Temperature field prediction network training, equipment temperature field prediction method and related equipment

Abstract

The present application discloses a method for training a temperature field prediction network, and equipment for temperature field prediction, and related equipment, relating to the field of temperature field prediction. The method includes: obtaining temperature data of observation points; constructing a distance adjacency graph using observation points and target points as nodes; establishing heat transfer paths between nodes to obtain a heat network adjacency graph; weighting the distance adjacency graph and the heat network adjacency graph to obtain a weighted directed graph; masking the temperature data of some observation points in the weighted directed graph to obtain an interpolated data set; using the masked observation points and non-masked observation points in the interpolated data set to train a temperature field prediction network to obtain a temperature field prediction model; the temperature field prediction network is a relational graph convolutional network. The present application uses a weighted directed graph to predict the temperature field, and uses a relational graph convolutional network model to capture the spatial and temporal dependencies between nodes, achieving high-precision spatiotemporal dependency expression, thereby predicting the temperature of an unknown target point using only limited actual temperature data.

Description

Technical Field

[0001] The present application relates to the field of temperature field prediction, and in particular to a temperature field prediction network training, equipment temperature field prediction method and related equipment. Background Art

[0002] In equipment health maintenance, performance and safety are the core research contents. By monitoring the temperature field at key locations of the equipment and combining it with equipment operation and maintenance strategies, not only can the efficiency and performance of the equipment be improved, but the service life of the equipment can also be extended and safety hazards can be avoided. At present, temperature sensors are the best means to monitor the temperature of equipment. However, in order to obtain the detailed and true internal temperature distribution of the equipment, a large number of sensors are required, which increases the complexity of assembly and the difficulty of internal design. In the small space of the equipment, the arrangement and maintenance of sensors are difficult, resulting in unstable temperature monitoring. In actual applications, due to cost and technical limitations, the number of sensors is limited. Usually, the ratio of sensors to equipment components is about 1:10, which makes it difficult to measure the temperature of some key components directly using temperature sensors. Summary of the Invention

[0003] The purpose of this application is to provide a temperature field prediction network training, equipment temperature field prediction method and related equipment, which only relies on part of the actual temperature monitoring data to predict the temperature field of monitoring target points with economic management value but difficult to arrange temperature sensors.

[0004] To achieve the above objectives, this application provides the following solutions:

[0005] In a first aspect, the present application provides a temperature field prediction network training method, comprising:

[0006] Acquire temperature data of an observation point; the observation point is a temperature monitoring point of the equipment where the sensor is installed;

[0007] The observation point and the target point are used as nodes to construct a distance adjacency graph; the target point is a temperature field prediction target point;

[0008] Establishing heat transfer paths between the nodes to obtain a heat network adjacency graph;

[0009] Weighting the distance adjacency graph and the heat network adjacency graph to obtain a weighted directed graph;

[0010] masking the temperature data of some observation points in the weighted directed graph to obtain an interpolated data set;

[0011] A temperature field prediction network is trained using masked observation points and non-masked observation points in the interpolation data set to obtain a temperature field prediction model; the temperature field prediction network is a relational graph convolutional network; the loss function of the temperature field prediction network includes an interpolation loss function and a reconstruction loss function; the interpolation loss function is constructed using masked observation points; the reconstruction loss function is constructed using non-masked observation points; the masked observation points are masked observation points in the weighted directed graph; the non-masked observation points are non-masked observation points in the weighted directed graph.

[0012] In a second aspect, the present application provides a method for predicting equipment temperature field, comprising:

[0013] Get the temperature data of the observation point in the current stage;

[0014] The temperature data of the observation point in the current stage is input into a temperature field prediction model to obtain the temperature data of the predicted target point; the temperature field prediction model is trained by any of the temperature field prediction network training methods described above.

[0015] In a third aspect, the present application provides a computer device comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement any one of the temperature field prediction network training methods or equipment temperature field prediction methods described above.

[0016] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements any of the above-mentioned temperature field prediction network training methods or equipment temperature field prediction methods.

[0017] In a fifth aspect, the present application provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the above-mentioned temperature field prediction network training methods or equipment temperature field prediction methods.

[0018] According to the specific embodiments provided in this application, this application discloses the following technical effects:

[0019] The present application provides a temperature field prediction network training method, comprising: obtaining temperature data of an observation point; the observation point is a temperature monitoring point of an equipment where a sensor is installed; constructing a distance adjacency graph using the observation point and a target point as nodes; the target point is a temperature field prediction target point; establishing a heat transfer path between each of the nodes to obtain a heat network adjacency graph; weighting the distance adjacency graph and the heat network adjacency graph to obtain a weighted directed graph; masking the temperature data of some observation points in the weighted directed graph to obtain an interpolation data set; training a temperature field prediction network using the masked observation points and non-masked observation points in the interpolation data set to obtain a temperature field prediction model; the temperature field prediction network is a relational graph convolutional network; the loss function of the temperature field prediction network includes an interpolation loss function and a reconstruction loss function; the interpolation loss function is constructed using masked observation points; the reconstruction loss function is constructed using non-masked observation points; the masked observation points are masked observation points in the weighted directed graph; the non-masked observation points are unmasked observation points in the weighted directed graph. This application uses a weighted directed graph to predict temperature fields, and a relational graph convolutional network model to capture the spatial and temporal dependencies between nodes and automatically update the connection strength between nodes to achieve high-precision expression of spatiotemporal dependencies, thereby predicting the temperature of unknown target points using only limited actual temperature data. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0021] Figure 1 A flow chart of a temperature field prediction network training method provided in one embodiment of the present application;

