corresponding timet Temperature prediction using a low-order thermal network model Perform linear operations to obtain the calculated residuals , represented as:

[0021] Based on the constructed computational residuals, the loss function for network training is defined as the mean square error between the computational residuals and the predicted residuals of the network output, expressed as:

[0022] Where N represents the number of training set samples; For the first o Calculate the residuals for each sample; For the first o Individual sample prediction residuals; Introducing an L2 regularization term to constrain the magnitude of network parameters, and constructing the total loss function. , represented as:

[0023] in, For the first Layer weight matrix; It is the Frobenius norm; It is the regularization coefficient; The temperature estimation problem is defined as a combination of a physical reference temperature and data-driven prediction residuals, expressed as:

[0024] in, For the estimated temperature, For the prediction of the physical reference temperature obtained through the thermal network model; The prediction residuals are obtained by the data-driven method.

[0025] Secondly, embodiments of the present invention also provide a graph neural network-based motor thermal state sensing system that integrates physical mechanisms. Applying the above-described method, it achieves accurate estimation of the internal temperature of the motor. The system includes: a physical reference temperature acquisition module, a topology mapping module, a spatial feature extraction module, a predicted temperature residual module, and a temperature estimation module, wherein: The physical reference temperature acquisition module is used to identify key component nodes and heat transfer paths using a low-order thermal network model, and to solve the heat balance equation to obtain the physical reference temperature. The topology mapping module is used to map the physical topology of the thermal network to the structure of the graph neural network, establishing a physical correspondence between the information transmission mechanism and the heat transfer process. The spatial feature extraction module is used to extract spatial thermal coupling features through the message passing and node update mechanism of the graph convolutional layer, and to simulate the dynamic distribution process of heat inside the motor on the spatial topology. The temperature residual prediction module is used to feed the spatial thermal coupling features sequentially into a bidirectional gated recurrent network to capture the dynamic evolution features of the motor's thermal inertia and time dimension, and predict the temperature residual through the output layer. The temperature estimation module is used to linearly superimpose the residual correction value with the physical reference temperature to achieve accurate estimation of the internal temperature of the motor.

[0026] Thirdly, embodiments of the present invention also provide an electronic device, including a processor and a memory, wherein the memory stores machine-executable instructions that can be executed by the processor, and the processor executes the machine-executable instructions to implement the above-described method.

[0027] As can be seen from the above technical solution, the present invention discloses a graph neural network method and system for sensing the thermal state of a motor that integrates physical mechanisms, which has the following advantages compared with the prior art: 1. This invention constructs a physical information hybrid driving model that integrates low-order thermal networks and graph neural networks, unifying the physical mechanism of heat conduction in a graph topology. Based on a physical topology mapping method, an isomorphic relationship is established between the neural network information transmission path and the actual heat flow path. Graph convolutional layers are used to aggregate and extract the nonlinear spatial thermal coupling features between nodes. Furthermore, a bidirectional gated recurrent network is combined to deeply capture the dynamic evolution of the motor's thermal inertia and time dimension. Finally, based on a residual learning strategy and a physical mechanism-constrained loss function, systematic deviations generated by the low-order thermal network are identified and compensated, facilitating accurate estimation of the temperature of key components of the permanent magnet synchronous motor.

[0028] 2. This invention complements and enhances the explicitness of the physical mechanism model and the nonlinear fitting capability of the deep learning model within the same architecture, effectively overcoming the prediction accuracy decline problems caused by the difficulty in parameter identification, structural simplification, and aging drift of traditional low-order thermal network models. It also avoids the "black box" uninterpretability and poor generalization ability defects of purely data-driven models due to the lack of physical constraints. It significantly improves the robustness of temperature estimation under varying loads, dynamic operating conditions, and parameter perturbation environments. By introducing physical mechanism constraints, it ensures that the prediction results conform to the laws of thermodynamics, avoiding logical errors under abnormal operating conditions.

