Air compressor fault prediction and health degree evaluation method and system based on graph neural network

Abstract

The application discloses a kind of air compressor fault prediction and health degree evaluation method and system based on graph neural network, the convolution processing of different receptive field is carried out to multivariate time series to obtain multiscale embedding feature;Dynamic time series graph is established based on multi-head attention mechanism;Then graph convolution is used to discretize dynamic time series graph to obtain aggregated representation with instantaneous interval, and ordinary differential equation is used to solve it to obtain multiscale information representation;According to the obtained multiscale information representation, the multivariate time series is predicted;According to the predicted sequence, the health index of air compressor is calculated according to the health degree model, and then the fault prediction and health degree evaluation of air compressor are carried out.The application considers the dependency relationship between variables in the whole evolution process by establishing dynamic time series graph, captures the long-range dynamic relationship between variables in time series, predicts the detection variable value at future time, detects the fault of air compressor system, and makes health degree evaluation on air compressor.

Description

Technical Field

[0001] This invention belongs to the field of joint research technology of Neural Networks (NN) in Machine Learning and air compressor fault diagnosis, and relates to air compressor fault prediction based on neural networks, particularly to a method and system for air compressor fault prediction and health assessment based on graph neural networks. This invention mainly utilizes deep learning and graph neural network models to construct a learning model. This model combines Ordinary Differential Equations (ODEs) and Graph Neural Networks (GNNs) techniques, enabling it to capture the long-range internal relationships and interrelationships of time-series embedded sequences in time-series predictions. It can effectively utilize multi-scale representations to form continuous evolutionary dynamics and establish a health model to perform fault prediction and health assessment on the predicted data. In the task of fault detection of air compressors in medium- and low-pressure gas systems, this method can learn the rich relationships and patterns shared between different characteristic scales of air compressors, enhancing the robustness and accuracy of the prediction results. Background Technology

[0002] In current air compressor systems, most equipment is designed only for its rated range without considering the system's normal operating data range, and lacks detection and alarm functions. This necessitates heavy reliance on on-site inspections and data collection by operation and maintenance personnel to analyze related problems. Current system operation monitoring still relies on traditional manual management methods, with significant room for improvement in intelligent detection. Furthermore, current system detection indicators and evaluation methods are too simplistic or fragmented, with outdated evaluation methods that cannot be dynamically adjusted based on operational conditions. Evaluation results are overly dependent on the skill level of the relevant personnel.

[0003] Because multivariate time series data are ubiquitous in life and have a significant impact on many aspects of daily life, multivariate time series forecasting plays a crucial role in numerous modern applications such as climate analysis, transportation and urban mobility, the power industry, and financial markets. Much research has been conducted to improve the performance of multivariate time series forecasting. Air compressor systems also exhibit a large number of multivariate time series characteristics, such as: exhaust pressure and temperature in the air circuit system; lubricating oil temperature and pressure in the oil circuit system; and cooling water temperature and pressure in the water circuit system. Successfully predicting the fault attributes of air compressors based on these multivariate characteristics is crucial for the health assessment of air compressor systems.

[0004] In existing literature, multivariate time series forecasting is based on autoregressive models, which have the ability to capture sequence features. For example, models based on ARIMA, Recurrent Neural Networks (RNNs), and Gaussian processes have been used for various real-world time series data. However, these models either ignore global information and fail to explicitly consider the relationships between different variables, or lack interpretability. Recent advances in deep learning, especially the rapid development of GNNs, have spurred research into modeling multivariate time series by leveraging the ability of GNNs to learn neighborhood context.

[0005] Despite the encouraging results achieved by existing GNN-based work, several issues limit their performance: (1) Previous work typically utilizes shallow layers of GNN models, with most achieving optimal performance using two-layer networks. This is because deep stacking of layers leads to a significant performance degradation (also known as oversmoothing). However, in a two-layer design, each node in the graph only aggregates two-hop neighbors, ignoring long-distance associations with other nodes, which often contain expressive features beneficial for prediction. Some work attempts to address this issue by introducing an infinite number of neural units, but this makes it extremely difficult for the model to converge while learning graph features. (2) Due to the above limitations, most GNN-based methods only consider static graphs, employing time-invariant structures between variables and building graphs without dynamic relationships at different time steps. However, such designs ignore the interaction between local information and global evolution. (3) Real-world time series often exhibit various patterns in multi-scale observations. Existing methods for extracting multi-scale representations often simply fuse them through concatenation or linear transformations. Generally, this operation is difficult to provide interpretability and learn information-rich cross-scale interactions.

[0006] This invention overcomes the oversmoothing problem by incorporating neural network constant differential equations into graph neural networks, capturing long-distance semantic relationships. This enables the model to simultaneously learn global co-evolutions and interdependencies. By adjusting the continuous evolution trajectory based on the theory of control differential equations, it achieves successful prediction of air compressor system fault attributes based on the multivariate characteristic properties of the air compressor. Summary of the Invention

[0007] The purpose of this invention is to address the aforementioned problems in the prior art by providing a method and system for air compressor fault prediction and health assessment based on graph neural networks. This method uses a graph neural network model and ordinary differential equations (ODEs) to model time series in multi-scale representations. It leverages the potential connections between the graph structure and the differential equations to assist in the inference and prediction of time evolution. Simultaneously, it simulates the dynamics of multivariate time series in continuous space, enabling the prediction of multivariate time series in the air compressor system. Furthermore, by analyzing the predicted attributes, it achieves the assessment of the air compressor system's health based on fault attributes.

