A device residual life prediction method based on hypergraph neural network

Abstract

The application discloses a kind of equipment residual life prediction methods based on hypergraph neural network.First, the data set is obtained according to the equipment type to be predicted, then the hypergraph structure is dynamically constructed, the high-order interaction relationship between nodes is captured, and the node embedding representation with spatial structure information is learned using the hypergraph neural network layer;The node embedding representation with spatial structure information obtained is processed by the time information capture layer, and finally the equipment residual life prediction result is obtained by the full connection layer output.The present application takes into account the high-order interaction relationship between sensor nodes, effectively utilizes the spatial structure information between sensor nodes, and uses an efficient MLP model in the time information capture layer, which has obvious advantages over traditional models in predicting the residual life of equipment.

Description

Technical Field

[0001] This invention belongs to the field of mechanical equipment failure prediction technology, specifically relating to a method for predicting the remaining life of equipment based on a hypergraph neural network and a spatiotemporal fusion mechanism. Background Technology

[0002] With the development of the Industrial Internet of Things (IIoT), predictive health management (PHM) for industrial system maintenance has attracted significant attention from researchers. Accurate prediction of remaining useful life (RUL) is crucial in PHM. Thanks to the rapid advancements in modern sensor technology, condition monitoring data can now be easily obtained from operating equipment. Condition-based maintenance (CBM) primarily revolves around developing maintenance plans based on the health status of monitored components. By analyzing sensor degradation signals, CBM can effectively identify early indicators of potential component failures. Therefore, it reduces maintenance costs and minimizes downtime in industrial systems caused by maintenance operations or malfunctions, thereby improving overall productivity.

[0003] Typically, two main approaches are used in researching Remaining Useful Life (RUL) prediction: model-based approaches and data-driven approaches. Model-based techniques involve dynamically modeling the degradation process of equipment using accurate physical models. However, due to the complex structure of modern large industrial equipment, characterized by intricate nonlinear interactions between components, building a single physical failure model presents significant challenges. Data-driven approaches for RUL prediction involve extracting feature information from monitoring data signals, thus avoiding the expertise required to build complex models that conform to the physical characteristics of the equipment. Some traditional machine learning techniques, such as Support Vector Regression (SVR) and Random Forest (RF), have been applied to RUL prediction tasks. However, with the surge in data volume, the effectiveness of traditional machine learning methods decreases, prompting the adoption of deep learning (DL)-based methods. DL-based methods excel at handling large amounts of data, automatically extracting relevant information from degradation signals, and performing well in RUL prediction tasks.

[0004] Deep neural network models effectively capture the temporal and spatial information present in degraded signals. Convolutional Neural Networks (CNNs) are renowned for their proficiency in extracting multidimensional spatial information and are widely used in RUL prediction tasks. Another popular model is the Recurrent Neural Network (RNN), which excels at extracting valuable information from time-series data—a capability that CNNs lack. However, RNN models have limitations in effectively learning long-term dependencies in sequential data. The Transformer architecture, initially developed for natural language processing tasks, has gained attention due to its ability to effectively handle sequential data. Numerous studies have demonstrated the Transformer's advantages in handling long sequences, utilizing self-attention mechanisms to comprehensively model the interactions between data elements. Some research has already applied the Transformer to RUL prediction tasks. However, directly applying the Transformer model ignores the inherent structured information in sensor data.

[0005] To address these challenges, recent research has proposed a spatiotemporal fusion strategy. This strategy utilizes graph neural networks (GNNs) to capture structured information between sensor nodes. Subsequently, a model capable of processing time-series data is used to effectively capture temporal information. However, most existing models rely on graph neural networks (GNNs), which typically capture simple pairwise interactions in graph structures and struggle to handle complex high-order interactions between multiple sensors. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method for predicting the remaining lifespan of equipment based on a hypergraph neural network. The aim is to improve the accuracy of predicting the remaining lifespan of equipment, minimize downtime of industrial systems caused by maintenance operations or malfunctions, and thereby improve overall productivity.

[0007] A method for predicting the remaining lifetime of equipment based on a hypergraph neural network includes the following steps:

[0008] S1. Select sensors to collect relevant data according to the type of device to be predicted, and perform data preprocessing operations. The preprocessing includes: removing useless data, data clustering, and data normalization. After preprocessing, perform sliding window processing and divide the training set and test set.