[0029] 3. The present invention has a compact overall structure, high precision and strong real-time performance, and is suitable for embedded motor control systems with limited computing power. It can significantly improve the reliability of full-range temperature monitoring of permanent magnet synchronous motors without the need to add additional temperature sensors, and has extremely high engineering application value.

[0030] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0031] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0032] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0033] Figure 1 This is a schematic flowchart of a graph neural network method for sensing the thermal state of a motor that integrates physical mechanisms, provided as an embodiment of the present invention.

[0034] Figure 2 This diagram illustrates a performance comparison between the present invention and different temperature estimation methods.

[0035] Figure 3 This is a schematic diagram comparing the stator winding temperature estimation results of the present invention and different models on the test set.

[0036] Figure 4 This is a schematic diagram comparing the permanent magnet temperature estimation results of the present invention and different models on the test set.

[0037] Figure 5 This is a schematic diagram of the electronic device structure provided by the present invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0039] In the description of this invention, it should be noted that some processes described in this application specification and drawings include multiple operations that appear in a specific order. However, it should be clearly understood that these operations may be performed in any order or in parallel. Furthermore, various numbers are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0040] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0041] See Figure 1 As shown in the figure, this invention discloses a graph neural network method for sensing the thermal state of a motor that integrates physical mechanisms, mainly including the following steps: S1. Use a low-order thermal network model to identify key component nodes that characterize the thermal dynamics of the motor and the heat transfer paths between them, and solve the thermal balance equation of the thermal network model to obtain the physical temperature prediction of each thermal node. S2. Map the physical topology of the thermal network to the structure of the graph neural network, and establish the physical correspondence between the information transmission mechanism of the neural network and the heat transfer process. S3. Spatial thermal coupling features are extracted through the message passing and node update mechanism of graph convolutional layers to simulate the dynamic distribution process of heat inside the motor on the spatial topology. S4. Spatial thermal coupling features are sequentially fed into a bidirectional gated recurrent network to capture the dynamic evolution features of motor thermal inertia and time dimension, and the temperature residual is predicted through the output layer. S5. Introduce a residual learning strategy, construct a composite loss function containing physical mechanism constraints for optimization, and linearly superimpose the predicted residual (residual correction value) with the physical reference temperature for compensation, so as to achieve accurate estimation of the internal temperature of the motor.

[0042] This invention constructs a graph neural network architecture based on the topology and thermodynamic structure of a low-order thermal network, and adopts a residual learning strategy to learn the system deviations in the physical model. It combines physical mechanism knowledge with data-driven error correction capabilities to develop a graph neural network method and system for sensing the thermal state of a motor that integrates physical mechanisms to estimate the internal temperature of the motor. When estimating the motor temperature, it can obtain an accurate temperature for control.

[0043] This embodiment uses the temperature of the stator windings and permanent magnets in a motor as an example for illustration, as detailed below: Step 1: Based on the motor geometry and thermodynamic mechanism, identify key thermal dynamic nodes, including stator windings, stator core, permanent magnets and housing, and determine the thermal resistance path that characterizes the heat exchange capacity between each node. Based on the law of conservation of energy, establish the thermal balance equation for each node.

[0044] Step 2: Construct a graph neural network input structure that is topologically isomorphic to the low-order thermal network model. Map the identified key component nodes to vertices of the graph structure and the thermal resistance paths to edges of the graph structure. Extract the thermal resistance physical parameters from the low-order thermal network model and encode them inversely. Embed these parameters as edge weight features into the adjacency relationship of the graph neural network, thereby establishing a one-to-one mapping relationship between the neural network information transmission channel and the actual physical heat transfer process, and realizing the constraint and initialization of the network connection structure by physical prior knowledge.

[0045] Step 3: Collect the current, voltage, speed and power loss data of the motor as node feature input, and use the graph convolutional layer to execute the message passing mechanism based on physical constraints; by calculating the feature differences between neighboring nodes and the central node and combining the edge weight features for weighted aggregation, simulate the heat conduction behavior between different components based on temperature difference and thermal resistance. Then, update the node state through a nonlinear activation function to extract the spatial thermal coupling features that reflect the complex nonlinear heat exchange law inside the motor.