[0008] The idea behind this invention is to construct a novel Continuous Graph Multivariate Forecasting (CGMF) framework that combines graph neural networks and neural ordinary differential equations for multivariate time series forecasting in air compressor systems, and to combine the predicted variables for fault detection. This framework provides a direct method to simulate the dynamics of multivariate time series in continuous space. Specifically, (1) the multivariate time series monitored in the air compressor system is transformed into multi-scale embeddings for subsequent inference; (2) a dynamic multivariate time series graph structure of each scale embedding is established, and ODE is incorporated into the GNN, which enables the model to learn the global common evolution and interdependencies simultaneously; (3) multi-scale information is fully utilized to generate time series forecast results; (4) a health model is established to evaluate the predicted data and determine the failure rate of the predicted data to complete the health assessment of the air compressor system.

[0009] Based on the above inventive ideas, this invention proposes a method for air compressor fault prediction and health assessment based on graph neural networks, which includes the following steps:

[0010] S1 performs convolution processing on multivariate time series with different receptive fields to obtain multi-scale embedding features;

[0011] S2 establishes a dynamic temporal graph based on a multi-head attention mechanism; it uses graph convolution to discretize the dynamic temporal graph to obtain an aggregated representation with instantaneous intervals, and uses ordinary differential equations to solve it to obtain a multi-scale information representation.

[0012] S3 uses the obtained multi-scale information representation to predict multivariate time series;

[0013] Based on the predicted sequence, S4 calculates the health index of the air compressor according to the following health model, and then performs fault prediction and health assessment of the air compressor:

[0014]

[0015]

[0016] Among them, W n This represents the weight of the health status of the given nth observation variable, where n = 1, 2, ..., N; N represents the number of observation variables. The air compressor health index represents the nth attribute variable at the predicted τ time points. and These represent the maximum and minimum values ​​of the nth observed variable in the input time series, i.e., the maximum and minimum values ​​of observed variable n when the air compressor is working normally; The standard value of the nth variable in the input time series is represented by the average value of the nth observed variable in the input time series.

[0017] In step S1 above, multivariate time series are modeled at multiple scales. By inputting the multivariate time series, multi-scale embedding features are obtained. Given a multivariate time series... Where T represents the length of the multivariate time series, N represents the N observed variables at time t (the observed variables of the air compressor include exhaust pressure, exhaust temperature, lubricating oil temperature, lubricating oil pressure, cooling water temperature, etc.), and M represents the observation dimension of each variable. The goal of this step is to generate a series of embedded features. For further processing. Each Where S = {s1, s2, ..., s} k} represents the set of all scales. d represents a time step at scale s (specifically, dividing a time series of length T into T / s data segments at scale s), where d represents the embedding dimension. This indicates rounding down. In terms of implementation, this invention employs multiple two-dimensional convolutional units to achieve this transformation. Specifically, it utilizes convolutional units with different receptive fields (i.e., different convolutional kernels) to generate the following multi-scale embedding representation, with the two-dimensional convolutional representation as follows:

[0018]

[0019] The input and output channels of this 2D convolution are M and d, respectively, corresponding to the sizes of the input features and the output embedded representations. The kernel size is 1×s, meaning that the time series of each variable is embedded at a scale s. With this setup, multi-scale embedded representations can ultimately be generated. And used for subsequent reasoning.

[0020] In step S2 above, the multi-scale embedding representation H obtained in S1 is first used, and the adjacency matrix of the dynamic temporal graph is calculated through a multi-head attention mechanism algorithm to construct the dynamic graph structure. In addition, this invention also improves the transition smoothing problem in traditional GNN by introducing neural ODE to learn the global common evolution and interdependencies at the same time, and overcomes the transition smoothing problem.

[0021] First, a dynamic time series graph is constructed using multi-scale embedding features. This time series graph consists of T×N nodes, representing variables in all time series with a time step of T, and the adjacency matrix... Record the relationships between them. Then, use a GNN to learn the relationships between nodes and their neighbors. A classic GNN updates the representation of each node by iteratively aggregating information about itself and its neighbors. Given a dynamic temporal graph, a discrete information propagation layer on scale s can be described as follows:

[0022]

[0023] in, The information embedding representation of the l-th aggregation, Let I denote the adjacency matrix with self-loop operations, and let I denote the identity matrix. express The degree matrix is ​​f(·), which represents a feedforward network typically composed of dense layers and activation functions, and L is the number of discrete propagation layers. However, a large number of discrete propagation layers l→∞ will lead to the oversmoothing problem. To address this issue, we incorporate neural ODEs into CGMFs, causing them to aggregate almost infinitely to capture long-range interiors and interrelationships, thus overcoming the oversmoothing problem.