[0009] S2. Dynamically construct hypergraph structures and capture high-order interaction relationships between nodes;

[0010] S3. Based on the dynamically constructed hypergraph structure, node embedding representations with spatial structure information are learned by hypergraph neural network layers.

[0011] S4. The node embedding representation with spatial structure information obtained by the time information capture layer is processed, and finally the device remaining lifetime prediction result is output through the fully connected layer.

[0012] S5. The equipment remaining life prediction model is trained using the training set constructed in S1.

[0013] The device remaining life prediction model consists of a hypergraph neural network layer, a time information capture layer, and a fully connected layer.

[0014] S6. After the model training is completed, the remaining lifespan will be predicted and the model will be evaluated based on the input serialized data.

[0015] Furthermore, step S1 is specifically performed as follows:

[0016] Select sensors based on the type of equipment to be predicted, and construct a custom dataset using the data collected by the sensors in real time. The requirements for the custom dataset are: it needs to collect data from a variety of different sensors. The more types of sensors collected, the higher the accuracy of the final remaining life prediction. The sensor collection is related to the type of equipment. For example, rotating machinery (such as motors, pumps, fans, etc.) requires vibration sensors, temperature sensors, acceleration sensors, noise sensors, etc.; internal combustion engines and gas turbines require pressure sensors, temperature sensors, emission sensors, vibration sensors.

[0017] Since some sensor data may remain unchanged throughout the degradation process and will not affect the final prediction results, the first step is to delete these irrelevant sensor data (i.e., sensor data that remains unchanged during degradation). Then, considering the unit differences in the data returned by different sensors, data normalization is necessary. For multi-mode data, clustering methods are used followed by normalization, with the K-nearest neighbor algorithm employed for clustering. For single-mode data, direct normalization is performed. The operating mode refers to the characteristics and behavioral patterns exhibited by the equipment under specific conditions and operating states. In actual production environments, for equipment operating in multiple states—that is, equipment that may be in different characteristic and behavioral modes at different times (e.g., no load, partial load, full load, etc.)—clustering methods followed by normalization are required; otherwise, direct normalization is performed.

[0018] The formula for normalization is as follows:

[0019]

[0020] Where S represents the set of sensors, O represents the set of operating modes, and s i，j Represents the original input, μ i，j σ represents the mean of the input data. i，jz represents the variance of the input data. i，j This indicates the score.

[0021] Furthermore, the sliding window processing in step S1 is as follows:

[0022] The raw data is processed using a sliding window. The training data is divided into smaller multidimensional time series using a fixed window length with a step size of one, and the corresponding remaining lifetime is used as the prediction label. The length of the sliding window is a customizable hyperparameter. For each data point processed by the sliding window, there is a remaining lifetime label value (RUL). The device's RUL follows a piecewise function, initially remaining constant (125). However, the device's RUL gradually decreases over time, decreasing by 1 unit per second. Therefore, the time point at which the RUL begins to decrease can be calculated from the initial RUL value. The piecewise function is as follows: Figure 1 The horizontal axis represents time points, and the vertical axis represents RUL;

[0023] Furthermore, in step S2, the process of dynamically constructing the hypergraph structure is as follows:

[0024] Hypergraph uses Let V represent this, where V t Denotes the set of nodes, E t Let H represent the set of hyperedges, and H... t The adjacency matrix of the hypergraph is represented by the subscript 't', indicating time 't'. Since the number of sensor nodes remains constant, but the relationships between sensors differ at different times, the hypergraph structure should be dynamically changing. The data obtained through sliding window processing is obtained by sliding a fixed-size window through the time-series data, extracting data segments. Each window corresponds to a time period and a RUL label value, reflecting the device's RUL value within that time period. In actual processing, the position of the last time point in the sliding window is used to represent the entire time period, and the RUL value is based on the RUL value of the last time point in that window. The hypergraph structure differs at different time points; therefore, each data obtained through sliding window processing has a corresponding hypergraph structure. The change in the hypergraph structure is mainly reflected in the changes in the number of nodes contained in different hyperedges, while the overall number of nodes remains constant. Nodes represent sensors in the feature space. The node information contained in a hyperedge at a certain time can be represented by e. t ={v1, v2, ..., v nThe hyperedge is represented by}, meaning it contains n distinct nodes, and the K-nearest neighbor algorithm is used to obtain the hyperedge. Furthermore, to represent the information contained in all hyperedges, the adjacency matrix H of the hypergraph is needed. t In H t Each element h(v, e) on the array is represented by the following formula:

[0025]

[0026] Adjacency matrix H t It is a matrix composed of 0s and 1s, where a value of 1 indicates that the node is contained on the hyperedge, and a value of 0 indicates that the node does not belong to the hyperedge. Therefore, the degree of a node is defined as:

[0027]

[0028] The degree of a hyperedge is defined as:

[0029]

[0030] The final degree of all nodes will be represented by an n x n diagonal matrix D. v (This represents the number of nodes, where n is the number of nodes, and the diagonal elements are d(v), which is the degree of node v; the degree of the hyperedge will be expressed using an m x m diagonal matrix D.) e Let m be the number of hyperedges and d(e) be the element on the diagonal, which is the degree of hyperedge e.

[0031] Furthermore, in step S3, the hypergraph learning process is as follows:

[0032] Each data point captured by the sliding window corresponds to a hypergraph. Information propagation between nodes in the hypergraph occurs through shared hyperedges. This message passing requires the use of a hypergraph convolution mechanism. The formula for hypergraph convolution is as follows:

[0033] X (l+1) =σ(H t WH t T X (l) P)

[0034] Where σ() is the activation function, H t Let H be the adjacency matrix of the hypergraph, and H be the adjacency matrix of the hypergraph. t T This is the transpose of the adjacency matrix. Let represent the feature matrix of the l-th layer sensor node, that is, the feature matrix of the sensor node after all nodes have completed l rounds of message passing. Let F represent the feature matrix of the (l+1)th layer sensor node. Since the sensor nodes have the same size across different layers, F... (l) =F(l+1) W is a diagonal matrix of hyperedge weights, where the diagonal elements represent the weights of the corresponding hyperedges. Initially, it is assumed that all hyperedges are equally important, i.e., all hyperedge weights are 1. This represents the learnable parameter matrix. The final hypergraph convolution process will undergo symmetric normalization, and the final formula after symmetric normalization is as follows:

[0035] X (l+1) =σ(D v -1 / 2 H t D e -1 H t T D v -1 / 2 X (l) P)

[0036] The activation function σ() can be either ReLU or LeakyReLU;

[0037] After completing three message passes between sensor nodes, different sensor nodes will contain information about their third-order neighbor nodes, resulting in a matrix composed of node embedding representations with spatial structure information. The embedding representation of each node is a vector containing the structure and feature information of the node and its neighbor nodes. These vectors can be stored in a matrix, where each row corresponds to the embedding representation of a node, and the columns represent the various dimensions of the vector. The data is in the form of a matrix, with rows representing different sensor nodes and columns representing the various dimensions of the node embedding vector.

[0038] Furthermore, in step S4, the time information capture layer process is as follows:

[0039] The temporal information capture layer uses N sequentially connected multilayer perceptron (MLP) modules to learn temporal information from features. Data with node embedding representations containing spatial structure information is input into the MLP modules for learning. The MLP module contains two distinct sub-modules: the MLP Time Module and the MLP Feature Module. The MLP Time Module is used to capture temporal information, while the MLP Feature Module is used to capture feature domain information.

[0040] The MLP Time Module is a 2-layer MLP structure consisting of a fully connected (FC) layer, a ReLU layer, a Dropout layer, an FC layer, and a Dropout layer. Considering the influence of independent variables on the experiment, another MLP module is needed to enhance the model. The MLP Feature Module is a 3-layer MLP structure consisting of a fully connected (FC) layer, a ReLU layer, a Dropout layer, an FC layer, a Dropout layer, an FC layer, and a Dropout layer to capture information in the feature domain. To improve model performance, layer normalization is added before each MLP sub-module. Furthermore, residual connections are used between layer normalization and the MLP sub-modules. The use of residual connections between layer normalization and the MLP module significantly improves the training efficiency and performance of the neural network, making the model more suitable for complex tasks and datasets.