[0046] Step four: Construct a bidirectional gated recurrent unit network. Input the extracted spatial thermal coupling features into the network in time series. Use the forward and reverse hidden layer structures to bidirectionally model the historical thermal state and future evolution trend, capturing the thermal inertia characteristics and time lag law of the motor due to the thermal capacity effect under dynamic load changes. Then, perform nonlinear mapping on the spatiotemporal fusion features through the fully connected output layer to predict the residual correction value used to compensate for the physical model parameter offset and structural simplification error.

[0047] Step five introduces a residual learning strategy, deconstructing the temperature estimation task into a process of correcting the systemic bias between the main physical temperature rise process calculated by the thermal network model and the systemic bias fitting by the data-driven model. A composite loss function containing a mean squared error term and a regularization constraint term is constructed to guide network parameter optimization by penalizing prediction results that violate the law of conservation of energy. The trained network is then used to perform forward calculations on the real-time acquired motor operating characteristics, outputting the temperature residual correction values ​​for each node. These residual correction values ​​are then linearly superimposed and reconstructed to achieve high-precision estimation of the temperature of key motor components.

[0048] The specific implementation method of this embodiment will be described in detail below: Step one involves using a low-order thermal network model to identify key component nodes characterizing the motor's thermal dynamics and the heat transfer paths between them, and solving the thermal balance equations of the thermal network model to obtain the predicted physical reference temperature for each thermal node; this includes: Step 1.1: Based on the thermal network topology, apply the principle of energy conservation to any thermal node. i The rate of temperature change is described by the heat balance equation, expressed as:

[0049] in, For nodes i The equivalent heat capacity, For nodes j temperature, For time, For nodes i temperature, For nodes i Heating power at the point, For nodes i With nodes j The equivalent thermal resistance between them Represents nodes i The set of adjacent nodes that are thermally connected.

[0050] Step 1.2: Based on a four-node low-order thermal network model of coolant, stator core, stator winding, and permanent magnet, the coolant temperature is used... As boundary conditions, the heat balance equation is expressed in state space as follows:

[0051] Where T is the state vector, P is the heat source vector, and C is the heat capacity matrix, expressed as:

[0052] in, Indicates the temperature of the stator core. This indicates the temperature of the stator winding. This indicates the temperature of the permanent magnet. This indicates the power loss of the stator core. This indicates the power loss of the stator winding. This indicates the power loss of the permanent magnet. This represents the equivalent heat capacity of the stator core. This represents the equivalent heat capacity of the stator winding. This represents the equivalent heat capacity of a permanent magnet.

[0053] In low-order thermal network models, the heat source terms are physically related to the motor's operating variables. Stator winding losses... Generated by Joule heating, it is proportional to the sum of the squares of the dq-axis currents, expressed as:

[0054] in, Stator phase resistance, and These are the d-axis and q-axis current components, respectively.

[0055] Stator core loss Including hysteresis loss and eddy current loss, it is approximated by the Steinmetz empirical formula, expressed as:

[0056] in, It is the electrical frequency. B It is the amplitude of magnetic flux density. and These are the hysteresis and eddy current loss coefficients, respectively. It is the Steinmetz index, which typically ranges from 1.6 to 2.0.

[0057] Permanent magnet loss Primarily caused by harmonic magnetic fields, it is approximately proportional to the square of the rotational speed, expressed as:

[0058] in, It is the rotor's mechanical angular velocity. The loss coefficient is determined through experimental calibration or finite element analysis.