[0024] We construct a dynamic time series graph and connect variables using a multi-head attention mechanism, as this focuses on the most relevant and important parts of the variables in the time series. We provide the multi-scale temporal embeddings for each variable. First, they are reshaped into two-dimensional matrices, and the resulting multi-scale time embedding H... 2d It is expressed as follows:

[0025]

[0026] Then, through a linear transformation, they are mapped to a higher-order query matrix and key matrix, resulting in C parallel headers, represented as follows:

[0027]

[0028] Among them, all The learning parameter, d aLet represent the dimensions of the query and key matrices; then, dot multiplication is applied to calculate the attention weights between each pair of embeddings, using the following formula:

[0029]

[0030] in, It is the adjacency matrix of the dynamic time series graph. Finally, sort the matrix elements by size, keep the first p weights, and set the other weights to 0, because a few dot product weights contribute to the main attention, while the influence of other weights is negligible.

[0031] Therefore, this limitation can be overcome by performing a large number of aggregations over several discrete propagation steps. Specifically, let l′ = L / K represent the intermediate variable between K and L, and denote the aggregation interval, where l′ = 1 in a typical discrete GNN. K is the aggregation number, representing the information within K-hop (i.e., range K) neighbors that can be considered. Based on the above setting, an aggregation with an interval l′ at scale s can be expressed as follows:

[0032]

[0033] We use To simplify, we can now set L as a very small constant. As K→∞, l′→0. We obtain a continuous information propagation process with instantaneous intervals. When l=0, according to equation (1.6) above, we obtain an aggregated representation with instantaneous intervals.

[0034] The multi-scale information representation is obtained by solving the ordinary differential equation (neural ODE solver). for:

[0035]

[0036] in, G represents the embedding at each scale. s (·) can be any input The feedforward network. Here we let No additional learning parameters are required. This not only improves the computational efficiency of integral approximation but also aggregates linear features across all nodes. After infinite aggregation, It includes long-distance relationships between nodes in a graph structure and avoids over-smoothing.

[0037] Subsequently, patterns can be further extracted from the multivariate sequences, and nonlinearity can be introduced. The multi-scale information representation of the nodes is then updated through biomimetic transformations and activation functions, as follows:

[0038]

[0039] Where σ(·) represents the PReLU activation function, b represents the learned parameters.

[0040] Then, the multi-scale information representation of all nodes at scale s.

[0041] In step S3 above, prediction is performed on the multivariate time series. This is done to obtain the multi-scale information representation in formula (1.8). Then, a basic multi-layer perceptron (MLP) is used as a predictor to generate multivariate time series predictions for τ time steps.

[0042]

[0043] Here, θ represents the learnable parameters. For model optimization, general optimizers such as Adam and RMSprop can be used to calculate gradients and update all learned parameters. MSE loss is used as the optimization objective because the prediction task can be viewed as a linear regression task.

[0044] In step S4 above, after obtaining the multivariate time series predicted values ​​from S3, a health model is established based on the predicted data to perform fault prediction and health assessment. The health index is a concept proposed to describe the operating status of power equipment. Its basic idea is to reflect the health status of the evaluated equipment in numerical form, based on the comprehensive mathematical transformation of each evaluated parameter and the calculation of this quantitative index. The specific calculation method for the single-variable health index is as follows:

[0045]

[0046] in, It is the air compressor health index of the nth variable at the predicted τ time points; and These represent the maximum and minimum values ​​of the nth observed variable in the input time series, i.e., the maximum and minimum values ​​of observed variable n when the air compressor is working normally; This represents the standard value of the nth variable in the input time series.

[0047] The operational health index of an air compressor is defined as follows:

[0048]

[0049] Among them, W n This represents the weight of the given nth variable in terms of health, where n = 1, 2, ..., N.

[0050] Calculated HV T+τ The results range from 0 to 1. Research has revealed the following relationship between the air compressor's operational health index and its failure rate: when the operational health index is between 0.7 and 1, the equipment is considered healthy with a very low failure rate; when the operational health index is between 0.6 and 0.7 (excluding the extreme values ​​of 0.6 and 0.7), the equipment is considered sub-healthy with a low failure rate; and when the operational health index is between 0 and 0.6, the equipment is considered faulty with a high failure rate. Therefore, we can obtain the failure information and health status of the air compressor system based on the predicted data.

[0051] Thus, this invention realizes a multivariate time series prediction method based on graph neural networks and combined with neural ODEs. This method overcomes the oversmoothing problem, captures long-distance semantic associations, and can simultaneously learn global co-evolutions and interdependencies. It successfully predicts multivariate time series monitoring data in air compressor systems and achieves fault diagnosis and health assessment by constructing a health model. It should be noted that this invention is not limited to this problem but can be widely applied to real-world time series scenarios, particularly time series prediction problems involving multivariate data types.

[0052] Based on the above inventive concept, this invention proposes an air compressor fault prediction and health assessment system based on graph neural networks, which includes:

[0053] A multivariate time series forecasting model, which includes:

[0054] The multi-scale embedding feature acquisition module is used to perform convolution processing on multivariate time series with different receptive fields to obtain multi-scale embedding features;

[0055] The multi-scale information representation acquisition module is used to build dynamic temporal graphs based on multi-head attention mechanisms; the dynamic temporal graphs are discretized using graph convolution to obtain aggregated representations with instantaneous intervals, and the multi-scale information representations are obtained by solving ordinary differential equations.