[0041] Furthermore, the fully connected layer in step S4 adopts a single-layer MLP structure, consisting of a fully connected (FC) layer, a ReLU layer, a Dropout layer, and an FC layer. It aims to learn the mapping relationship between sequence data and RUL prediction values ​​by learning the intermediate hidden layers, and output the device remaining lifetime prediction results through the fully connected layer.

[0042] Furthermore, the specific training method for the S5 model is as follows:

[0043] The device remaining lifetime prediction model, consisting of a hypergraph neural network layer and a time information capture layer, was trained using the training set constructed through S1. Furthermore, considering that predicting RUL is a classic regression task, mean squared error (MSE) was used as the final loss function.

[0044] The formula is as follows:

[0045]

[0046] Where y i This represents the actual remaining lifespan value, while This represents the predicted remaining lifetime value, and N represents the number of samples in the training set.

[0047] The backpropagation algorithm is used to adjust the overall model training parameters, so that the loss value calculated by the loss function continuously decreases, and the model training ends when the loss value eventually stabilizes.

[0048] All model parameters are jointly trained using the batch gradient descent algorithm, with an early stopping mechanism to prevent overfitting. The batch gradient descent algorithm relies on backpropagation to calculate gradients, making it one of the core optimization algorithms for neural network training. Together, they work in the neural network training process to achieve the goals of parameter optimization and loss function minimization.

[0049] Furthermore, the specific method for step S6 is as follows:

[0050] The test set data is input into the trained model to predict remaining lifetime and evaluate model performance;

[0051] Two evaluation metrics are used to assess the model's performance. The first metric is the root mean square error (RMSE), which is widely used for prediction tasks. The specific calculation formula is as follows:

[0052]

[0053] in, y represents the predicted remaining lifespan. i The first metric is the true remaining lifetime value, and N represents the number of samples in the training set. The second metric is the scoring function (SF). The core of the SF metric is that it more severely penalizes predictions that are greater than the true value. This is achieved by changing the base, and the specific calculation formula is as follows:

[0054]

[0055] The beneficial effects of this invention are as follows:

[0056] This invention is a device remaining lifetime prediction method based on hypergraph neural networks. Compared with existing technologies, it takes into account the high-order interaction relationships between sensor nodes and effectively utilizes the spatial structure information between sensor nodes. At the same time, it adopts an efficient MLP model in the time information capture layer, which has significant advantages over traditional models in predicting device remaining lifetime. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of a piecewise function;

[0058] Figure 2 This is a flowchart of an embodiment of the present invention;

[0059] Figure 3 This is a diagram illustrating the overall model training process in an embodiment of the present invention.

[0060] Figure 4 This is a schematic diagram of the structure and principle of the hypergraph neural network layer in an embodiment of the present invention;

[0061] Figure 5 This is a schematic diagram of the structure of the time information capture layer in an embodiment of the present invention. Detailed Implementation

[0062] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0063] In this embodiment, the C-MAPSS dataset is used. The C-MAPSS dataset was generated by NASA using simulation software. This dataset contains four sub-datasets: FD001, FD002, FD003, and FD004. Each sub-dataset has different operating and failure modes.

[0064] Table 1. Description of the composition of the C-MAPSS dataset

[0065] Training set engine 100 260 100 249 Test set engine 100 259 100 248 Work mode 1 6 1 6 Failure Mode 1 1 2 2

[0066] Each dataset includes a training set and a test set, where data from sensors records the entire degradation process of each turbofan engine from normal operation to failure. Statistics from C-MAPSS are shown in Table 1.

[0067] like Figure 2 As shown, this invention is a flowchart of a method for predicting the remaining life of equipment based on a hypergraph neural network. The specific steps are as follows:

[0068] Step 1: Perform data preprocessing based on the data input from the sensor;

[0069] Data preprocessing includes: removing useless data, data clustering, data normalization, label construction, and splitting the training and test sets. Specific operations:

[0070] (1) Removal of useless data: There are 21 sensors used to monitor the operating status of the turbofan engine. Observation of the raw data revealed that 7 of the sensors remained unchanged during the engine degradation process. These 7 sensors do not provide information for the final prediction, so this data should be removed.