[0022] Figure 2 A schematic diagram showing comparison of regression curves after interpolation of target points and artificially masked observation points provided in one embodiment of the present application;

[0023] Figure 3 A schematic flow chart of a method for predicting equipment temperature field provided in another embodiment of the present application;

[0024] Figure 4 A schematic diagram of a technical route for a method for predicting equipment temperature field provided in another embodiment of the present application;

[0025] Figure 5 A schematic diagram of raw data provided for another embodiment of the present application;

[0026] Figure 6 A schematic diagram of a heat network adjacency graph and a weighted adjacency graph provided in another embodiment of the present application.

[0027] Figure 7 A schematic diagram of the structure of a computer device provided in one embodiment of the present application. DETAILED DESCRIPTION

[0028] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0029] Performance and safety are core research areas in equipment health maintenance. Monitoring temperature fields at key locations on equipment and integrating them with equipment operation and maintenance strategies can not only improve equipment efficiency and performance, but also extend its service life and mitigate potential safety hazards. Currently, temperature sensors are the best means of monitoring equipment temperature. However, obtaining a detailed and accurate picture of the equipment's internal temperature distribution requires a large number of sensors, increasing assembly complexity and internal design challenges. Sensor placement and maintenance are difficult within the confined space of equipment, leading to unstable temperature monitoring. This also makes it difficult to directly measure the temperatures of some key components using temperature sensors. In practical applications, cost and technical constraints limit the number of sensors, typically with a sensor-to-equipment component ratio of approximately 1:10. Due to the stability of thermodynamics, heat transfer within equipment is analyzable, meaning that temperature field predictions can be performed using a limited number of monitoring points where temperature sensors are easily accessible. Current research on temperature field prediction primarily focuses on solving partial differential equations (PDEs) using methods such as finite difference and finite element methods. However, these methods are often computationally expensive, time-consuming, and lack timely response. Data-driven approaches have also been applied to solving PDEs and predicting temperature fields. For example, multilayer perceptrons and convolutional neural networks are used to learn mappings from limited measurement data to high-dimensional states, but they often overlook the hidden temporal and spatial correlations between temperature monitoring data, reducing prediction accuracy. To reduce reliance on large amounts of data, some studies have combined deep learning with physical information to propose physically informed neural networks (PINNs). PINNs estimate the parameters of PDEs using solutions to inverse problems, and their success relies heavily on the accuracy of the physical equations used. However, in practical applications, accurately handling boundary and initial conditions is a significant challenge. The handling of these conditions directly impacts the accuracy and stability of predictions.

[0030] The Relational Graph Convolutional Network (RGCN) has attracted much attention due to its ability to simultaneously capture spatial and temporal series features and connectivity relationships. RGCN utilizes a graph structure to represent the relationships between different monitoring points and effectively captures the spatiotemporal dependencies between nodes through a recurrent neural network with a graph convolution gating mechanism. In temperature field prediction, RGCN can process node or edge features in the graph and extract neighborhood information with higher precision. Currently, most temperature field predictions use simulated data to construct datasets. However, simulated data is generated based on mathematical models or assumptions and may not be fully applicable to the actual environment. Real sensor data records temperature changes under actual working conditions, and this information is crucial for model training and prediction. However, the noise, outliers, and random fluctuations contained in real data are often simplified or ignored in simulated data. Therefore, using real data for interpolation can ensure that the dataset is closer to reality and reflects the true patterns of temperature changes.

[0031] In summary, a method for predicting the temperature field of equipment based on limited measured data and spatiotemporal interpolation is proposed. This method utilizes real sensor data from a limited number of monitoring points and explores the heat transfer characteristics between temperature nodes by constructing spatial dynamic adjacency relationships that account for real heat transfer information. A prediction model that considers spatiotemporal dependencies is constructed based on a relational graph convolutional network to predict the temperature field of key components located in confined areas of the equipment. This method not only facilitates real-time monitoring of the temperature state of key equipment locations, preventing thermal runaway events and extending the service life of the equipment, but also eliminates reliance on traditional physical field simulation results, avoiding idealized prediction results and saving computational and time costs. Therefore, it is of great significance to propose a method for predicting the temperature field of monitoring target points that are economically valuable but difficult to deploy temperature sensors, relying only on a portion of actual temperature monitoring data.

[0032] The present application provides a method for predicting the temperature field of equipment based on limited measured data and spatiotemporal relationship interpolation. First, the actual temperature sensor data of the key positions of the equipment is obtained. Then, based on the physical heat transfer information and spatiotemporal dependency characteristics, a static graph of the heat network and a spatial topological adjacency graph are constructed. The relational graph convolutional network model is used to capture the spatial and temporal dependencies between nodes, and the connection strength between nodes is automatically updated to achieve high-precision spatiotemporal dependency expression. The trained model is used to predict the temperature field of key components in small positions of the equipment. This method addresses the problem that the temperature at key positions of equipment is difficult to monitor due to factors such as vibration, dust, and small space. Only limited actual temperature data is used to predict the temperature of unknown target points, avoiding reliance on traditional physical field simulation and saving calculation and time costs. The method based on the dynamic adjacency graph enables the model to adapt to different working conditions, accurately capture the complex heat exchange mechanism inside the equipment, and improve the flexibility and accuracy of the prediction. This method is suitable for real-time or near real-time health monitoring of equipment and is convenient for use in engineering practice.

