Motor temperature trend prediction method based on GCN-LSTM multi-sensor data fusion

Abstract

The application discloses a motor temperature trend prediction method based on GCN-LSTM multi-sensor data fusion, and specifically comprises the following steps: step 1, a plurality of multivariate time series collected by a plurality of different monitoring motor temperature state sensors are constructed into a sensor network, and a space-time graph is generated on the network; step 2, a GCN-LSTM model is constructed to obtain the final prediction value of the motor temperature. In the application, the multi-source sensor fusion affecting the motor temperature trend is abstracted into a non-directional graph structure by using GCN, the neighborhood attribute coding of the node is taken as the weight of the self-loop through the local clustering coefficient, the node can obtain more information of the neighbor nodes in the message aggregation process of the graph neural network, the unique space features of the sensor network can be effectively mined, and meanwhile, the LSTM is introduced to extract the space-time features of the motor temperature, so that the motor trend prediction accuracy is effectively improved.

Description

Technical Field

[0001] This invention relates to a method for predicting motor temperature conditions based on multi-sensor data fusion using GCN-LSTM. Background Technology

[0002] Electric motors are crucial power inputs for equipment such as belt conveyors and crushers in open-pit coal mines, and their safe and reliable operation is of great significance for ensuring a stable coal supply. Therefore, conducting research on the condition monitoring and evaluation of mining electric motors, and timely and accurately sensing their operating status, provides technical support for optimizing the operation and maintenance strategies of coal mining enterprises, thereby reducing unplanned downtime, extending equipment life, and saving costs.

[0003] Deep learning, with its powerful feature learning capabilities and accurate results, has gradually become a powerful tool in many industrial applications. Its significant advantage lies in its ability to model multi-dimensional features, learn complex relationships within data, and thus improve prediction accuracy. Although deep learning-based monitoring and prediction methods have achieved considerable success, there are still some problems to be solved: (1) Compared with single sensor signals, multi-sensor signals have more meaningful features. Multi-sensor feature fusion can achieve higher prediction accuracy in prediction models. Few researchers have considered using multi-sensor feature fusion for situation prediction, which limits the performance of the model; (2) Current deep learning methods used for situation prediction either consider temporal correlation or spatial correlation, but do not combine temporal and spatial correlation simultaneously, resulting in overfitting or low prediction accuracy in the constructed prediction model.

[0004] In recent years, Graph Neural Networks (GNNs) have developed rapidly, capable of effectively learning from non-Euclidean data and uncovering complex spatial relationships within it. GNNs have been successfully applied in many fields. Because GNNs can establish interdependencies between data and embed them into extracted features, researchers have increasingly applied them to equipment failure prediction and health management. Inspired by this, time series data collected by multiple sensors are typically regular data, but sensor networks constructed from multiple sources are considered irregular data. This invention constructs a situation prediction model based on GCN-LSTM. On one hand, it utilizes Graph Convolutional Networks (GCNs) to abstract the fusion of multiple source sensors affecting motor temperature situation into an undirected graph structure. Local clustering coefficients are used to encode the neighborhood attributes of nodes as weights for self-loops, enabling nodes to acquire more information from neighboring nodes during message aggregation in the graph neural network, thus effectively uncovering the unique spatial features of the sensor network. On the other hand, it introduces Long Short-Term Memory (LSTM) networks to extract the spatiotemporal features of motor temperature, thereby effectively improving the accuracy of motor situation prediction. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a method for predicting motor temperature conditions based on multi-sensor data fusion using GCN-LSTM.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: a method for predicting motor temperature status based on multi-sensor data fusion using GCN-LSTM, as detailed below:

[0007] Step 1. Construct a graph based on multivariate time series.

[0008] A multivariate time series data collected by multiple sensors monitoring the temperature status of the motor is constructed into a sensor network, and a space-time map is generated on this network; the specific content includes:

[0009] For a multivariate time series X=[X1, X2, …, X…] composed of k sensors… κ First, each sensor is treated as a node in a sensor network to construct a graph with k nodes; second, the signals in each sensor are normalized; finally, cosine similarity is used to estimate the distance between each sensor, and a threshold is defined. If the cosine similarity is greater than the threshold, then there will be an edge between the two nodes; node x is obtained as follows: i Neighbors:

[0010] (A)

[0011] In the formula, yes The neighboring node, y i These are the labels for the sensor network, and ε is the selected radius. Here, the threshold ε is set to 0. compute nodes and Cosine similarity;

[0012] The weights between every two nodes are calculated using a threshold Gaussian kernel weighting function.