[0056] The multivariate time series prediction module is used to predict multivariate time series based on the obtained multi-scale information representation.

[0057] An air compressor failure prediction and health assessment model is used to calculate the air compressor's health index based on the prediction sequence and the following health model, thereby performing failure prediction and health assessment of the air compressor:

[0058]

[0059]

[0060] Among them, W nThis represents the weight of the health status of the given nth observation variable, where n = 1, 2, ..., N; N represents the number of observation variables. The air compressor health index represents the nth attribute variable at the predicted τ time points. and These represent the maximum and minimum values ​​of the nth observed variable in the input time series, respectively. The standard value of the nth variable in the input time series is represented by the average value of the nth observed variable in the input time series.

[0061] The aforementioned multi-scale embedding feature acquisition module performs operations according to step S1; it includes:

[0062] Multivariate time series partitioning unit, used to divide the input multivariate sequence into several time periods according to different scales;

[0063] Convolutional units are used to perform convolution processing on the time series of each divided time period to generate multi-scale embedded representations.

[0064] For different scales, the convolutional unit has a different receptive field (i.e., kernel size). For scale s, the kernel size of the convolutional unit is 1×s.

[0065] The aforementioned multi-scale information representation acquisition module performs operations according to step S2 above; it includes:

[0066] The dynamic sequence graph creation unit is used to create dynamic sequence graphs based on a multi-head attention mechanism.

[0067] The graph convolution processing unit is used to discretize the dynamic time series graph to obtain an aggregated representation with instantaneous intervals;

[0068] The ODE solver is used to solve aggregated representations with instantaneous intervals to obtain multi-scale information representations.

[0069] The above dynamic timing diagram establishment unit obtains the adjacency matrix A of the dynamic timing diagram according to the above formula (1.5).

[0070] The graph convolution processing unit described above obtains an aggregated representation with instantaneous intervals according to formula (1.6).

[0071] The ODE solver described above obtains the multi-scale information feature representation by solving according to the above formula (1.7).

[0072] The aforementioned multi-scale information representation acquisition module also includes an update unit, which updates the multi-scale information features through biomimetic transformation and activation functions. Specifically, it updates the multi-scale information feature representation obtained by the ODE solver according to formula (1.8).

[0073] The multivariate time series prediction module described above uses a multi-layer perceptron (MLP).

[0074] The aforementioned multivariate time series prediction model also includes an optimizer for optimizing the model parameters. The optimizers used in this invention, such as Adam and RMSprop, are used to calculate gradients and update all learned parameters; and the MSE loss is used as the optimization objective.

[0075] Compared with existing technologies, the air compressor fault prediction and health assessment method and system based on graph neural networks provided by this invention have the following beneficial effects:

[0076] 1. This invention proposes a method for learning continuous multivariate time series forecasting (CGMF) based on graph neural networks and combined with neural ordinary differential equations. It considers the dependencies between variables throughout the entire evolution process by establishing a dynamic time series graph, capturing the long-range dynamic relationships between variables in the time series. By utilizing multivariate monitoring data from the air compressor system, including exhaust pressure, exhaust temperature, lubricating oil temperature, lubricating oil pressure, and cooling water temperature, it predicts the values ​​of the detected variables at future moments. Through analysis and judgment of the predicted data, it detects faults in the air compressor system and makes a health assessment.

[0077] 2. This invention models multivariate time series at multiple scales, explores rich patterns at each scale, and constructs dynamic time series graphs to dynamically extract cross-scale knowledge and generate multivariate predictions.

[0078] 3. Compared with traditional manual experience-based judgment methods, this invention realizes the intelligence of the detection system and the accuracy of fault prediction, which has a guiding role in the intelligent assessment of the normal operation and health of the air compressor system.

[0079] 4. Based on this invention, the method of this invention was verified on a time series dataset of multivariate air compressor attributes collected in the real world; the experimental results show that the prediction method CGMF of this invention has superior performance compared with the existing state-of-the-art models. Attached Figure Description

[0080] Figure 1 This is a schematic diagram of the process for the air compressor fault prediction and health assessment method based on graph neural networks of the present invention.

[0081] Terminology Explanation

[0082] RNN is an abbreviation for Recurrent Neural Network. It is a type of recurrent neural network that takes sequential (such as time series) data as input, recursively processes data in the direction of sequence evolution, and connects all nodes in a chain-like manner. It can capture the connections between different parts of a sequence of data and plays an important role in tasks such as natural language processing and time series prediction. Its theoretical basis can be found in the reference [A. Sagheer and M. Kotb. Unsupervised Pre-Training of ADeep Lstm-based Stacked Autoencoder for Multivariate Time Series Forecasting Problems. Scientific Reports, 2019, 9(1): 1-16.].