[0059] Thermal conductivity matrix Based on the network topology, its elements are determined by the thermal conductivity between nodes, and are expressed as:

[0060] in, This indicates the thermal conductivity between the coolant and the stator core. Indicates the thermal conductivity between the stator core and the stator windings. This indicates the thermal conductivity between the stator core and the permanent magnet. This indicates the thermal conductivity between the coolant and the permanent magnet; The boundary input matrix B includes the effect of coolant temperature, and B is represented as:

[0061] Given an initial state (0) and the time series of input variables at time t It can be achieved through numerical integral equations Obtain the predicted physical baseline temperature for each thermal node.

[0062] Step two involves mapping the physical topology of the low-order thermal network to the structure of a graph neural network, establishing a physical correspondence between the neural network's information transmission mechanism and the heat transfer process; this includes: Step 2.1: Construct the adjacency matrix of the graph neural network based on the connection relationships of the low-order thermal network model, and establish a one-to-one mapping isomorphic relationship between graph nodes and motor components, and between graph edges and heat transfer paths; formally represent the network as a graph with attributes. ,in It is a set of nodes. It is an edge set. It is the node feature matrix. It is the edge feature matrix.

[0063] Step 2.2: To enable the graph neural network to capture the physical state of each hot node, a node feature vector containing thermodynamic information was designed. For each node... Its eigenvectors Defined as:

[0064] in, This represents the temperature rise of a node relative to its environment. This represents the nodes predicted by the hot network model. i temperature, Indicates ambient temperature. It is the average temperature gradient between the node and its neighbors. It is the power of the heat source. This represents the time derivative of the predicted temperature in a heat network, which describes the dynamic trend of temperature change. This represents the motor operating variables associated with each node, including current amplitude, speed, and dq-axis current components. The feature vectors of all nodes are stacked row-wise to form the node feature matrix. , Represents the dimension of node features.

[0065] Step 2.3, for connecting nodes and The edge ( , This will encode the feature vector of the edge features of the physical properties of the heat transfer path between nodes. Defined as:

[0066] in, This represents the temperature difference between nodes, reflecting the driving force of heat transfer. This represents the nodes predicted by the hot network model. j temperature, It is a heat flow direction indicator; its symbol reflects the direction of heat flow. It is used to regulate temperature difference Regarding the direction of heat flow The degree of sensitivity, It is a thermal resistance type code used to distinguish different heat transfer mechanisms (such as conduction, air gap convection, and surface convection). This represents the heat source power of the source node. The eigenvectors of all edges are stacked to form the edge feature matrix. , This represents all effective thermal resistance paths between the connecting nodes. Represents the edge feature dimension.

[0067] Step 3: Extract spatial thermal coupling features through the message passing and node update mechanism of the graph convolutional layer to simulate the dynamic distribution of heat inside the motor on the spatial topology; including: Step 3.1, calculate the features combined with the source node. Sum of edge features The physical properties of the integrated heat transfer path are represented as follows:

[0068] in, Indicates the first In the layer, from the source node j Transmitted to the target node i The message vector; It is the ReLU activation function; , These are learnable parameters. Represents the feature dimension of a node. The dimension representing the edge features; This represents vector concatenation. This represents the characteristics of neighboring nodes. This allows physical information such as temperature gradients and heat flow directions encoded in the edge features to directly modulate the information flow between nodes.

[0069] Step 3.2, aggregate nodes from all neighboring nodes. The information is updated by combining this information with its own characteristics, and is represented as:

[0070] in, Indicates the updated number Layer nodes i The feature vector represents the extracted spatial thermal coupling features; Indicates the first Layer center node i Current characteristics; and These are learnable parameters; layer normalization and dropout regularization are applied after each graph convolutional layer to enhance training stability.

[0071] Step 3.3, in After each graph convolutional layer, the features of all nodes are processed. Global average pooling is used to obtain a graph-level representation, which is represented as:

[0072] in, It is in time The graph-level eigenvectors.