[0013] (B)

[0014] in It is the bandwidth variance of the Gaussian function;

[0015] use An undirected graph representing a sensor network monitoring the temperature status of a motor, wherein... Represents a set of nodes. , This represents the number of sensors on the graph. Denotes the set of edges. Represents the adjacency matrix;

[0016] After obtaining the constructed sensor network, a spatial-temporal map is generated by assigning time series collected at different timestamps as node features.

[0017] Step 2. Construct a GCN-LSTM model to obtain the final predicted value of the motor temperature.

[0018] The problem of predicting motor temperature status based on sensor networks is described as follows: Predicting the motor temperature at the next time step t by collecting historical time data (s) from multiple sensors and the spatial relationships A between the sensors, i.e.:

[0019] (I)

[0020] In the formula, Let t represent the motor temperature at time t; A represents the spatial influence relationship between various sensors in the sensor network, i.e., the adjacency matrix.

[0021] F represents the processing method using the GCN-LSTM model;

[0022] The GCN-LSTM model includes at least two layers of graph convolutional neural network (GCN), at least one layer of long short-term memory network (LSTM), and at least two fully connected layers (FC).

[0023] Graph Convolutional Neural Network (GCN) consists of convolutional layers, pooling layers, and readout layers. Convolutional layers iteratively gather information from their neighbors to obtain new nodes. Pooling layers iteratively merge nodes on each edge to form new nodes, preserving the connections between merged and new nodes. The readout layers fold the node representation of the subgraph into a graph representation. Long Short-Term Memory (LSTM) network consists of an input layer, hidden layers, and an output layer. The first fully connected layer accepts the LSTM output as input, reducing the feature dimension of nodes and extracting more representative features. The second fully connected layer accepts the output of the first fully connected layer as input and outputs the final predicted value.

[0024] First, a multivariate historical time series dataset of length s is input into the model. A graph convolutional neural network (GCN) is then used to analyze the topology of the sensor network, extract spatial features, and perform mapping, as detailed below:

[0025] A graph convolutional neural network (GCN) in the spatial domain is used to capture the spatial correlations between nodes in the spatial-temporal graph obtained in step 1:

[0026] The input to graph convolution is multivariate time series data X and adjacency matrix A. The transmission method between convolutional layers is as follows:

[0027] (1)

[0028] in, This represents the sum of adjacent matrices and the identity matrix. ; for The degree matrix; This represents the feature matrix of the l-th layer; This represents the feature matrix of the (l+1)th layer; This represents the initial matrix, whose characteristic matrix is ​​the multivariate time series characteristic matrix; This represents the weight matrix of the l-th layer; For activation functions; , , Represented as:

[0029] (2)

[0030] (3)

[0031] (4)

[0032] in, This represents the connection relationship between two nodes in a sensor network;

[0033] The edge pooling method is used for pooling, which iteratively merges nodes on each edge to form a new node, while preserving the connection relationship between the merged node and the new node;

[0034] Then, the extracted time-series data with spatial features is input into the Long Short-Term Memory (LSTM) network to learn the temporal features and perform mapping, outputting the final predicted value; specifically as follows:

[0035] The Long Short-Term Memory (LSTM) neural network consists of several units, with each time step in the temperature sequence corresponding to one unit. Each unit contains an input gate, an output gate, a forget gate, and a cell state.

[0036] The input gate structure is i (t) The output gate structure is o (t) The forget gate structure is f (t) The calculation formula is as follows:

[0037] (5)

[0038] (6)

[0039] (7)

[0040] In the formula, This is an important characteristic of the t-1 time step; It is the original time feature at time step t, and it is the historical time series with spatial features after GCN processing; and These are the weights and bias terms of the input gate, respectively; and These are the weights and bias terms of the output gate, respectively; and These are the weights and biases of the forget gate, respectively;

[0041] Each unit in a Long Short-Term Memory (LSTM) network has initial temporal features based on the current time step. Key time features of the previous time step Two temporal features, using the forget gate Deleting control time features using an input gate The candidate time features are supplemented, and the cumulative impact of motor temperature on the temperature at multiple historical moments is quantified to obtain the time features of the current time step. These features are then processed by the output gate. Output the key temporal features at the current time step, with cell state c. (t) The calculation process is as follows:

[0042] ; (9);

[0043] (10);

[0044] In the formula, The time feature at time step t; The time feature to be predicted at time step t; and These are the weights and bias terms of the time features to be predicted, respectively;

[0045] The key time features of the current time step output will be input into the next unit to participate in the time feature extraction of the next time step, and so on, until the last unit completes the time feature extraction;

[0046] Long Short-Term Memory (LSTM) networks incorporate key temporal features from the last time step. Mapping this to the motor temperature at the predicted time allows for motor temperature prediction, i.e.:

[0047] (11)

[0048] In the formula, The output is the mapping; W and b are the weights and biases of the mapping, respectively.

[0049] As a preferred embodiment, the GCN-LSTM model comprises a two-layer graph convolutional neural network (GCN), a two-layer long short-term memory network (LSTM), and a two-layer fully connected layer (FC).

[0050] The beneficial effects of this invention are:

[0051] (1) The present invention abstracts the sensor network formed by multiple source sensors into an undirected graph, and fuses the data of multiple source sensors by constructing a graph, automatically extracts features and fuses the features of multiple source sensors to realize motor status prediction.

[0052] (2) The proposed GCN-LSTM model considers both the spatial correlation and temporal dependence of the input features. Specifically, GCN is used to capture the spatial dependence between multi-source sensor networks, while LSTM is used to learn temporal features from time series curves.

[0053] (3) Compared with the prediction results of traditional CNN, LSTM, GCN, etc., the proposed method has more robust performance and higher prediction accuracy. Attached Figure Description

[0054] Figure 1 This is a spatiotemporal diagram of the sensor network constructed by multi-source sensing in this invention;

[0055] Figure 2 This is the GCN-LSTM model structure of the present invention;

[0056] Figure 3 This invention relates to the pooling method;

[0057] Figure 4 This is a visualization of the nodes and edges of the GCN-LSTM model of this invention;

[0058] Figure 5 This is a comparison of evaluation parameters for different baseline models in this invention;

[0059] Figure 6 This is a comparison of losses between different models of this invention;

[0060] Figure 7 These are the motor temperature prediction results for the next 5 minutes using different models of this invention. Detailed Implementation

[0061] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0062] This invention takes a drive motor of a belt conveyor in an open-pit coal mine as the research object. Pearson correlation analysis revealed a strong correlation between motor temperature and bearing temperature, three-phase winding temperature, and ambient temperature. Therefore, second-level data were collected from February 24, 2021 to March 6, 2021.

[0063] Before data analysis, the collected multivariate time series signals should be cleaned, deduplicated, aligned, downsampled, and have planned and unplanned downtime removed. A sliding window is used to label the multivariate time series data; this involves dividing the data into fixed-size windows, using the data within each window as input, and the next or subsequent data points within the window as labels. Furthermore, to facilitate data use and eliminate the negative impact of outlier samples on the overall data, the labeled data undergoes Min-max normalization. The normalization formula is as follows: ;

[0064] In the formula, This is the raw temperature data; The normalized temperature data; is the maximum value in the original temperature data. This is the minimum value in the original temperature data.

[0065] A method for predicting motor temperature status based on multi-sensor data fusion using GCN-LSTM is described below:

[0066] Step 1: Constructing a graph based on multivariate time series

[0067] For a multivariate time series X=[X1, X2, X3, X4, X5] consisting of 5 sensors, firstly, each sensor is treated as a node in a sensor network to construct a graph with 5 nodes; secondly, the signals from each sensor are normalized; finally, cosine similarity is used to estimate the distance between each sensor, and a threshold is defined. If the cosine similarity is greater than a threshold, then there will be an edge between the two nodes. Therefore, node x can be obtained as follows: i Neighbors:

[0068] (A)

[0069] In the formula, yes The neighboring node, y i These are the labels for the sensor network, and ε is the selected radius. Here, the threshold ε is set to 0. compute nodes and Cosine similarity;

[0070] The weights between every two nodes are calculated using a threshold Gaussian kernel weighting function.