[0083] GNN is an abbreviation for Graph Neural Network, which is a class of algorithms that use neural networks to learn graph-structured data, extract and discover features and patterns in graph-structured data, and meet the needs of graph learning tasks such as clustering, classification, prediction, segmentation, and generation. Related theories can be found in the following reference: [F. Scarselli, M. Gori, ACTsoi, et al. The Graph Neural Network Model. IEEE Transactions on Neural Networks, 2008, 20(1):61-80.]

[0084] ODEs are an abbreviation for Ordinary Differential Equations. Ordinary differential equations parameterize the derivatives of hidden states using neural networks, rather than the discrete sequences of hidden layers used in traditional models (such as ResNet and RNNs). They offer a balance between numerical accuracy and computational cost, while significantly reducing memory usage. The theoretical basis can be found in the reference [RT Chen, Y. Rubanova, J. Bettencourt, and D. Duvenaud. Neural ordinary differential equations. In NeurIPS, 2018, pp. 6572–6583.]. Detailed Implementation

[0085] The present invention will be further described with reference to the accompanying drawings.

[0086] Example 1

[0087] The air compressor fault prediction and health assessment method based on graph neural networks provided in this embodiment, such as Figure 1 As shown, it includes:

[0088] S1 performs convolution processing on multivariate time series with different receptive fields to obtain multi-scale embedding features.

[0089] Given a multivariate time series Where T represents the length of the multivariate time series, N represents the N observed variables at time t (the observed variables of the air compressor include exhaust pressure, exhaust temperature, lubricating oil temperature, lubricating oil pressure, cooling water temperature, etc.), and M represents the observation dimension of each variable.

[0090] Multivariate time series X at different scales s 1:T Divide, divide into Data segments, for any time period t s The input time series is processed by convolution units, as shown below:

[0091]

[0092] The input and output channels of this two-dimensional convolution are M and d, respectively, which correspond to the size of the input features and the output embedded representation. The kernel size is 1×s.

[0093] Therefore, by processing each time period sequentially, a multi-scale embedding representation corresponding to scale s can be generated.

[0094] Based on different scales s = s1, s2, ..., s k By following the steps described above, a series of embedded features can be generated.

[0095] S2 establishes a dynamic temporal graph based on a multi-head attention mechanism; it uses graph convolution to discretize the dynamic temporal graph to obtain an aggregated representation with instantaneous intervals, and uses ordinary differential equations to solve it to obtain a multi-scale information representation.

[0096] First, based on multi-scale temporal embeddings for each variable Reshape them into a two-dimensional matrix, the resulting multi-scale time embedding H 2d It is expressed as follows:

[0097]

[0098] Then, through a linear transformation, they are mapped to a higher-order query matrix and key matrix, resulting in C parallel headers, represented as follows:

[0099]

[0100] Among them, all The learning parameter, d a Let represent the dimensions of the query and key matrices; then, dot multiplication is applied to calculate the attention weights between each pair of embeddings, using the following formula:

[0101]

[0102] in, It is the adjacency matrix of the dynamic time series graph. Finally, sort the matrix elements by size, keep the first p weights, and set the other weights to 0, because a few dot product weights contribute to the main attention, while the influence of other weights is negligible.

[0103] Therefore, this limitation can be overcome by performing a large number of aggregations over several discrete propagation steps. Specifically, let l′ = L / K represent the intermediate variable between K and L, and denote the aggregation interval. Based on the above setting, an aggregation with interval l′ at the s-scale can be expressed as follows:

[0104]

[0105] We use To simplify, we can now set L as a very small constant. As K→∞, l′→0. We obtain a continuous information propagation process with instantaneous intervals. When l=0, according to the above formula, we obtain an aggregated representation with instantaneous intervals.

[0106] The multi-scale information representation is obtained by solving the ordinary differential equation (neural ODE solver). for:

[0107]

[0108] in, G represents the embedding at each scale. s (·) can be any input The feedforward network. Here we let No additional learning parameters are required. This not only improves the computational efficiency of integral approximation but also aggregates linear features across all nodes. After infinite aggregation, It includes long-distance relationships between nodes in a graph structure and avoids over-smoothing.

[0109] Subsequently, patterns can be further extracted from the multivariate sequences, and nonlinearity can be introduced. The multi-scale information representation of the nodes is then updated through biomimetic transformations and activation functions, as follows:

[0110]

[0111] Where σ(·) represents the PReLU activation function, b represents the learned parameters.

[0112] Then, the multi-scale information representation of all nodes at scale s.

[0113] S3 uses the obtained multi-scale information representation to predict multivariate time series.

[0114] In this step, a multilayer perceptron is used as the predictor to generate multivariate time series predictions for τ time steps.

[0115]

[0116] Here, θ represents the learnable parameters.

[0117] Based on the prediction sequence, S4 calculates the health index of the air compressor according to the health model, and then performs fault prediction and health assessment of the air compressor.

[0118] The specific calculation method for a single-variable health index is as follows:

[0119]

[0120] in, The air compressor health index represents the nth attribute variable at the predicted τ time points. and These represent the maximum and minimum values ​​of the nth observed variable in the input time series, i.e., the maximum and minimum values ​​of observed variable n when the air compressor is working normally; This represents the standard value of the nth variable in the input time series.