[0033] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0034] In an exemplary embodiment, Figure 1 As shown, a temperature field prediction network training method is provided. The method is executed by a computer device, specifically a computer device such as a terminal or a server, or a terminal and a server. In the embodiment of the present application, the method is applied to a server as an example for explanation, including the following steps S1 to S6. Among them:

[0035] This embodiment uses a set of 185 hours of recorded public data sets of permanent magnet synchronous motors (PMSMs) to experimentally verify the temperature field prediction network training method proposed in this application.

[0036] This publicly available dataset of 185 hours of recorded permanent magnet synchronous motor (PMSM) data consists of multiple sensor data collected from a PMSM deployed on a test bench. The test bench measurements were collected by the LEA department at Paderborn University. All data was sampled at 2 Hz. Features in this dataset include two dimensions of voltage and current, motor speed, coolant temperature, ambient temperature, motor torque, stator winding temperature measured with thermocouples, stator tooth temperature measured with thermocouples, stator yoke temperature measured with thermocouples, permanent magnet temperature measured with thermocouples and wirelessly transmitted via a thermal imaging device, and the sequence number of each measurement, for a total of 13 feature dimensions.

[0037] like Figure 1 As shown in FIG, a 185-hour public data set of records of a permanent magnet synchronous motor (PMSM) used in the experimental example is used. This embodiment includes the following steps:

[0038] S1. Acquire temperature data of an observation point; the observation point is a temperature monitoring point of an equipment where a sensor is installed.

[0039] In this embodiment, the positions of equipment temperature monitoring observation points (abbreviated as observation points) and temperature field prediction target points (abbreviated as target points) are marked in combination with equipment layout and operating conditions.

[0040] The specific steps for marking observation and target points are: select locations on key equipment components that are economically valuable but difficult to place temperature sensors, set several evenly distributed target points, and select observation points at spatial topological locations adjacent to the target points or locations on other adjacent components that are suitable for placing temperature sensors.

[0041] In this example, the stator yoke temperature feature of the permanent magnet synchronous motor described above is selected as the target point, which is considered to be truly missing and does not participate in any training. Observation points are divided into two categories: stator tooth temperature and stator winding temperature are selected to simulate artificially masked observation points, and power, motor speed, coolant temperature, ambient temperature, and permanent magnet temperature are selected to simulate unmasked observation values.

[0042] Place temperature sensors at observation points of key equipment components, and record temperature monitoring timing data and related operating condition information.

[0043] When arranging temperature sensors, specifically, select temperature sensors that are appropriate for the measurement range, response time, and accuracy requirements and can withstand the equipment's operating conditions (such as vibration, shock, and temperature fluctuations), such as thermocouples, RTDs, or infrared temperature sensors. Record temperature-related operating parameters, such as the equipment's ambient temperature, speed, power, coolant temperature, and lubricant oil temperature.

[0044] S2. Constructing a distance adjacency graph using the observation point and the target point as nodes; the target point is a temperature field prediction target point.

[0045] Specifically, the method includes: taking the observation point and the target point as nodes; calculating the Euclidean distance between each of the nodes; normalizing the Euclidean distance to obtain the topological weight of the distance adjacency graph; calculating the Pearson correlation coefficient between each of the nodes; adding the topological weight of the distance adjacency graph and the Pearson correlation coefficient and normalizing them to obtain the weight value of each edge in the distance adjacency graph; and constructing a distance adjacency graph using the weight values ​​of the nodes and the edges.

[0046] In this example, a distance adjacency graph is constructed: the observation point and the target point are used as nodes in the adjacency graph. The spatial topological distance between the nodes is calculated to obtain the distance adjacency graph topological weight. The data correlation between the nodes is calculated to obtain the distance adjacency graph correlation weight. The two are added and normalized to obtain the final weight value of each edge in the distance adjacency graph.

[0047] Specifically, the spatial topological distance between nodes is calculated by using, but not limited to, Euclidean distance and performing min-max normalization (deviation standardization) to obtain the topological weight d i ' ,j ; Three-dimensional coordinates (x i ,y i ,z i ) Euclidean distance d i,j and topological weight d i ' ,j The formula is as follows:

[0048]

[0049] Specifically, the data correlation between nodes is calculated by using, but not limited to, the Pearson correlation coefficient r to calculate the data correlation between observation points and obtain the correlation weight. The Pearson correlation coefficient r is calculated as follows: for the kth feature T of node i ik , calculate T ik The mean Standard deviation σ Ti ; Calculate the covariance Cov(T i ,T j ), and finally get the Pearson correlation coefficient r between nodes i,j .

[0050]

[0051]

[0052] Calculate and get the final element a of the distance adjacency graph A i,j Specifically, the topological weight and the correlation weight are added together and min-max normalization is performed again.