[0073] Step four involves sequentially feeding the spatial thermal coupling features into a bidirectional gated recurrent network to capture the dynamic evolution characteristics of the motor's thermal inertia and time dimension, and predicting the temperature residual through the output layer; this includes: Step 4.1: After the graph convolutional layer, a bidirectional gated recurrent unit is introduced to capture the time dependence of motor temperature. The bidirectional hidden states of the final time step are concatenated to obtain the temporal feature vector. , represented as:

[0074] in, This represents the forward hidden state, indicating the state from the beginning of the sequence. Start scanning and continue until the current time. Subsequent accumulated time characteristic information, This indicates a reverse hidden state, representing the time evolution information from the current moment back to the past; Indicates the length of the time window.

[0075] Step 4.2: The initial features of the stator windings and permanent magnet nodes are directly transferred to the output layer, and after linear transformation, they are fused with the timing features to obtain the fused features. , represented as:

[0076] in, This represents the initial splicing characteristics of the stator winding and permanent magnet nodes. This represents the characteristic vector of the stator winding at the initial moment. This represents the feature vector of a permanent magnet node at the initial moment; and These are learnable parameters.

[0077] Step 4.3: Map the fused features to the residual prediction using a multilayer perceptron, as shown below:

[0078] in, This represents the residual prediction of the stator winding and permanent magnet temperatures. This represents the predicted temperature residual of the stator winding. This represents the predicted temperature residual of the permanent magnet. and This represents the learnable weight matrix of the first and second layers of the network. and This represents the bias terms of the first and second layers of the network. This is the ReLU activation function.

[0079] Step 5 involves introducing a residual learning strategy to construct a composite loss function that includes physical mechanism constraints. The residual correction value is then linearly superimposed with the physical reference temperature to achieve high-precision estimation of the motor's internal temperature. This includes: Step 5.1, In actual operation, the actual temperature of key motor components. Baseline predictions from low-order thermal network models Prediction errors are inevitable. This invention systematically deconstructs the prediction error into four physical components:

[0080] in, This represents the actual motor temperature. This indicates the predicted temperature of the thermal network. Indicates parameter identification deviation. This represents the structural simplification error caused by the lumped parameter assumption. This represents the nonlinear effect caused by temperature-dependent thermal parameters. This represents the measurement noise. The first three error components are deterministically correlated with the motor operating conditions and constitute a systematic error:

[0081]

[0082] in, It is a nonlinear mapping that describes the unmodeled dynamics of a thermal network. It is the running state vector. It is a temperature state vector.

[0083] Transform its objective function into fitting the aforementioned systematic error. Since the thermal network model has captured the main evolutionary trends of the motor's thermal dynamics, the fluctuation range of the residuals is […]. min , max The range is much smaller than the actual temperature range [T]. min T max ],in, This represents the maximum fluctuation of the residual. This represents the minimum fluctuation of the residual. This represents the maximum value of the actual temperature. It is the minimum value of the true temperature; that is, it satisfies:

[0084] This significantly compresses the output mapping space of the neural network, effectively reducing the dependence of deep learning models on massive amounts of data and network complexity, while accelerating the convergence process of the model under varying conditions.

[0085] Step 5.2 defines the loss function for network training as the mean square error between the predicted residual and the calculated residual, expressed as:

[0086] in, Let N represent the mean squared error loss function, and let N represent the training set containing N samples. , for the first o The residuals for each sample are calculated. For the first o The measured temperature of each sample For the first o Temperature prediction using a low-order thermal network model for a single sample; For the first o The prediction residuals of each sample.

[0087] Step 5.3 introduces an L2 regularization term to constrain the magnitude of the network parameters, and combines these two terms to represent the total loss function. , represented as:

[0088]

[0089] in, It is the regularization loss. For the first Layer weight matrix; It is the Frobenius norm; It is the regularization coefficient.

[0090] Step 5.4 defines the temperature estimation problem as a combination of physical foundations and data-driven residual correction:

[0091] in, For the estimated temperature, To ensure that the temperature baseline prediction (physical reference temperature) obtained through the low-order thermal network model adheres to the physical bottom line such as energy conservation under various operating conditions; The correction values ​​(i.e. prediction residuals) obtained from the data-driven method for the thermal network model prediction accurately compensate for the systematic prediction bias caused by the aging of physical parameters and unmodeled nonlinear dynamics.