[0071] (2) Data Clustering: According to the original dataset description, the sensor data collected in different working modes are independent. Therefore, for the subsets with multiple working modes, it is necessary to cluster the data according to their respective working modes. The K-nearest neighbor algorithm is used for clustering, which clusters different data based on data characteristics.

[0072] (3) Data Normalization: Based on a fundamental assumption that data in the same subset of data under the same working mode are independent and identically distributed, normalization is mainly performed using the mean and variance of the data. The formula is as follows:

[0073]

[0074] Where S represents the set of registers, O represents the set of operating modes, and s i，j Represents the original input, μ i，j σ represents the mean of the input data. i，j z represents the variance of the input data.i，j This indicates the score.

[0075] (4) Sliding window processing, remaining lifetime label setting, and splitting the training and test sets: The original data is processed using a sliding window. A fixed window length with a step size of one is used to divide the training data into smaller multidimensional time series, and the corresponding remaining lifetime is used as the prediction label. The length of the sliding window is a customizable hyperparameter. For the data obtained through sliding window processing, each data point has a remaining lifetime label (RUL). The device's RUL conforms to a piecewise function, remaining constant initially (corresponding to the initial RUL obtained through sliding window processing remaining constant), but gradually decreasing over time (corresponding to the RUL obtained through sliding window processing gradually decreasing over time), with a decrease rate of 1 for every time step. A test set is then split from the original training set, with a ratio of 80% to 20%.

[0076] Step 2: Dynamically construct the hypergraph structure and capture the high-order interaction relationships between nodes;

[0077] (1) Construct a dynamic hypergraph structure to capture high-order interactions between nodes: Each sensor node should be a node in the hypergraph, and the dimension of the node is the cut length of the sliding window. K-nearest neighbor clustering is used to obtain hyperedges, where the number of points contained in the hyperedge is the K-nearest neighbor clustering coefficient. Duplicate hyperedges need to be removed in this step. Once the hypergraph structure is constructed, high-order interactions exist between different nodes, which will be captured through message passing between nodes.

[0078] Step 3: Obtain node embedding representations with spatial structure information through hypergraph neural network learning;

[0079] The model training process is as follows Figure 3 It mainly consists of a hypergraph neural network layer responsible for capturing spatial structure information and a temporal information capture layer responsible for capturing temporal structure information. Details of the hypergraph neural network layers are as follows... Figure 4 The specific implementation process is as follows:

[0080] Information propagation between nodes occurs through shared hyperedges. This message passing requires the use of a hypergraph convolution mechanism. The formula for hypergraph convolution is as follows:

[0081] X (l+1) =σ(H t WH t T X (l) P)

[0082] Where H t Let H be the adjacency matrix of the hypergraph, and H be the adjacency matrix of the hypergraph. tT This is the transpose of the adjacency matrix. Let represent the feature matrix of the l-th layer sensor node, that is, the feature matrix of the sensor node after all nodes have completed l rounds of message passing. Let F represent the feature matrix of the (l+1)th layer sensor node. Since the sensor nodes have the same size across different layers, F... (l) =F (l+1) W is a diagonal matrix of the weights of the hyperedges, where the diagonal elements represent the weights of the corresponding hyperedges. For simplicity, we assume that all hyperedges are equally important at the initial moment, i.e., all hyperedge weights are 1. This represents the learnable parameter matrix. During training, the equation may increase the risk of gradient explosion / vanishing due to numerical instability, therefore it will be normalized. The final formula after symmetric normalization is as follows:

[0083] X (l+1) =σ(D v -1 / 2 H t D e -1 H t T D v -1 / 2 X (l) P)

[0084] The activation function σ() can be either ReLU or LeakyReLU;

[0085] Step 4: The obtained node embedding representation with spatial structure information is processed by the time information capture layer, and finally processed by the fully connected layer to obtain the device remaining lifetime prediction result. The time information capture module is as follows: Figure 5 As shown. Specific process:

[0086] (1) The obtained node embedding representation with spatial structure information is processed through the time information capture layer:

[0087] The temporal information in the learned features utilizes an MLP module, primarily by inputting data containing spatial structure information into the MLP module for learning. The temporal information capture layer uses N sequentially connected MLP modules to learn the temporal information in the features, inputting data with node embeddings containing spatial structure information into the MLP module for learning, where N is a subset of the total number of nodes. Figure 5 The MLP module contains two different sub-modules: the MLP Time Module and the MLP Feature Module. The MLP Time Module is used to capture time information, while the MLP Feature Module is used to capture feature domain information.