[0053]

[0054] Among them, the results of the Pearson coefficient are not normalized; the numerical calculation of the distance adjacency graph: after adding the topological weight and the elements at the same position in the correlation weight matrix, the new matrix is ​​normalized to obtain a new matrix with a 0-1 distribution.

[0055] S3. Establish heat transfer paths between the nodes to obtain a heat network adjacency graph. Based on the principles of thermodynamics and heat transfer, determine whether a heat transfer path exists between the nodes to obtain a heat transfer determination result. Calibrate the heat transfer direction in the heat transfer determination result to obtain a calibrated heat transfer result. Based on the calibrated heat transfer result, construct a heat network adjacency matrix with a 0-1 distribution to obtain a heat network adjacency graph.

[0056] In this embodiment, a heat network adjacency graph is constructed. Based on the lumped parameter heat network method, a heat transfer path between nodes is established. The specific steps to obtain the heat transfer path between nodes are: based on the lumped parameter heat network method, reasonable boundary conditions (such as temperature, heat flux density, etc.) are determined, the heat transfer path is reasonably simplified and assumed based on the actual situation, and the nonlinear factors in the heat transfer process are processed as necessary (the heat exchange across nodes is ignored in this embodiment). According to the principles of thermodynamics and heat transfer, it is determined whether there is a heat transfer path between nodes and the heat transfer direction is calibrated (the determination here is manual and requires expert knowledge), and a 0-1 distributed heat network adjacency matrix B is constructed, where the element b in B i,j It is expressed as follows:

[0057]

[0058] Among them, the reasonable boundary condition determination process requires the assistance of expert knowledge. This embodiment relies on the thermal network construction process in the paper Short-Term Prediction Method of Transient Temperature Field Variation for PMSM inElectric Drive Gearbox Using Spatial-Temporal Relational Graph ConvolutionalThermal Neural Network (doi:10.1109 / TIE.2023.3303650). The ambient temperature is regarded as a node in the paper, and there is thermal radiation (heat exchange) with other nodes; therefore, the boundary conditions are not considered in this embodiment. However, it is not guaranteed that the boundary conditions can be ignored in other cases.

[0059] All “graphs” in this embodiment refer to adjacency matrices.

[0060] S4. Weighting the distance adjacency graph and the heat network adjacency graph to obtain a weighted directed graph.

[0061] Construct a weighted directed graph: The distance adjacency graph and the heat network adjacency graph are weighted to obtain an asymmetric adjacency matrix that covers the spatial topology information and heat transfer rules between nodes as the final weighted directed graph.

[0062] The specific construction of the weighted directed graph is as follows: based on an adjustable weight, the distance adjacency graph and the heat network adjacency graph are weighted and normalized to obtain an asymmetric adjacency matrix that covers the spatial topology information and heat transfer rules between nodes. i,j It is expressed as follows, where κ is an adjustable hyperparameter:

[0063]

[0064] S5. Mask the temperature data of some observation points in the weighted directed graph to obtain an interpolated data set.

[0065] In this example, the interpolated dataset was constructed by artificially masking the temperature data at some observation points. The target points were considered true missing locations, the masked observation points were considered artificially missing locations, and the unmasked observation points were considered non-missing locations. After preprocessing the data, the dataset was divided into training, validation, and test sets.

[0066] In the specific implementation process of this embodiment, observation points and target points are regarded as nodes and are distinguished during the processing. In the program, a new column is created to use numbers to distinguish different nodes. For example, the three types of nodes, target points, masked observation points and non-masked observation points, are marked with 0, 1 and 2 respectively.

[0067] Specifically, the interpolation data set is constructed by artificially masking some observation point data. The masking strategy includes but is not limited to random sampling of observation points or sampling guidance combined with expert knowledge. The target point is regarded as a real missing position, the masked observation point is regarded as an artificially missing position, and the non-masked observation point is regarded as a non-missing position. The missing indicator matrix M and the masked indicator matrix H are designed to mark the data at each position. Specifically, the data is pre-processed by filling the missing data of the target point with zero according to the dimension of the observation point data; dividing the time window and constructing the label set: the non-masked observation value after the time window is divided is used as the training data, the masked observation value corresponding to the last moment of each time window is used as the interpolation label, and the non-masked observation value corresponding to the last moment of each time window is used as the reconstructed label. For node i, time window interval t, the element M in the missing indicator matrix i,t and the element H in the mask indicator matrix i,t It is expressed as follows:

[0068]

[0069] S6. Use the masked observation points and non-masked observation points in the interpolation data set to train the temperature field prediction network to obtain a temperature field prediction model; the temperature field prediction network is a relational graph convolutional network; the loss function of the temperature field prediction network includes an interpolation loss function and a reconstruction loss function; the interpolation loss function is constructed using the masked observation points; the reconstruction loss function is constructed using the non-masked observation points; the masked observation points are the masked observation points in the weighted directed graph; the non-masked observation points are the observation points that are not masked in the weighted directed graph.

[0070] The training process is as follows: inputting the interpolated data set into the temperature field prediction network; using the temperature field prediction network to predict the temperature data of the masked observation point to obtain predicted data; using the predicted data and the real temperature data of the corresponding masked observation point in combination with the loss function to adjust the parameters of the relationship graph convolution to obtain a relationship graph convolution model.

[0071] In this embodiment, an interpolation loss is designed based on masked observation points, a reconstruction loss is designed based on non-masked observation points, and a weighted temperature field prediction loss function is designed based on the two.