[0092] As shown above, in this embodiment, when using a graph neural network to perceive the thermal state of a motor by integrating physical mechanisms, firstly, a hybrid driving model integrating low-order thermal networks and graph neural networks is constructed to uniformly represent the physical mechanism of motor heat conduction within a graph topology. Based on the physical topology mapping method, an isomorphic relationship is established between the neural network information transmission path and the actual heat flow path. Graph convolutional layers are used to aggregate and extract the nonlinear spatial thermal coupling features between nodes. Then, a bidirectional gated recurrent network is combined to deeply capture the dynamic evolution of motor thermal inertia and the time dimension. Finally, based on a residual learning strategy and a physical mechanism-constrained loss function, systematic deviations are identified and compensated for, achieving accurate estimation of the temperature of key components of the permanent magnet synchronous motor.

[0093] Furthermore, the performance comparison results of this invention with different temperature estimation methods can be found in [reference]. Figure 2 As shown, the method of this invention (Physics-Informed Graph Neural Network, PI-GNN) is compared with residual prediction models without graph structure, including the Gated Recurrent Unit (LPTN-GRU), Lumped Parameter Thermal Network-Long Short-Term Memory (LPTN-LSTM), BackPropagation Neural Network (LPTN-BPNN), and Random Forest (LPTN-RF) algorithms based on low-order thermal networks. Furthermore, these physically-driven residual models are compared with data-driven direct prediction models LSTM, GRU, BPNN, and RF, demonstrating the advantages of the residual learning framework guided by low-order thermal networks. As can be seen from the comparison, the method of the present invention exhibits better temperature estimation performance, indicating that the introduction of graph structure can more effectively characterize the internal thermal coupling relationship of the system, while the residual learning framework guided by thermal network can further take into account both physical prior constraints and data fitting ability.

[0094] Further, see Figure 3 and Figure 4As shown, a visual comparison of temperature estimation results from different models on the test set is presented. Even under non-ideal conditions such as difficulties in parameter identification, structural simplification errors, material aging drift, and complex variable load conditions in low-order thermal networks, this invention can still achieve accurate sensing of the stator winding and rotor temperatures of a motor, possessing significant theoretical research value and extremely high engineering application potential. It achieves high-precision, robust, and physically consistent estimation of the temperatures of key internal components of a permanent magnet synchronous motor, significantly improving the prediction accuracy of stator winding and rotor temperatures while maintaining physical consistency.

[0095] Furthermore, this embodiment of the invention also provides a graph neural network-based motor thermal state sensing system that integrates physical mechanisms. Applying the above-described method, it achieves accurate estimation of the motor's internal temperature. The system includes: a physical reference temperature acquisition module, a topology mapping module, a spatial feature extraction module, a predicted temperature residual module, and a temperature estimation module, wherein: The physical reference temperature acquisition module is used to identify key component nodes and heat transfer paths using a low-order thermal network model, and to solve the heat balance equation to obtain the physical reference temperature. The topology mapping module is used to map the physical topology of the thermal network to the structure of the graph neural network, establishing a physical correspondence between the information transmission mechanism and the heat transfer process. The spatial feature extraction module is used to extract spatial thermal coupling features through the message passing and node update mechanism of the graph convolutional layer, and to simulate the dynamic distribution process of heat inside the motor on the spatial topology. The temperature residual prediction module is used to feed the spatial thermal coupling features sequentially into a bidirectional gated recurrent network to capture the dynamic evolution features of the motor's thermal inertia and time dimension, and predict the temperature residual through the output layer. The temperature estimation module is used to linearly superimpose the residual correction value with the physical reference temperature to achieve accurate estimation of the internal temperature of the motor.