[0121] The air compressor's operational health index is:

[0122]

[0123] Among them, W n This represents the weight of the given nth variable in terms of health, where n = 1, 2, ..., N.

[0124] Calculated HV T+τThe results range from 0 to 1. Research has revealed the following relationship between the air compressor's operational health index and its failure rate: when the operational health index is between 0.7 and 1, the equipment is considered healthy with a very low failure rate; when the operational health index is between 0.6 and 0.7 (excluding the extreme values ​​of 0.6 and 0.7), the equipment is considered sub-healthy with a low failure rate; and when the operational health index is between 0 and 0.6, the equipment is considered faulty with a high failure rate. Therefore, we can obtain the failure information and health status of the air compressor system based on the predicted data.

[0125] Example 2

[0126] This embodiment provides an air compressor fault prediction and health assessment system based on graph neural networks, which includes a multivariate time series prediction model and an air compressor fault prediction and health assessment model.

[0127] A multivariate time series prediction model, which includes a multi-scale embedded feature acquisition module, a multi-scale information representation acquisition module, a multivariate time series prediction module, and an optimizer.

[0128] The multi-scale embedding feature acquisition module is used to perform convolution processing on multivariate time series with different receptive fields to obtain multi-scale embedding features.

[0129] This module performs the operation according to step S1 in Example 1; it includes:

[0130] Multivariate time series partitioning units are used to divide an input multivariate sequence into several time periods based on different scales.

[0131] Multivariate time series are divided into units at different scales s for the multivariate time series X. 1:T Divide, divide into Data segments.

[0132] Convolutional units are used to perform convolution processing on the time series of each divided time period to generate multi-scale embedded representations.

[0133] For different scales, the convolutional unit has a different receptive field (i.e., kernel size). For scale s, the kernel size of the convolutional unit is 1×s.

[0134] The convolutional units are used to perform convolution processing on the time series of each time period obtained at the corresponding scale according to formula (1.1) to generate the corresponding multi-scale embedding representation.

[0135] The multi-scale information representation acquisition module is used to establish a dynamic temporal graph based on a multi-head attention mechanism; the dynamic temporal graph is discretized using graph convolution to obtain an aggregated representation with instantaneous intervals, and the multi-scale information representation is obtained by solving the ordinary differential equation.

[0136] This module performs operations according to step S2 above; it includes:

[0137] The dynamic sequence graph building unit is used to build a dynamic sequence graph based on the multi-head attention mechanism; specifically, the adjacency matrix A of the dynamic sequence graph is obtained according to the above formula (1.5).

[0138] The graph convolution processing unit is used to discretize the dynamic time series graph to obtain an aggregated representation with instantaneous intervals; specifically, the aggregated representation with instantaneous intervals is obtained according to the above formula (1.6).

[0139] The ODE solver is used to solve aggregated representations with instantaneous intervals to obtain multi-scale information representations; specifically, it solves for multi-scale information feature representations according to the above formula (1.7).

[0140] The update unit is used to update the multi-scale information features through biomimetic transformation and activation function; specifically, it updates the multi-scale information feature representation obtained by the ODE solver according to formula (1.8).

[0141] The multivariate time series prediction module is used to predict multivariate time series based on the obtained multi-scale information representation.

[0142] This module uses a multi-layer perceptron (MLP).

[0143] An optimizer is used to optimize the model parameters. In this embodiment, the Adam algorithm is used to calculate the gradient and update all learned parameters; and the MSE loss is used as the optimization objective.

[0144] An air compressor failure prediction and health assessment model is used to calculate the air compressor's health index based on the prediction sequence and the following health model, thereby performing failure prediction and health assessment of the air compressor:

[0145]

[0146]

[0147] Among them, W n This represents the weight of the health status of the given nth observation variable, where n = 1, 2, ..., N; N represents the number of observation variables. The air compressor health index represents the nth attribute variable at the predicted τ time points. and These represent the maximum and minimum values ​​of the nth observed variable in the input time series, respectively. The standard value of the nth variable in the input time series is represented by the average value of the nth observed variable in the input time series.

[0148] Application examples

[0149] The CGMF (Graph Neural Network-based Fault Prediction and Health Assessment) method for air compressors, as provided in Example 1, was used in experiments on a real-world multivariate time-series dataset of air compressors. The multivariate observed variables included: current (A), voltage (volts), exhaust pressure (kPa), exhaust temperature (degrees Celsius), cooling water temperature (degrees Celsius), and cooling water pressure (kPa), with a granularity of 1 ms / sample. Simultaneously, this invention was compared with five different baseline models: the statistical model ARIMA and neural network models (GRU, LatentODE, MTGNN, and STODE).

[0150] To meet the requirements of deep learning model learning problems, the dataset required for each test is divided into a training set, a validation set, and a test set in a ratio of 6:2:2. The training and validation sets are used to adjust the parameters and hyperparameters of the multivariate time series prediction model provided in Example 2, while the test set is used to test the air compressor fault prediction and health assessment system based on graph neural networks provided in Example 2 according to the method provided in Example 1. Furthermore, to verify the effectiveness of predicting air compressor system faults and assessing health at different time periods, multiple values ​​for τ are set. Considering the significance of time, predictions are made and results are reported for three future time periods: one hour, one day, and one week.

