An engine fault diagnosis method of a physically driven recurrent graph neural network

Abstract

A physical-driven recurrent graph neural network (RTB) method for engine fault diagnosis is disclosed. The method includes: using prior knowledge of the physical effects of multi-sensor interference, constructing a physical information graph of sensor signals by acquiring data from a high-fidelity model. For the constructed physical information graph, a RTB is used for modeling and parameter training to extract temporal and spatial information, giving the proposed model the advantage of considering both spatial and temporal information. The proposed method simultaneously considers signals from multiple sensors and, based on physical priors, can achieve fault diagnosis in both temporal and spatial dimensions for large and complex systems.

Description

Technical Field

[0001] This invention belongs to the field of complex system fault diagnosis, and relates to a method for diagnosing complex system faults using a physical-driven recurrent graph neural network. Background Technology

[0002] Fault diagnosis and prediction are crucial for the safe testing and operation of complex systems such as liquid rocket engines. Extreme operating conditions (such as high pressure and combustion temperature) and the complexity of mechanical systems significantly increase the likelihood of catastrophic failures. Unlike component-level diagnostics, complex system diagnostics prioritizes developing system-level diagnostic strategies with multiple sensors, providing operators with sufficient time to make decisions. To date, typical health monitoring systems are redline cutoff systems, primarily responding to system failures through manually set fixed thresholds. This approach was developed early on to improve rocket engine safety and was largely constrained by computational resources on flight or test benches. With the development of high-performance computing, there is an urgent need to research innovative technologies that integrate multi-sensor information.

[0003] Generally, fault diagnosis and prediction methods for mechanical systems fall into two categories: model-based methods and data-driven methods. Model-based methods, due to the clear physical processes, have been extensively studied for complex systems. They utilize linear models and methods such as Kalman filters to determine parameters from historical data. The obtained parameters are then used to predict sensor outputs. By comparing these parameters with real-time measurements, residuals are obtained, enabling fault detection. Richardson et al. explored the use of Kalman filters to analyze and validate sensor measurements of the RS-25 engine, developing a set of Kalman filters to generate estimates of sensor measurements and using the residuals between the predicted and sensor outputs to detect faults. Results showed that the Kalman filter algorithm can isolate faults between two redundant sensors but cannot address hardware redundancy. Kawatsu et al. proposed a model-based method for diagnosing electromechanical actuator faults to reduce the affordable maintenance costs and rapid turnaround time of reusable launch vehicles. This method involves multi-physics system-level modeling and simulation, and employs a dynamic time-warping algorithm to evaluate the residuals between previous datasets and in-flight collected values.

[0004] Model-based techniques rely heavily on accurate system models, which may be impractical for complex systems such as rocket engines. Another approach is data-driven methods, widely studied in most industrial sectors because features are extracted directly from monitoring data without requiring a physical model of the system. Nie et al. combined cloud models and backpropagation neural networks for fault diagnosis in liquid rocket engines. Their proposed algorithm achieved fault detection accuracy down to the second, meeting real-time requirements. Tsutsumi et al. proposed a fault detection method based on phase plane trajectories of feature vectors obtained from two sensors through principal component analysis. Their ability to detect sensor faults was demonstrated through static ignition tests. However, in the field of rocket engine fault diagnosis, the amount of fault state data is limited due to the relatively few launches or hot-fire tests per year. Therefore, more physical knowledge should be incorporated when developing fault diagnosis technologies.

[0005] These data-driven methods typically focus on extracting features individually from multiple sensors, neglecting the relationships between the different sensors in a system. However, in many real-world scenarios, the signals collected from sensors possess inherent characteristics. To model the local information from multiple sensors, the physical knowledge of sensor interference effects between them should be considered.

[0006] The information disclosed in the background section is only intended to enhance the understanding of the background of the present invention, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention proposes a physical-driven recursive graph neural network-based engine fault diagnosis method. The method uses a graph neural network to model the relationships between different sensors and extract temporal and spatial information.

[0008] The objective of this invention is achieved through the following technical solutions:

[0009] A physical-driven recurrent graph neural network-based engine fault diagnosis method includes the following steps:

[0010] Step 1: Install multiple sensors in the engine;

[0011] Step 2: Multiple sensors measure sensor monitoring data of the engine in normal and fault states in real time;

[0012] Step 3: Based on the sensor monitoring data and the prior physical knowledge that there is mutual interference between multiple sensors, construct an adjacency matrix to form a physical information graph.