[0096] The system provided in this embodiment of the invention has the same implementation principle and technical effects as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the system embodiment can be referred to the corresponding content in the aforementioned method embodiment, and will not be repeated here.

[0097] As described in the above embodiments, this invention provides a graph neural network-based method and system for sensing the thermal state of a motor, integrating physical mechanisms. First, a hybrid driving model integrating a low-order thermal network and a graph neural network is constructed, unifying the physical mechanism of heat conduction in the motor within a graph topology. Based on a physical topology mapping method, an isomorphic relationship is established between the neural network information transmission path and the actual heat flow path. Graph convolutional layers are used to aggregate and extract the nonlinear spatial thermal coupling features between nodes. Then, a bidirectional gated recurrent network is combined to deeply capture the dynamic evolution of the motor's thermal inertia and time dimension. Finally, based on a residual learning strategy and a physical mechanism-constrained loss function, systematic deviations generated by the low-order thermal network are identified and compensated, achieving accurate estimation of the temperature of key components of the permanent magnet synchronous motor.

[0098] This invention complements and enhances the explicitness of physical mechanism models and the nonlinear fitting capabilities of deep learning models within the same architecture, effectively overcoming the prediction accuracy degradation problems caused by difficulties in parameter identification, structural simplification, and aging drift in traditional low-order thermal network models. It also avoids the "black box" uninterpretability and poor generalization ability of purely data-driven models due to the lack of physical constraints. It significantly improves the robustness of temperature estimation under varying loads, dynamic operating conditions, and parameter perturbations. By introducing physical mechanism constraints, it ensures that the prediction results conform to the laws of thermodynamics, avoiding logical errors under abnormal operating conditions. This method has a compact overall structure, combining high accuracy and strong real-time performance, making it suitable for embedded motor control systems with limited computing power. It can significantly improve the reliability of full-domain temperature monitoring of permanent magnet synchronous motors without adding additional temperature sensors, and has extremely high engineering application value.

[0099] Additionally, refer to Figure 5 As shown, this embodiment of the invention also provides an electronic device that can perform the above-described method. The electronic device may include a processor 10, a memory 11, a communication bus 12, and a communication interface 13, and may also include a computer program stored in the memory 11 and capable of running on the processor 10.

[0100] In some embodiments, the processor 10 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 10 is the control unit of the electronic device, connecting various components of the entire electronic device through various interfaces and lines. It executes programs or modules stored in the memory 11 and calls data stored in the memory 11 to perform various functions of the electronic device and process data.

[0101] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, electronic devices, or computer program products, etc. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention 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.

[0102] It should be noted that the word "comprising" does not exclude the presence of components or steps not listed in the claims. The words "a" or "an" preceding a component do not exclude the presence of a plurality of such components. This invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer.

[0103] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0104] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A graph neural network method for sensing the thermal state of a motor that integrates physical mechanisms, characterized in that, The method includes the following steps: S1. Use the thermal network model to identify the key component nodes that characterize the thermal dynamics of the motor and the heat transfer paths between them, and solve the thermal balance equation of the thermal network model to obtain the physical reference temperature prediction of each thermal node. S2. Map the physical topology of the thermal network to the graph structure of the graph neural network, configure input features for the nodes and edges of the graph neural network, and establish the physical correspondence between the neural network information transmission mechanism and the heat transfer process. S3. Based on the input features, spatial thermal coupling features are extracted through the message passing and node update mechanism of the graph convolutional layer of the graph neural network to simulate the dynamic distribution process of heat inside the motor on the spatial topology. S4. The spatial thermal coupling features are sequentially fed into a bidirectional gated recurrent network to capture the dynamic evolution features of motor thermal inertia and time dimension, and the prediction residual is obtained through the output layer. S5. Introduce a residual learning strategy, construct a composite loss function containing physical mechanism constraints for optimization, and linearly superimpose the predicted residual with the physical reference temperature for compensation, so as to achieve accurate estimation of the internal temperature of the motor.