[0088] The MLP Time Module is a 2-layer MLP structure consisting of a fully connected (FC) layer, a ReLU layer, a Dropout layer, an FC layer, and a Dropout layer. Considering the influence of independent variables on the experiment, another MLP module is needed to enhance the model. The MLP Feature Module is a 3-layer MLP structure consisting of a fully connected (FC) layer, a ReLU layer, a Dropout layer, an FC layer, a Dropout layer, an FC layer, and a Dropout layer to capture information in the feature domain. To improve model performance, layer normalization is added before each MLP module. Finally, residual connections are used after layer normalization and the MLP module. The use of residual connections between layer normalization and the MLP module significantly improves the training efficiency and performance of the neural network, making the model more suitable for complex tasks and datasets.

[0089] (2) Processed through a fully connected layer:

[0090] The final fully connected layer is a single-layer MLP model. The structure of the fully connected layer includes: FC layer, ReLU layer, Dropout layer, and FC layer. It aims to learn the mapping relationship between sequence data and RUL prediction values ​​by learning the intermediate hidden layers, and output the device remaining lifetime prediction results through the fully connected layer.

[0091] Step 5: Train the equipment remaining life prediction model using the training set constructed in Step 1.

[0092] (1) Model training:

[0093] The model training uses mean squared error (MSE) as the final loss function.

[0094] The formula is as follows:

[0095]

[0096] Where y i This represents the actual remaining lifespan value, while This represents the predicted remaining lifetime value, and N represents the number of samples in the training set.

[0097] After the fully connected layer, the model will use the backpropagation algorithm to adjust the training parameters, so that the loss value calculated by the loss function will continuously decrease, and the training of the model will end when the loss value tends to stabilize.

[0098] All models are built using the PyTorch framework, all model parameters are jointly trained using the batch gradient descent algorithm, and an early stopping mechanism is used to prevent overfitting.

[0099] The hyperparameters in the experiment were set as follows: each hyperedge in the hypergraph contained 4 nodes, the sliding window length was 30 / 45 / 30 / 40, the dropout rate was 0.2, the batch size was 128, the learning rate was 0.001, the hidden layer dimension was 256, and the number of MLP modules N was 2.

[0100] Step 6: After the model training is completed, the remaining lifetime will be predicted and the model will be evaluated after data preprocessing based on the input serialized data.

[0101] In a real-world production environment, only existing data needs to be used. The training error on this data is then trained until the loss value stabilizes. Subsequent data collection requires preprocessing before being input into the trained model to obtain a predicted remaining lifespan. When the remaining lifespan is shorter than a set threshold, an alarm system is triggered, alerting the user to perform timely equipment maintenance. In the experimental dataset, only the test set data needs to be input into the model to obtain the predicted remaining lifespan. This predicted lifespan is then compared with the actual remaining lifespan, and finally, evaluation metrics are used to assess the model's performance.

[0102] Two evaluation metrics are used to assess the model's performance. The first metric is the root mean square error (RMSE), which is widely used for prediction tasks. The specific calculation formula is as follows:

[0103]

[0104] in, y represents the predicted remaining lifespan. i The first metric is the true remaining lifetime value, and N represents the number of samples in the training set. The second metric is the scoring function (SF). The core of the SF metric is that it more severely penalizes predictions that are greater than the true value. This is achieved by changing the base, and the specific calculation formula is as follows:

[0105]

[0106] The final training results of the model and its comparison with other models are shown in Table 2:

[0107]

[0108] Compared to newer methods, the main advantage of this invention is that it significantly outperforms other models on datasets with multiple working modes, and outperforms most models on single-working-condition datasets.

[0109] The above description, in conjunction with specific / preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. Those skilled in the art can make various substitutions or modifications to these described embodiments without departing from the inventive concept, and all such substitutions or modifications should be considered within the scope of protection of the present invention.

[0110] The parts of this invention not described in detail are well-known to those skilled in the art.