[0071] (B)

[0072] in It is the bandwidth variance of the Gaussian function;

[0073] To obtain sufficient training samples, the window length is set to 30. Six days of data are used as the training set, one day as the test set, and one day as the validation set.

[0074] use An undirected graph representing a sensor network monitoring the temperature status of a motor, wherein... Represents a set of nodes. , This represents the number of sensors on the graph. Denotes the set of edges. Represents the adjacency matrix. For example... Figure 1 As shown, a spatial-temporal map can be generated by assigning node features with different timestamps to the sensor network.

[0075] Step 2: Construct the GCN-LSTM model to obtain the final predicted value of the motor temperature.

[0076] The problem of predicting motor temperature status based on sensor networks can be described as follows: predicting the motor temperature at the next time t by collecting historical time data of duration s from multiple sensors and the spatial relationship A between the sensors, i.e.:

[0077]

[0078] In the formula, Let t represent the motor temperature at time t; A represents the spatial influence relationship between various sensors in the sensor network, i.e., the adjacency matrix; F is the processing method, which is the GCN-LSTM model in this embodiment.

[0079] like Figure 2 As shown, the proposed GCN-LSTM model has 6 layers, namely two GCN layers, two LSTM layers, and two fully connected layers (FC).

[0080] Graph Convolutional Neural Network (GCN) consists of convolutional layers, pooling layers, and readout layers. Convolutional layers iteratively gather information from their neighbors to obtain new nodes. Pooling layers iteratively merge nodes on each edge to form new nodes, preserving the connections between merged and new nodes. The readout layers fold the node representation of the subgraph into a graph representation. Long Short-Term Memory (LSTM) network consists of an input layer, hidden layers, and an output layer. The first fully connected layer accepts the LSTM output as input, reducing the feature dimension of nodes and extracting more representative features. The second fully connected layer accepts the output of the first fully connected layer as input and outputs the final predicted value.

[0081] First, a multivariate historical time series dataset of length s is input into the model. A graph convolutional neural network (GCN) is then used to analyze the topology of the sensor network, extract spatial features, and perform mapping, as detailed below:

[0082] A graph convolutional neural network (GCN) in the spatial domain is used to capture the spatial correlations between nodes in the spatial-temporal graph obtained in step 1:

[0083] The input to graph convolution is multivariate time series data X and adjacency matrix A. The transmission method between convolutional layers is as follows:

[0084] (1)

[0085] in, This represents the sum of adjacent matrices and the identity matrix. ; for The degree matrix; This represents the feature matrix of the l-th layer; This represents the feature matrix of the (l+1)th layer; This represents the initial matrix, whose characteristic matrix is ​​the multivariate time series characteristic matrix; This represents the weight matrix of the l-th layer; This is the activation function. , , Represented as:

[0086] (2)

[0087] (3)

[0088] (4)

[0089] in, This indicates the connection relationship between two nodes in a sensor network.

[0090] With the help of a graph pooling layer, this invention employs edge pooling, which iteratively merges nodes on each edge to form a new node, while preserving the connection between the merged node and the new node. This reduces the number of nodes in the graph, thereby lowering computational complexity. The specific process is as follows: Figure 3 As shown.

[0091] Then, the extracted time-series data with spatial features is input into a Long Short-Term Memory (LSTM) network to learn the temporal features and perform mapping, outputting the final predicted value. LSTM is a special recurrent neural network structure used to extract the temporal features of motor temperature. It extracts the temporal features of the actual temperature at the predicted time from the actual temperature sequence and maps them to the actual temperature at the predicted time, thus achieving the mapping from the actual temperature sequence to the actual temperature at the predicted time. Since the motor temperature is affected by the cumulative effect of temperatures at multiple historical times, this cumulative effect should be considered when extracting the temporal features. Specifically:

[0092] The Long Short-Term Memory (LSTM) neural network consists of several units, with each time step in the temperature sequence corresponding to one unit. Each unit contains an input gate, an output gate, a forget gate, and a cell state.

[0093] The input gate structure is i (t) The output gate structure is o (t) The forget gate structure is f (t) The calculation formula is as follows:

[0094] (5)

[0095] (6)

[0096] (7)

[0097] In the formula, This is an important characteristic of the t-1 time step; It is the original time feature at time step t, and it is the historical time series with spatial features after GCN processing; and These are the weights and bias terms of the input gate, respectively; and These are the weights and bias terms of the output gate, respectively; and These are the weights and biases of the forget gate, respectively.