[0151] The Root Relative Squared Error (RSE) and Empirical Correlation Coefficient (CORR), widely used in time series forecasting, are used as evaluation metrics (the lower the RSE value, the better the model; the higher the CORR value, the better the model). Here, Ours represents the results of the model in this invention, with the best results indicated in bold.

[0152] Table 1: Prediction results for future time τ on all datasets in the application example.

[0153]

[0154] The remaining methods in the table are described below:

[0155] ARIMA is a typical time series forecasting model based on statistics. It consists of three parts: an autoregressive model, a moving average model, and the order of differencing. It only requires endogenous variables and does not need to rely on other exogenous variables to complete the forecast. It can predict water flow based on historical observations. The implementation process can be found in the reference [Wang, Wen-chuan, et al. "Improving forecasting accuracy of annual runoff time series using ARIMA based on EEMD decomposition." Water Resources Management 29.8(2015):2655-2675.].

[0156] GRU: A variant based on an improved Recurrent Neural Network (RNN). Compared to the original RNN, GRU adds some gated logic units, enabling the network to capture dependencies over longer periods and handle long-term sequence events. It also alleviates, to some extent, the gradient explosion and vanishing problems encountered by the original RNN during training. This type of model is widely used in time series forecasting problems. For implementation details, please refer to the following references: [Yang, Shuyu, et al. Real-time reservoir operation using recurrent neural networks and inflowforecast from a distributed hydrological model. Journal of Hydrology 579(2019):124229.][Apaydin, Halit, et al. Comparative analysis of recurrent neural network architectures for reservoir inflow forecasting. Water 12.5(2020):1500.]

[0157] LatentODE: It generalizes discrete RNNs to continuous-time hidden dynamics defined by ODE. It treats the latent representation as a time-series variable within the RNN, enabling it to handle arbitrary time intervals between observations. For its implementation details, refer to [Y. Rubanova, R.T. Chen, and D. Duvenaud. Latent odes for irregularly sampled timeseries. In NeurIPS, 2019.].

[0158] STODE: Based on the neural network of ordinary differential equations, this paper proposes a continuous graph neural network with a restart distribution for time-space data to capture long-range spatial correlations in multivariate traffic flow forecasting. For implementation details, please refer to [Z. Fang, et al. Spatial-temporal graph ode networks for trafficflow forecasting. Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining. 2021.].

[0159] MTGNN: Based on graph neural networks, this paper proposes a novel hybrid jump propagation layer and an expanding initiation layer to capture the spatial and temporal dependencies within time series by extracting undirected relationships between variable graphs and combining graph and temporal convolutional modules. The implementation process and details are referenced in the following literature: [Z. Wu, S. Pan, G. Long, et al., C. Connecting the dots: Multivariate time series forecasting with graph neural networks. In Proceedings of the 26th ACM SIGKDD international conference on knowledge discovery & data mining, 2020, pp. 753-763.]

[0160] As can be seen from the experimental results in Table 1, the air compressor fault prediction and health assessment method based on graph neural networks proposed in this invention can be used for multivariate time series prediction tasks of air compressor systems. After accurately predicting data for future time periods, a health assessment model is established to transform the predicted attributes into a health index, which can successfully predict faults. Furthermore, compared with other deep learning methods, this invention shows significant performance improvements in both relative square root error and empirical correlation coefficient.

[0161] Therefore, the graph neural network-based air compressor fault prediction and health assessment method proposed in this invention has been successfully applied to multivariate fault prediction and health assessment of air compressor systems. It integrates long-range internal and interrelationships between variables, as well as expressive cross-scale time information. By inputting multivariate time-series monitoring data of the air compressor system, the operating status values ​​of the air compressor in future time periods can be predicted relatively accurately, and fault analysis and health assessment can be performed. Compared with manual methods, this invention provides guidance for the intelligent management and fault and health assessment of air compressor systems. Simultaneously, this invention also provides a general model for time-series prediction tasks.

[0162] Those skilled in the art will recognize that the embodiments described herein are intended to help the reader understand the principles of the invention, and should be understood that the scope of protection of the invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations based on the technical teachings disclosed in this invention without departing from the spirit of the invention, and these modifications and combinations are still within the scope of protection of this invention.