[0072] Designing interpolation loss based on masked observation points is as follows: Constructing interpolation loss function based on masked observation points based on mean absolute error (MAE) and the reconstruction loss function based on non-masked observation points And design the weighted parameters, add the two weightedly, and get the temperature field prediction loss function and The mathematical representation is as follows, where X imp is the interpolation label, Xrec is the reconstructed label, X′ is the interpolation result, X″ is the reconstruction result, N is the total number of nodes, and λ is the loss weight hyperparameter adjustable in the range of 0-1.

[0073]

[0074] Build a temperature field prediction network. Design an encoder based on the RGCN network (Relational Graph Convolutional Network), and build a learnable weight matrix in the encoder to dynamically update the weighted directed graph. Design a decoder based on the RGCN network and flatten the decoder output.

[0075] The temperature field prediction network constructed in S8 is specifically as follows: taking the weighted adjacency graph and the divided training data set as input, after initializing the network parameters, constructing an encoder based on RGCN, the encoder is a single-layer GRU network (gated recurrent unit), and all matrix multiplications of the update gate, reset gate and candidate hidden state in the GRU are performed by the GCN operator. The calculation process of the RGCN layer is as follows:

[0076]

[0077] Among them, σ and tanh nonlinear transformation activation function, θ u ,θ r ,θ c Update gate u (K) , reset gate r (K) and candidate hidden states The learnable weight matrix of u , b r , b c are the update gate, reset gate and bias vector of the candidate hidden state respectively; x is the input vector, h is the hidden state vector, and K is the number of iterations.

[0078] A learnable weight matrix W is introduced in the RGCN layer. Its dimension is the same as the adjacency matrix and is used to dynamically update the weighted directed graph G. It is initialized to a constant value. The parameter update process of the weight matrix W is as follows:

[0079]

[0080] Where W l is the weight matrix of the lth layer in the GCN operator, is the loss function, η is the learning rate. The weighted directed graph update process is as follows, where w i,j is the element at the corresponding position in the weight matrix W:

[0081]

[0082] The encoder uses the ReLU activation function to perform a nonlinear transformation on the RGCN output. A decoder based on the RGCN is designed, with the decoder's RGCN layer and activation layer structure consistent with the encoder's. The decoder output is flattened to obtain the model output.

[0083] The results are replayed at different scales, and the interpolation results of the target points are extracted as the final temperature field prediction results.

[0084] The specific steps of rescaling the results are as follows: rescaling the results based on the pre-order normalization method to obtain the prediction results at the real scale.

[0085] In this embodiment, the PMSM public data set stator yoke temperature T is selected. sy Feature simulation target point, namely, visual T sy This is a real loss and does not participate in any training. Select the stator tooth temperature T st , stator winding temperature T sw Simulate artificial masking of observation points, that is, visual T st 、T sw is a known value, but is artificially masked and used to measure interpolation accuracy. d 2 +i q 2 ), motor speed n, coolant temperature T w1 、Ambient temperature T air1 , permanent magnet temperature T pm Simulate unmasked observations to measure data reconstruction accuracy. The motor speed n is taken to the power of 1.3, 2, 2.86, and 3 respectively to simulate several heat sources, such as frictional heat. d 2 and i q 2 They respectively represent the measured value of the current d component and the measured value of the current q component in the dq coordinate system during the measurement process.

[0086] In addition, in this embodiment, the Adam optimization algorithm is used to optimize the network weights and biases. The initial learning rate is set to 0.005 and decays at a rate of 0.5 every 10 cycles. The ratio of training set, validation set, and test set is 6:3:1. The early stopping method is used to avoid overfitting. That is, in the 20 training cycles after the current training cycle, the error on the validation set and the error on the validation set in the previous training cycle do not differ by more than 1×10 -6 , then stop training and select the training model in the current cycle as the final prediction model. In addition, set the maximum model training cycle to 200.

[0087] In actual application (i.e., prediction), the monitored temperature at the unmasked observation point is input, and the predicted temperature at the target point is output. If the device or operating conditions are changed (i.e., new operating conditions not included in the existing dataset), the model needs to be retrained using the above method.

[0088] In this embodiment, the absolute mean error and median error of the predicted value after scale playback are selected to evaluate the solution results. The evaluation indicators of the experimental results are shown in Table 1:

[0089] Table 1 Scheme result evaluation data table

[0090]

[0091] As can be seen from Table 1, the average temperature error of the interpolation prediction results of the masked observation points and target points is around 2 degrees, and the median error is around 1 degree, which meets the accuracy requirements of actual engineering. Figure 2 (a) Figure 2 (b) and Figure 2 (c) shows the comparison between the predicted values ​​and the true values ​​of this embodiment, demonstrating that the regression fit of this embodiment method is good. Furthermore, the standard deviation of the regression errors across the seven experiments is small, indicating minimal fluctuations in the results and reliable experimental results. These experimental results demonstrate that this embodiment is feasible and effective for temperature field prediction at target points on key equipment components where it is difficult to place temperature sensors.