2. The method according to claim 1, characterized in that, Step S1 includes: Based on the thermal network topology, the principle of energy conservation is applied to any hot node. i The rate of temperature change is described by the heat balance equation, expressed as: in, For nodes i The equivalent heat capacity, For nodes i temperature, For time, For nodes j temperature, For nodes i Heating power at the point, For nodes i With nodes j The equivalent thermal resistance between them Represents nodes i The set of adjacent nodes that are thermally connected; Based on a four-node low-order thermal network model of coolant, stator core, stator windings, and permanent magnets, using coolant temperature... As boundary conditions, the heat balance equation is expressed in state space as follows: Where C is the heat capacity matrix and T is the state vector. Let P be the thermal conductivity matrix, and P be the heat source vector. This represents the boundary input matrix.

3. The method according to claim 1, characterized in that, In step S2, establishing the physical correspondence between the neural network information transmission mechanism and the heat transfer process specifically includes: Based on the connection relationships of the thermal network model, an adjacency matrix of a graph neural network is constructed to establish a one-to-one mapping isomorphic relationship between graph nodes and motor components, and between graph edges and heat transfer paths; the thermal network is represented as a graph with attributes. ,in It is a set of nodes. It is an edge set. It is the node feature matrix. It is the edge feature matrix.

4. The method according to claim 3, characterized in that, Step S3 includes: Combining source node features Sum of edge features The physical properties of the integrated heat transfer path are represented as follows: in, Indicates the first In the layer, from the source node j Transmitted to the target node i The message vector; It is the ReLU activation function; , These are learnable parameters. Represents the feature dimension of a node. The dimension representing the edge features; This indicates vector concatenation; Aggregate nodes from all neighboring nodes The information is updated by combining this information with its own characteristics, and is represented as: in, Indicates the updated number Layer nodes i The feature vector represents the extracted spatial thermal coupling features; Indicates the first Layer center node i Current characteristics; and These are learnable parameters; Global average pooling is performed on all node features to obtain a graph-level representation.

5. The method according to claim 1, characterized in that, Step S4 includes: A bidirectional gated recurrent unit is introduced after the graph convolutional layer to capture the time dependence of motor temperature, and the bidirectional hidden states of the final time step are concatenated to obtain the temporal feature vector. The initial characteristics of the stator windings and permanent magnet nodes are directly transmitted to the output layer and then fused with the timing characteristics after linear transformation. The fused features are mapped to residual predictions through a multilayer perceptron to obtain residual correction values, i.e., prediction residuals.

6. The method according to claim 5, characterized in that, Step S5 includes: Represent the time Measured temperature With the corresponding time Physical reference temperature prediction of the thermal network model Perform a linear operation to obtain the calculation residual , which is expressed by the formula as: Based on the constructed computational residuals, the loss function for network training is defined as the mean square error between the computational residuals and the predicted residuals of the network output, expressed as: Where N represents the number of training set samples; For the first o Calculate the residuals for each sample; For the first o The prediction residuals for each sample; Introducing an L2 regularization term to constrain the magnitude of network parameters, and constructing the total loss function. , represented as: in, For the first Layer weight matrix; It is the Frobenius norm; It is the regularization coefficient; The temperature estimation problem is defined as a combination of a physical reference temperature and data-driven prediction residuals, expressed as: in, For the estimated temperature, For the prediction of the physical reference temperature obtained through the thermal network model; The prediction residuals are obtained by the data-driven method.

7. A graph neural network-based motor thermal state sensing system integrating physical mechanisms, characterized in that, The system includes: a physical reference temperature acquisition module, a topology mapping module, a spatial feature extraction module, a predicted temperature residual module, and a temperature estimation module. It uses the method described in any one of claims 1-6 to achieve accurate estimation of the internal temperature of the motor.

8. An electronic device, characterized in that, It includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, the processor executing the machine-executable instructions to perform the method as described in any one of claims 1-6.

Images