Claims

1. A method for air compressor fault prediction and health assessment based on graph neural networks, characterized in that, Includes the following steps: S1 performs convolution processing on multivariate time series with different receptive fields to obtain multi-scale embedding features; Specifically, given a multivariate time series Where T represents the length of the multivariate time series, N represents the N observed variables at time t, and M represents the observation dimension of each variable; Multivariate time series X at different scales s 1:T Divide, divide into Data segments, for any time period t s The input time series is processed by convolution units, as shown below: ； The input and output channels of this two-dimensional convolution are M and d, respectively, which correspond to the size of the input features and the output embedded representation. The kernel size is 1×s. Therefore, by processing each time period sequentially, a multi-scale embedding representation corresponding to scale s can be generated. , ; Based on different scales By following the steps described above, a series of embedded features can be generated. ; S2 establishes a dynamic temporal graph based on a multi-head attention mechanism; the dynamic temporal graph is discretized using graph convolution to obtain an aggregated representation with instantaneous intervals, and the multi-scale information representation is obtained by solving ordinary differential equations. First, based on multi-scale temporal embeddings for each variable They are reshaped into a two-dimensional matrix, and the resulting multi-scale time embeddings are matrixed. It is expressed as follows: ； Then, through a linear transformation, they are mapped to a higher-order query matrix and key matrix, resulting in C parallel headers, represented as follows: ； ； Among them, all Represents the learning parameters, Let represent the dimensions of the query and key matrices; then, dot multiplication is applied to calculate the attention weights between each pair of embeddings, using the following formula: ； in, It is the adjacency matrix of a dynamic time sequence graph; Then, a scale of s has intervals The aggregation can be expressed in the following way: ； use To put it simply, when l=0, the above formula yields an aggregated representation with instantaneous intervals. ; The multi-scale information representation is obtained by solving ordinary differential equations. for: ； in, Represents the embedding at each scale. Any input is acceptable Feedforward network; here ; S3 uses the obtained multi-scale information representation to predict multivariate time series; S4 calculates the air compressor's health index based on the predicted sequence and the following health model, and then performs fault prediction and health assessment of the air compressor: ； ； in, This represents the weight of the health status of the given nth observation variable, where n = 1, 2, ..., N; N represents the number of observation variables. Indicates prediction The air compressor health index of the nth observed variable at a given time; and These represent the maximum and minimum values ​​of the nth observed variable in the input time series, respectively. This represents the standard value of the nth observed variable in the input time series.

2. The method for air compressor fault prediction and health assessment based on graph neural networks according to claim 1, characterized in that, Sort the matrix elements by size, keep the first p weights of A, and set the other weights to 0.

3. The method for air compressor fault prediction and health assessment based on graph neural networks according to claim 1 or 2, characterized in that, The multi-scale information representation of nodes is further updated through biomimetic transformations and activation functions, as shown below: ； in, Represents the PReLU activation function. and Indicates the learned parameters; Then, the multi-scale information representation of all nodes at scale s. , .

4. The method for air compressor fault prediction and health assessment based on graph neural networks according to claim 3, characterized in that, In step S3, a multilayer perceptron is used as the predictor to generate the predicted values. Multivariate time series forecasts at each time step: ； in, This represents the learnable parameters.

5. The method for air compressor fault prediction and health assessment based on graph neural networks according to claim 1, characterized in that, In step S4, the relationship between the operating health index of the air compressor and its failure rate is as follows: when the operating health index is between 0.7 and 1, the equipment is healthy and the failure rate is very low. When the operating health index is between 0.6 and 0.7, excluding the extreme values ​​of 0.6 and 0.7, the equipment health is considered sub-healthy, with a low failure rate. When the operating health index is between 0 and 0.6, the equipment health status is faulty, and the failure rate is high.

6. A fault prediction and health assessment system for air compressors based on graph neural networks, characterized in that, To implement the method according to any one of claims 1 to 5, comprising: A multivariate time series forecasting model, which includes: A multi-scale embedding feature acquisition module is used to perform convolution processing on multivariate time series with different receptive fields to obtain multi-scale embedding features; this module includes: Multivariate time series partitioning unit, used to divide the input multivariate sequence into several time periods according to different scales; Convolutional units are used to perform convolution processing on the time series of each divided time period to generate multi-scale embedded representations. A multi-scale information representation acquisition module is used to build dynamic temporal graphs based on a multi-head attention mechanism. It discretizes the dynamic temporal graph using graph convolution to obtain an aggregated representation with instantaneous intervals, and solves it using ordinary differential equations to obtain the multi-scale information representation. This module includes: The dynamic sequence graph creation unit is used to create dynamic sequence graphs based on a multi-head attention mechanism. The graph convolution processing unit is used to discretize the dynamic time series graph to obtain an aggregated representation with instantaneous intervals; The ODE solver is used to solve aggregated representations with instantaneous intervals to obtain multi-scale information representations. The multivariate time series prediction module is used to predict multivariate time series based on the obtained multi-scale information representation; this module uses a multilayer perceptron. An air compressor failure prediction and health assessment model is used to calculate the air compressor's health index based on the prediction sequence and the following health model, thereby performing failure prediction and health assessment of the air compressor: ； ； in, This represents the weight of the health status of the given nth observation variable, where n = 1, 2, ..., N; N represents the number of observation variables. Indicates prediction The air compressor health index of the nth observed variable at a given time; and These represent the maximum and minimum values ​​of the nth observed variable in the input time series, respectively. This represents the standard value of the nth observed variable in the input time series.

7. The air compressor fault prediction and health assessment system based on graph neural networks according to claim 6, characterized in that, The multi-scale information representation acquisition module further includes an update unit, which is used to update the multi-scale information features through biomimetic transformation and activation function.

8. The air compressor fault prediction and health assessment system based on graph neural networks according to claim 6, characterized in that, The multivariate time series forecasting model also includes an optimizer for optimizing the model parameters.

Images