[0092] This embodiment introduces the interpolation idea for temperature field prediction for the first time, analyzes the correlation between nodes based on physical heat transfer information and the spatiotemporal dependency characteristics of temperature characteristics, constructs a static graph of the thermal network and a spatial topological adjacency graph, and improves the interpretability and reliability of the model by introducing physical heat transfer information and spatial topological information and following physical laws. It automatically updates the connection strength between nodes based on real-time data to more accurately reflect actual changes in physical and environmental conditions and achieve high-precision expression of dependencies between nodes. A relational graph convolutional network is used to predict missing temperature data using dependencies between nodes. This method makes the model easy to implement and expand, and is particularly suitable for real-time or near real-time application scenarios. It can be widely used in health status monitoring of various equipment.

[0093] In the field of temperature field prediction of key components of equipment, this embodiment introduces the interpolation idea for temperature field prediction for the first time, avoiding the idealization of prediction results in simulation methods and saving prediction calculation and time costs. By constructing a static diagram of the thermal network to introduce physical heat transfer information, the interpretability and reliability of the model are improved; the connection strength between nodes is automatically updated according to real-time data to more accurately reflect the actual physical and environmental conditions. High-precision expression of the dependency relationship between nodes is achieved. A relational graph convolutional network is used to utilize the dependency relationship between nodes to predict missing temperature data. The method of this embodiment is novel and suitable for real-time or near real-time application scenarios, making it easy to use in engineering practice.

[0094] The above embodiment is a tuning result of the present invention on a publicly available PMSM dataset. However, the specific implementation of the present invention is not limited to the above embodiment. The description of the above embodiment is only intended to facilitate understanding of the method and core concept of the present invention. At the same time, those skilled in the art will be able to modify the specific implementation and application scope based on the concept of the present invention. In summary, the content of this specification should not be construed as limiting the present invention.

[0095] In an exemplary embodiment, a method for predicting an equipment temperature field is provided, comprising:

[0096] Step A1: Obtain the temperature data of the observation point in the current stage.

[0097] Step A2: input the temperature data of the observation point in the current stage into the temperature field prediction model to obtain the temperature data of the predicted target point; the temperature field prediction model is trained by the temperature field prediction network training method described above.

[0098] In another exemplary embodiment of the present application, Figure 3 As shown in the figure, a method for predicting the temperature field of equipment is provided. This method is a method for predicting the temperature field of equipment based on limited measured data and interpolation of spatiotemporal relationships. It introduces the interpolation idea for temperature field prediction for the first time and only uses limited actual temperature monitoring data to predict the temperature of unknown target points. It can be widely used in the health status monitoring of various equipment. The technical route diagram of this embodiment is shown in FIG. Figure 4 As shown, the following steps are included:

[0099] B1. Based on the equipment layout and operating conditions, mark the locations of the equipment temperature monitoring observation points (referred to as observation points) and the temperature field prediction target points (referred to as target points).

[0100] Specifically, the locations of equipment temperature observation points and target points in B1 are selected from key equipment components that are economically valuable but difficult to place temperature sensors on, and several evenly distributed target points are set. Observation points are selected from spatial topological locations adjacent to the target points or locations on other adjacent components suitable for placing temperature sensors.

[0101] B2. Place temperature sensors at observation points on key equipment components and record temperature monitoring time series data and related operating condition information.

[0102] The temperature sensor arrangement in B2 is specifically to select a temperature sensor that is suitable for the measurement range, response time and accuracy requirements and can withstand the equipment operating conditions (such as vibration, impact, temperature changes, etc.), such as a thermocouple, thermal resistor (RTD) or infrared temperature sensor. Record temperature-related operating parameters such as equipment working environment temperature, speed, power, coolant temperature, lubricating oil temperature, etc. The raw data obtained in this embodiment is shown in the schematic diagram. Figure 5 (a) Figure 5 (b) and Figure 5 As shown in (c).

[0103] B3. Construct a distance adjacency graph. Consider the observation point and target point as nodes in the adjacency graph. Calculate the spatial topological distance between nodes to obtain the distance adjacency graph topological weights. Calculate the data correlation between nodes to obtain the distance adjacency graph correlation weights. Add the two and normalize them to obtain the final weights for each edge in the distance adjacency graph.

[0104] Calculating the spatial topological distance between nodes in B3 specifically involves using, but not limited to, Euclidean distance and normalizing the distance to obtain a topological weight. Calculating the data correlation between nodes specifically involves using, but not limited to, the Pearson correlation coefficient to calculate the data correlation between observation points and obtain a correlation weight. Calculating and obtaining the final weight of the distance adjacency graph specifically involves adding the topological weight and the correlation weight and normalizing the result again.

[0105] B4. Construct a thermal network adjacency graph. Based on the lumped parameter thermal network method, establish the heat transfer paths between nodes.

[0106] Determining the heat transfer paths between nodes in B4 involves determining reasonable boundary conditions (such as temperature and heat flux) based on the lumped parameter thermal network method, making reasonable simplifications and assumptions about the heat transfer paths based on actual conditions, and addressing nonlinear factors in the heat transfer process as necessary. Based on the principles of thermodynamics and heat transfer, the presence of heat transfer paths between nodes is determined, the heat transfer direction is calibrated, and a 0-1 distribution heat network adjacency matrix is ​​constructed.

[0107] The schematic diagrams of the heat network adjacency graph and weighted adjacency graph in this embodiment are as follows: Figure 6 (a) and Figure 6 As shown in (b).