Claims

1. A method for predicting the remaining life of equipment based on a hypergraph neural network, characterized in that, Includes the following steps: S1. Select sensors to collect relevant data according to the type of device to be predicted, and perform data preprocessing operations. The preprocessing includes: removing useless data, data clustering, and data normalization. After preprocessing, perform sliding window processing and divide the training set and test set. S2. Dynamically construct hypergraph structures and capture high-order interaction relationships between nodes; Hypergraph uses To indicate, among which Represents a set of nodes. Let represent the set of superedges, and Represents the adjacency matrix of a hypergraph, with subscripts... Indicates the first At any given moment, since the number of sensor nodes remains constant, while the relationships between sensors differ at different times, the hypergraph structure should be dynamically changing. The data obtained through sliding window processing is segmented by sliding a fixed-size window through the time-series data. Each window corresponds to a time period and a RUL label value. The position of the last time point within the sliding window represents the entire time period, and the RUL value is based on the RUL value of the last time point within that window. The hypergraph structure differs at different time points; therefore, each piece of data obtained through sliding window processing has a corresponding hypergraph structure. The changes in the hypergraph structure are mainly reflected in the changes in the number of nodes contained in different hyperedges, while the overall number of nodes remains constant. Nodes represent sensors in the feature space. The node information contained in a hyperedge at a given moment can be used... This means that the hyperedge contains n distinct nodes, and the K-nearest neighbor algorithm is used to obtain the hyperedge. Furthermore, to represent the information contained in all hyperedges, the adjacency matrix of the hypergraph is needed. ;exist Each element on This can be expressed using the following formula: Adjacency matrix It is a matrix composed of 0s and 1s, where a value of 1 indicates that the node is contained on the hyperedge, and a value of 0 indicates that the node does not belong to the hyperedge; therefore, the degree of a node is defined as: The degree of a hyperedge is defined as: The final degree of all nodes will be represented by an n x n diagonal matrix. This represents the expression, where n is the number of nodes, and the elements on its diagonal are... , i.e., node The degree of the hyperedge will be determined using an m x m diagonal matrix; Let m be the number of hyperedges, and let m be the number of elements on its diagonal. That is, superedge The degree; S3. Based on the dynamically constructed hypergraph structure, node embedding representations with spatial structure information are learned by hypergraph neural network layers. S4. The node embedding representation with spatial structure information is processed by the time information capture layer, and finally the device remaining lifetime prediction result is output through the fully connected layer. S5. Train the equipment remaining life prediction model using the training set constructed in S1. The equipment remaining life prediction model consists of a hypergraph neural network layer, a time information capture layer, and a fully connected layer; S6. After the model training is completed, the remaining lifespan will be predicted and the model will be evaluated based on the input serialized data.

2. The method for predicting the remaining life of equipment based on a hypergraph neural network according to claim 1, characterized in that, The specific operation of step S1 is as follows: Select sensors based on the type of device to be predicted, and build a custom dataset using the data collected by the sensors in real time. The requirements for the custom dataset are: it needs to collect data from a variety of different sensors. The more types of sensors collected, the higher the accuracy of the final remaining lifetime prediction. Irrelevant sensor data, i.e., sensor data that remains unchanged during degradation, is removed. The data is then normalized. For multi-mode data, clustering is used followed by normalization, with the K-nearest neighbor algorithm employed for clustering. For single-mode data, direct normalization is performed. A working mode refers to the characteristics and behavioral patterns exhibited by the equipment under specific conditions and operating states. In actual production environments, for equipment operating in multiple states—that is, equipment that may exhibit different characteristics and behavioral patterns at different times—clustering followed by normalization is required; otherwise, direct normalization is performed. The formula for normalization is as follows: in Represents a collection of sensors. Represents a set of working modes. Indicates the original input. This represents the mean of the input data. This represents the variance of the input data. This indicates the score.

3. The method for predicting the remaining life of equipment based on a hypergraph neural network according to claim 2, characterized in that, The sliding window processing in step S1 is as follows: The raw data is processed using a sliding window; the training data is divided into smaller multidimensional time series using a fixed window length with a step size of one, and the corresponding remaining lifetime is used as the prediction label; the length of the sliding window is a custom hyperparameter; for the data processed by the sliding window, each data point has a remaining lifetime label value RUL, and the device's RUL conforms to a piecewise function.