[0098] Each unit in the LSTM network has initial time features based on the current time step. Key time features of the previous time step Two temporal features, using the forget gate Deleting control time features using an input gate The candidate time features are supplemented to quantify the cumulative impact of motor temperature on the temperature at multiple historical moments, thereby obtaining the time features of the current time step, which are then processed by the output gate. Output the key temporal features at the current time step, with cell state c. (t) The calculation process is as follows:

[0099] ; (9)

[0100] (10)

[0101] In the formula, The time feature at time step t; The time feature to be predicted at time step t; and These are the weights and bias terms of the time features to be predicted, respectively;

[0102] The key time features of the current time step are output and then input into the next unit to participate in the time feature extraction of the next time step. This process is repeated until the last unit completes the time feature extraction.

[0103] LSTM neural networks capture key temporal features at the last time step. Mapping this to the motor temperature at the predicted time allows for motor temperature prediction, i.e.: (11)

[0104] In the formula, The output is the mapping; W and b are the weights and biases of the mapping, respectively.

[0105] The method employed effectively extracts graph features. Node information is embedded in each layer, and features from the original data are extracted by graph convolutional layers. The number of nodes and edges in the graph is reduced through graph pooling. After node embedding, each node represents a vector, and the embedded node is represented by an embedding vector. Each layer of the network is visualized using Gephi, such as... Figure 4 As shown in the figure. This model can reduce the number of nodes in the graph, thereby reducing computational complexity.

[0106] Other parameter settings for the GCN-LSTM model are as follows:

[0107] The batch size was 64, the learning rate was 0.001, the number of training iterations was 150, and the detailed structure of the model is shown in Table 1.

[0108] The first layer of GCN 1 has a batch size of 64, a number of nodes of 320, a feature dimension of 1024, an activation function of ReLU, and 31744 parameters.

[0109] The second layer is GCN 2, with a batch size of 64, 192 nodes, a feature dimension of 1024, a ReLU activation function, and 1,049,600 parameters.

[0110] The third LSTM layer is 1, with a batch size of 64, 64 nodes, a feature dimension of 2048, a ReLU activation function, and 25182202 parameters.

[0111] The fourth layer is LSTM 2, with a batch size of 64, 64 nodes, a feature dimension of 2048, a ReLU activation function, and 33,570,816 parameters.

[0112] The fifth layer is FC 1, with a batch size of 64, a number of nodes of 64, a feature dimension of 2048, and a number of parameters of 4196352;

[0113] The sixth layer is FC 2, with a batch size of 64, a number of nodes of 64, a feature dimension of 5, and a number of parameters of 10245.

[0114] Table 1 Detailed structure of the model

[0115] 1 GCN 1 (64, 320, 1024) ReLU 31744 2 GCN 2 (64, 192, 1024) ReLU 1049600 3 LSTM 1 (64, 64, 2048) ReLU 25182202 4 LSTM 2 (64, 64, 2048) ReLU 33570816 5 FC 1 (64, 64, 2048) —— 4196352 6 FC 2 (64, 64, 5) —— 10245

[0116] Evaluation indicators

[0117] To better quantitatively evaluate the model's fitting and prediction performance, this paper uses three widely used metrics to assess the model's performance: root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The expression is:

[0118] (12)

[0119] (13)

[0120] (14)

[0121] In the formula, N is the sample size. Let the value be the true value of the i-th sample. Let represent the predicted value of the i-th sample. RMSE indicates the model's prediction accuracy, while MAE reflects the actual error in the prediction value. The smaller the RMSE and MAE values, the better the prediction model's performance. R 2 The range of the model's predicted output is the degree to which the input variables are controllable. The value is between 0 and 1. The closer the value is to 1, the better the model fits the data.

[0122] Results and Analysis

[0123] To compare and analyze the prediction performance of the prediction model, the proposed motor temperature trend prediction model is compared with existing prediction methods. The baseline models include the following three types: (1) Convolutional Neural Network (CNN); (2) Graph Convolutional Neural Network (GCN); and (3) Long Short-Term Memory Network (LSTM). The proposed model and the baseline model predict the motor temperature trend for the next 5 minutes on the above dataset. The comparison of various evaluation indicators is as follows: Figure 5 As shown in the figure, the GCN-LSTM algorithm has lower RMSE and MAE values ​​than other algorithms, and R... 2 The value is higher than other algorithms. The prediction performance based on deep learning is generally better than machine learning methods, indicating that deep learning has advantages in predicting motor temperature conditions.