[0108] B5. Construct a weighted directed graph. Weight the distance adjacency graph and the heat network adjacency graph to obtain an asymmetric adjacency matrix that includes the spatial topology information and heat transfer patterns between nodes, which serves as the final weighted directed graph.

[0109] The weighted directed graph constructed in B5 is specifically constructed by weighted addition and normalization of the distance adjacency graph and the heat network adjacency graph based on an adjustable weight, thereby obtaining an asymmetric adjacency matrix that covers the spatial topology information and heat transfer rules between nodes.

[0110] B6. Construct an interpolated dataset after artificially masking the temperature monitoring data at some observation points. Treat the target points as truly missing locations, the masked observation points as artificially missing locations, and the unmasked observations as non-missing locations. After preprocessing the data, divide it into training, validation, and test sets.

[0111] The interpolation data set constructed in B6 is specifically to artificially mask part of the observation point data. The masking strategy includes but is not limited to random sampling of the observation points or sampling guidance combined with expert knowledge. The target point is regarded as the real missing position, the masked observation point is regarded as the artificial missing position, and the non-masked observation point is regarded as the non-missing position. The missing indicator matrix and the masked indicator matrix are designed to mark the data at each position. The data is preprocessed specifically to fill the missing data of the target point with zero according to the dimension of the observation point data; divide the time window and construct the label set: the non-masked observation value after the time window is divided is used as the training data, the masked observation value corresponding to the last moment of each time window is used as the interpolation label, and the non-masked observation value corresponding to the last moment of each time window is used as the reconstructed label.

[0112] B7. Design interpolation loss based on masked observation points, design reconstruction loss based on non-masked observation points, and design a weighted temperature field prediction loss function based on the two.

[0113] Specifically, the interpolation loss designed based on the masked observation points in B7 is to construct an interpolation loss function based on the masked observation points and a reconstruction loss function based on the non-masked observation points based on the mean absolute error (MAE), and design weighting parameters to add the two together to obtain the temperature field prediction loss function.

[0114] B8. Build a temperature field prediction network. Design an encoder based on the RGCN network and construct a learnable weight matrix in the encoder to dynamically update the weighted directed graph. Design a decoder based on the RGCN network and flatten the decoder output.

[0115] The temperature field prediction network constructed in B8 is specifically constructed by taking the weighted adjacency graph and the divided training data set as input, initializing the network parameters, and then constructing an RGCN-based encoder. The encoder is a single-layer GRU network, and the update gate, reset gate, and all matrix multiplications of the candidate hidden state in the GRU are replaced by GCN operators. A learnable weight matrix with the same dimension as the adjacency matrix is ​​introduced into the RGCN layer for dynamically updating the weighted directed graph. The encoder uses the ReLU activation function to perform a nonlinear transformation on the RGCN output. A decoder based on RGCN is designed, and the decoder RGCN layer and activation layer structure are consistent with the encoder structure. The decoder output is flattened to obtain the model output.

[0116] B9. Replay the results at different scales and extract the interpolation results of the target points as the final temperature field prediction results.

[0117] In B9, the result is scaled and replayed, specifically: based on the pre-order normalization method, the result is scaled and replayed to obtain the prediction result at the real scale.

[0118] This embodiment introduces the interpolation idea for temperature field prediction for the first time, analyzes the correlation between nodes based on physical heat transfer information and the spatiotemporal dependency characteristics of temperature characteristics, constructs a static graph of the thermal network and a spatial topological adjacency graph, and improves the interpretability and reliability of the model by introducing physical heat transfer information and spatial topological information and following physical laws. It automatically updates the connection strength between nodes based on real-time data to more accurately reflect actual changes in physical and environmental conditions and achieve high-precision expression of dependencies between nodes. A relational graph convolutional network is used to predict missing temperature data using dependencies between nodes. This method makes the model easy to implement and expand, and is particularly suitable for real-time or near real-time application scenarios. It can be widely used in health status monitoring of various equipment.

[0119] Compared with the existing technology, this embodiment has the following advantages:

[0120] Compared with traditional methods, the proposed approach uses only limited actual temperature monitoring data to predict the temperature of unknown target points, eliminating reliance on traditional physical field simulation results. This avoids idealized predictions and reduces computational and time costs. Furthermore, the dynamic adjacency graph-based approach is more sophisticated than traditional thermal models, enabling the model to adapt to diverse operating conditions and accurately capture and utilize the complex heat exchange mechanisms within the equipment, improving both the flexibility and accuracy of predictions. This approach outperforms traditional interpolation and regression techniques, enabling adaptive learning and updating of the dynamic interactions between temperature nodes.

[0121] In the field of temperature field prediction of key components of equipment, the present invention introduces the interpolation concept for temperature field prediction for the first time, avoiding the idealization of prediction results in simulation methods and saving prediction calculation and time costs. By constructing a static diagram of the thermal network, physical heat transfer information is introduced to improve the interpretability and reliability of the model; the connection strength between nodes is automatically updated according to real-time data to more accurately reflect the actual changes in physical and environmental conditions and achieve high-precision expression of the dependencies between nodes. A relational graph convolutional network is used to utilize the dependencies between nodes to predict missing temperature data. The method of the present invention is novel and suitable for real-time or near real-time application scenarios, making it easy to use in engineering practice.