4. The method for predicting the remaining life of equipment based on a hypergraph neural network according to claim 3, characterized in that, In step S3, the hypergraph learning process is as follows: Each data point captured by the sliding window corresponds to a hypergraph. Information propagation between nodes in the hypergraph occurs through shared hyperedges. This message passing requires the use of a hypergraph convolution mechanism; the formula for hypergraph convolution is as follows: in, For activation function, Let be the adjacency matrix of the hypergraph, and This is the transpose of the adjacency matrix. Indicates the first The feature matrix of the layer sensor nodes, which completes the message transmission of all nodes. The feature matrix of the subsequent sensor nodes, and Indicates the first The feature matrix of the sensor nodes in each layer; since the sensor nodes have the same size in different layers, therefore ; It is a diagonal matrix of the weights of the hyperedges, that is, the diagonal elements of the matrix represent the weights of the corresponding hyperedges. At the initial time, it is assumed that all hyperedges are equally important, that is, all hyperedge weights are 1. This represents the learnable parameter matrix; the final hypergraph convolution process will undergo symmetric normalization, and the final formula after symmetric normalization is as follows: The activation function Use ReLU or LeakyReLU; After completing three message passes between sensor nodes, each sensor node will contain information about its third-order neighbor nodes, resulting in a matrix composed of node embedding representations with spatial structure information.

5. The method for predicting the remaining life of equipment based on a hypergraph neural network according to claim 4, characterized in that, In step S4, the time information capture layer process is as follows: The temporal information capture layer will use N sequentially connected multilayer perceptron (MLP) modules to learn temporal information in the features. Data with node embedding representations containing spatial structure information will be input into the MLP modules for learning. The MLP modules contain two different sub-modules: the MLP Time Module and the MLP Feature Module. The MLP Time Module is used to capture temporal information, while the MLP Feature Module is used to capture feature domain information. The MLP Time Module is a 2-layer MLP structure consisting of a fully connected FC layer, a ReLU layer, a Dropout layer, an FC layer, and a Dropout layer. Considering the influence of independent variables on the experiment, another MLP module is needed to enhance the model. The MLP Feature Module is a 3-layer MLP structure consisting of an FC layer, a ReLU layer, a Dropout layer, an FC layer, a Dropout layer, an FC layer, and a Dropout layer to capture information in the feature domain. To improve model performance, layer normalization is added before each MLP sub-module. In addition, residual connections are used from before layer normalization to after the MLP sub-module.

6. A method for predicting the remaining life of equipment based on a hypergraph neural network according to any one of claims 1-5, characterized in that, The fully connected layer in step S4 adopts a single-layer MLP structure, consisting of a fully connected FC layer, a ReLU layer, a Dropout layer, and an FC layer. It aims to learn the mapping relationship between sequence data and RUL prediction values ​​by learning the intermediate hidden layers, and output the device remaining lifetime prediction results through the fully connected layer.

7. The method for predicting the remaining life of equipment based on a hypergraph neural network according to claim 6, characterized in that, The specific method for training the S5 model is as follows: The training set constructed using S1 was used to train the device remaining life prediction model, which consists of a hypergraph neural network layer and a time information capture layer, and the mean squared error (MSE) was used as the final loss function. The formula is as follows: in This represents the actual remaining lifespan value, while This represents the predicted remaining lifespan value. Indicates the number of samples in the training set; The backpropagation algorithm is used to adjust the overall model training parameters so that the loss value calculated by the loss function continuously decreases, and the model training ends when the loss value eventually stabilizes. All model parameters are trained jointly using the batch gradient descent algorithm, and an early stopping mechanism is used to prevent overfitting.

8. The method for predicting the remaining life of equipment based on a hypergraph neural network according to claim 7, characterized in that, The specific method for step S6 is as follows: The test set data is input into the trained model to predict remaining lifetime and evaluate model performance; Two evaluation metrics are used to assess the model's performance; the first metric is the root mean square error (RMSE), calculated using the following formula: in, This represents the predicted remaining lifespan value. This represents the actual remaining lifespan value. The first metric represents the number of samples in the training set; the second metric is the scoring function SF, which is calculated using the following formula: 。

Images