[0124] Using the baseline model mean as a benchmark, the GCN-LSTM algorithm reduced RMSE and MAE values ​​by 31.3% and 38.7%, respectively. 2The value improved by 17.1%, as shown in Table 2. Because the baseline model is a time series prediction model, it only considers the temporal correlation between data points and not spatial correlation. GCN-LSTM fully integrates the advantages of various neural network methods, demonstrating significant advantages in all performance metrics. The good prediction performance of GCN-LSTM is attributed to its ability to effectively capture complex local spatiotemporal correlations through a carefully designed spatiotemporal synchronization modeling mechanism. Furthermore, the multiple time-segmented modules designed in the model also help to effectively capture different characteristics in the local spatiotemporal graph. The good prediction performance of GCN-LSTM also indicates the existence of different local spatiotemporal correlations between motor temperatures. This also verifies that introducing a graph structure in motor temperature trend prediction helps to uncover complex spatiotemporal relationships in temperature data and improve prediction accuracy.

[0125] Table 2 Comparison of evaluation parameters for different models

[0126] RMSE 1.7947 2.3101 1.9441 2.0163 1.386 MAE 1.496 1.9654 1.1524 1.6579 1.017 <![CDATA[R 2 ]]> 0.8095 0.6844 0.7765 0.7568 0.8864

[0127] Furthermore, we compared the training efficiency of the GCN-LSTM model with other baseline models. Figure 6 shows the relationship between the loss curve and training epochs. The loss of GCN-LSTM is the smallest among all models. Compared with the three existing methods, the GCN-LSTM model can achieve convergence quickly with fewer iterations.

[0128] Figure 7 The results show the prediction of motor temperature for the next 5 minutes using a GCN-LSTM model and three benchmark models, respectively, with a historical time window of 30 minutes. The predictions are then compared with actual temperature data. As shown in the figure, the temperature values ​​predicted by the GCN-LSTM model are similar to the original values ​​in most cases, even when there are sudden changes. Furthermore, the LSTM model's estimates differ the most from the measured data, failing to accurately adapt to sudden changes. It must be emphasized that the proposed GCN-LSTM model not only captures normal behavior better but also captures a series of atypical dynamics. This demonstrates the model's ability to predict extreme temperature behavior in time series data.

[0129] The above embodiments are merely illustrative of the principles and effects of the present invention, as well as some examples of its application, and are not intended to limit the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements are all within the scope of protection of the present invention.