[0122] In an exemplary embodiment, a computer device is provided. The computer device may be a server or a terminal. The internal structure diagram thereof may be as follows: Figure 7 As shown. The computer device includes a processor, a memory, an input / output interface (Input / Output, abbreviated as I / O) and a communication interface. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, the above-mentioned temperature field prediction network training method or equipment temperature field prediction method is implemented.

[0123] Those skilled in the art will understand that Figure 7 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.

[0124] In an exemplary embodiment, a computer device is further provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiments when executing the computer program.

[0125] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0126] In an exemplary embodiment, a computer program product is provided, including a computer program. When the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0127] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0128] Those skilled in the art will understand that all or part of the processes in the above-mentioned embodiment methods can be implemented by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, database or other media used in the embodiments provided in this application may include at least one of non-volatile and volatile memory. Non-volatile memory may include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory may include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0129] The databases involved in the various embodiments provided herein may include at least one of a relational database and a non-relational database. Non-relational databases may include, but are not limited to, distributed databases based on blockchains. The processors involved in the various embodiments provided herein may include, but are not limited to, general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic units, data processing logic units based on quantum computing, and the like.

[0130] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0131] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A temperature field prediction network training method, characterized in that: The temperature field prediction network training method includes: Acquire temperature data of an observation point; the observation point is a temperature monitoring point of the equipment where the sensor is installed; The observation point and the target point are used as nodes to construct a distance adjacency graph; the target point is the temperature field prediction target point; Establishing heat transfer paths between the nodes to obtain a heat network adjacency graph; Weighting the distance adjacency graph and the heat network adjacency graph to obtain a weighted directed graph; masking the temperature data of some observation points in the weighted directed graph to obtain an interpolated data set; A temperature field prediction network is trained using masked observation points and non-masked observation points in the interpolation data set to obtain a temperature field prediction model; the temperature field prediction network is a relational graph convolutional network; the loss function of the temperature field prediction network includes an interpolation loss function and a reconstruction loss function; the interpolation loss function is constructed using masked observation points; the reconstruction loss function is constructed using non-masked observation points; the masked observation points are masked observation points in the weighted directed graph; the non-masked observation points are non-masked observation points in the weighted directed graph.

2. The temperature field prediction network training method according to claim 1, characterized in that: The observation point and the target point are used as nodes to construct a distance adjacency graph, which specifically includes: Taking the observation point and the target point as nodes; Calculating the Euclidean distance between each of the nodes; Normalizing the Euclidean distance to obtain a distance adjacency graph topology weight; Calculating the Pearson correlation coefficient between each of the nodes; The distance adjacency graph topology weight and the Pearson correlation coefficient are added and normalized to obtain a weight value of each edge in the distance adjacency graph; A distance adjacency graph is constructed using the weight values ​​of the nodes and the edges.

3. The temperature field prediction network training method according to claim 1, characterized in that: The step of establishing the heat transfer paths between the nodes to obtain a heat network adjacency graph specifically includes: According to the principles of thermodynamics and heat transfer, determine whether there is a heat transfer path between each node and obtain the heat transfer determination result; calibrating the heat transfer direction in the heat transfer determination result to obtain a calibrated heat transfer result; A heat network adjacency matrix with 0-1 distribution is constructed according to the calibrated heat transfer results to obtain a heat network adjacency graph.

4. The temperature field prediction network training method according to claim 1, characterized in that: The temperature field prediction network is trained using the masked observation points and the non-masked observation points in the interpolation data set to obtain a temperature field prediction model, specifically including: Inputting the interpolation data set into a temperature field prediction network; Predicting the temperature data of the masked observation point using the temperature field prediction network to obtain predicted data; The predicted data and the actual temperature data of the corresponding masked observation point are combined with a loss function to adjust the parameters of the relationship graph convolution to obtain a temperature field prediction model.

5. The temperature field prediction network training method according to claim 1, characterized in that: The temperature field prediction network includes: an encoder based on the RGCN network and a decoder based on the RGCN; The encoder includes a weight matrix; the weight matrix is ​​used to dynamically update the weighted directed graph.

6. The temperature field prediction network training method according to claim 1, characterized in that: The process of constructing the loss function of the temperature field prediction network includes: Construct an interpolation loss function based on masked observation points based on mean absolute error; Constructing the reconstruction loss function based on non-masked observation points based on mean absolute error; The interpolation loss function and the reconstruction loss function are weighted to obtain the loss function of the temperature field prediction network.

7. A method for predicting equipment temperature field, characterized in that: The equipment temperature field prediction method includes: Get the temperature data of the observation point in the current stage; The temperature data of the observation point in the current stage is input into a temperature field prediction model to obtain the temperature data of the predicted target point; the temperature field prediction model is trained by the temperature field prediction network training method according to any one of claims 1 to 6.

8. A computer device comprising: A memory, a processor, and a computer program stored in the memory and runnable on the processor, characterized in that the processor executes the computer program to implement the temperature field prediction network training method described in any one of claims 1 to 6 or the equipment temperature field prediction method described in claim 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the temperature field prediction network training method according to any one of claims 1 to 6 or the equipment temperature field prediction method according to claim 7 is implemented.

10. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the temperature field prediction network training method according to any one of claims 1 to 6 or the equipment temperature field prediction method according to claim 7 is implemented.

Images