Claims

1. A method for predicting motor temperature status based on multi-sensor data fusion using GCN-LSTM, as detailed below: Step 1. Construct a graph based on multivariate time series. A multivariate time series collected by multiple sensors that monitor the temperature status of the motor is constructed into a sensor network, and a space-time map is generated on this network; The specific content includes: For a multivariate time series X=[X1, X2, …, X…] composed of k sensors… κ First, each sensor is treated as a node in a sensor network to construct a graph with k nodes. Second, the signals from each sensor are normalized. Finally, cosine similarity is used to estimate the distance between each sensor, and a threshold ε is defined. If the cosine similarity is greater than the threshold, there will be an edge between the two nodes. Node x is obtained as follows: i Neighbors: （A） In the formula, yes The neighboring node, y i These are the labels for the sensor network, and ε is the selected radius. Here, the threshold ε is set to 0. compute nodes and Cosine similarity; The weights between every two nodes are calculated using a threshold Gaussian kernel weighting function. （B） in It is the bandwidth variance of the Gaussian function; Let G = {V, E, A} represent an undirected graph of a sensor network for monitoring the temperature status of a motor, where V represents the set of nodes. , Let E represent the number of sensors on the graph, E represent the set of edges, and A represent the adjacency matrix. After obtaining the constructed sensor network, a spatial-temporal map is generated by assigning time series collected at different timestamps as node features; Step 2. Construct a GCN-LSTM model to obtain the final predicted value of the motor temperature. The problem of predicting motor temperature status based on sensor networks is described as follows: Predicting the motor temperature at the next time step t by collecting historical time data (s) from multiple sensors and the spatial relationships A between the sensors, i.e.: （I） In the formula, X t Let t represent the motor temperature at time t; A represents the spatial influence relationship between various sensors in the sensor network, i.e., the adjacency matrix. F represents the processing method using the GCN-LSTM model; The GCN-LSTM model includes at least two layers of graph convolutional neural network (GCN), at least one layer of long short-term memory network (LSTM), and at least two fully connected layers (FC). Graph Convolutional Neural Network (GCN) consists of convolutional layers, pooling layers, and readout layers. Convolutional layers iteratively gather information from their neighbors to obtain new nodes. Pooling layers iteratively merge nodes on each edge to form new nodes, preserving the connections between merged and new nodes. The readout layers fold the node representation of the subgraph into a graph representation. Long Short-Term Memory (LSTM) network consists of an input layer, hidden layers, and an output layer. The first fully connected layer accepts the LSTM output as input, reducing the feature dimension of nodes and extracting more representative features. The second fully connected layer accepts the output of the first fully connected layer as input and outputs the final predicted value. First, a multivariate historical time series dataset of length s is input into the model. A graph convolutional neural network (GCN) is then used to analyze the topology of the sensor network, extract spatial features, and perform mapping, as detailed below: A graph convolutional neural network (GCN) in the spatial domain is used to capture the spatial correlations between nodes in the spatial-temporal graph obtained in step 1: The input to graph convolution is multivariate time series data X and adjacency matrix A. The transmission method between convolutional layers is as follows: （1） in, This represents the sum of adjacent matrices and the identity matrix. ; for degree matrix; H (l) H represents the feature matrix of the l-th layer; (l+1) H represents the feature matrix of the (l+1)th layer; (o) W represents the initial matrix, whose characteristic matrix is ​​the multivariate time series characteristic matrix; (l) This represents the weight matrix of the l-th layer; For activation functions; , , Represented as: （2） （3） （4） Among them, A ij This represents the connection relationship between two nodes in a sensor network; The edge pooling method is used for pooling, which iteratively merges nodes on each edge to form a new node, while preserving the connection relationship between the merged node and the new node; Then, the extracted time-series data with spatial features is input into the Long Short-Term Memory (LSTM) network to learn the temporal features and perform mapping, outputting the final predicted value; specifically as follows: The Long Short-Term Memory (LSTM) neural network consists of several units, with each time step in the temperature sequence corresponding to one unit. Each unit contains an input gate, an output gate, a forget gate, and a cell state. The input gate structure is i (t) The output gate structure is o (t) The forget gate structure is f (t) The calculation formula is as follows: （5） （6） （7） In the formula, This is an important characteristic of the t-1 time step; It is the original temporal feature at time step t, and the historical time series with spatial features after GCN processing; W ig and b ig These are the weights and bias terms of the input gate, respectively; W og and b og These are the weights and bias terms of the output gate, respectively; W fg and b fg These are the weights and biases of the forget gate, respectively; Each unit in a Long Short-Term Memory (LSTM) network has initial temporal features based on the current time step. Key time features of the previous time step Two temporal features, using the forgetting gate f (t) The control time feature is deleted using the input gate i. (t) The candidate time features are supplemented, and the cumulative impact of motor temperature on the temperature at multiple historical moments is quantified to obtain the time features of the current time step. These features are then processed by the output gate. (t) Output the key temporal features at the current time step, with cell state c. (t) The calculation process is as follows: （9） （10） In the formula, The time feature at time step t; The time feature to be predicted at time step t; and These are the weights and bias terms of the time features to be predicted, respectively; The key time features of the current time step output will be input into the next unit to participate in the time feature extraction of the next time step, and so on, until the last unit completes the time feature extraction; Long Short-Term Memory (LSTM) networks incorporate key temporal features from the last time step. Mapping this to the motor temperature at the predicted time allows for motor temperature prediction, i.e.: （11） In the formula, Y is the mapping output; W and b are the mapping weights and bias terms, respectively.

2. The method for predicting motor temperature status based on GCN-LSTM multi-sensor data fusion as described in claim 1, characterized in that: The GCN-LSTM model consists of a two-layer graph convolutional neural network (GCN), a two-layer long short-term memory network (LSTM), and a two-layer fully connected layer (FC).

Images