[0013] Step 4: Use graph convolutional neural network to process the physical information graphs of normal and abnormal states, compare the processing results of normal and abnormal states, and obtain the fault diagnosis results.

[0014] In the method described above, in step 3, the physical information map is constructed as follows:

[0015] Based on the acquired monitoring data from multiple sensors, a physical information map is constructed according to the following rules:

[0016]

[0017] If a ij If the value is 1, then in the physical information graph, there is an edge between the nodes representing the two sensors; otherwise, there is no edge.

[0018] In the method described, the graph convolutional neural network used in step 4 is a recurrent graph neural network.

[0019] In the method described, in step 4, the graph convolutional neural network generates node embeddings based on the local neighborhood of each node. For any node in the graph convolutional neural network, its feature vector is considered as an aggregated mapping of the initial inputs of its surrounding nodes. When the graph convolutional neural network processes information from the physical information graph, it uses the adjacency matrix A∈{0,1} N×N Determine the adjacency set of a node, where A is the set representing whether there is an edge between two nodes, i.e., the adjacency matrix, and N represents the number of nodes. The adjacency matrix represents the effect of a node, which is calculated by the number of its adjacency nodes.

[0020] In the method described, in step 4, the graph convolutional neural network matrix is ​​defined as follows: in, A is an adjacent matrix, I N It is an ordered identifier matrix, D is composed of... The calculated diagonal matrix, f() is the activation function used in the neural network, x represents the input, and ω represents the learnable parameter matrix.

[0021] In the method described above, step 4 involves constructing a recursive operation to extract time information from the time series data. in, A is an adjacent matrix, I N It is an ordered identifier matrix, D is composed of... The calculated diagonal matrix, f() is the activation function used in the neural network, x represents the input, ω represents the learnable parameter matrix, and h t-1 This is the result of the previous learning session.

[0022] In the method described above, step 4 involves using a dense layer for prediction. The formula for the dense layer is as follows: Among them, h t The result is the processing result of the recurrent convolutional graph neural network. θ is the learnable parameter matrix, b is the learnable bias, and ReLU() is the activation function used. By learning to change the parameters of the dense layer, the trained dense layer is used to process the processing result of the recurrent graph neural network to obtain the prediction result.

[0023] Beneficial effects

[0024] This invention provides a method that integrates physical knowledge into a graphical recurrent neural network, extracting spatial and temporal features from multiple sensors for fault diagnosis, which aids in model interpretation. This method not only outperforms other methods in accuracy but also enables earlier fault detection. Attached Figure Description

[0025] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0026] In the attached diagram:

[0027] Figure 1 This is a schematic diagram of a liquid rocket engine structure.

[0028] Figure 2 The diagram shows the corresponding representations of the sensors on the engine.

[0029] Figure 3 This is a flowchart illustrating the process of using this method for fault diagnosis;

[0030] Figure 4 These are the results of parameter simulation under normal conditions;

[0031] Figure 5 The results are the parameter simulation results under leakage fault conditions.

[0032] The present invention will be further explained below with reference to the accompanying drawings and embodiments. Detailed Implementation

[0033] The following will refer to the appendix. Figures 1 to 5Specific embodiments of the invention will be described in more detail below. While specific embodiments of the invention are shown in the accompanying drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.

[0034] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this invention is determined by the appended claims.

[0035] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0036] The engine fault diagnosis method driven by a physical recursive graph neural network includes the following steps:

[0037] Step 1: Install multiple sensors in the engine;

[0038] Step 2: Multiple sensors measure sensor monitoring data of the engine in normal and fault states in real time;

[0039] Step 3: Based on the sensor monitoring data and the prior physical knowledge that there is mutual interference between multiple sensors, construct an adjacency matrix to form a physical information graph.

[0040] Step 4: Use graph convolutional neural network to process the physical information graphs of normal and abnormal states, compare the processing results of normal and abnormal states, and obtain the fault diagnosis results.

[0041] In a preferred embodiment of the method, in step 3, the physical information map is constructed as follows:

[0042] Based on the acquired monitoring data from multiple sensors, a physical information map is constructed according to the following rules:

[0043]

[0044] If a ijIf the value is 1, then in the physical information graph, there is an edge between the nodes representing the two sensors; otherwise, there is no edge.

[0045] In a preferred embodiment of the method, the graph convolutional neural network used in step 4 is a recurrent graph neural network.

[0046] In a preferred embodiment of the method, in step 4, the graph convolutional graph neural network generates node embeddings based on the local neighborhood of the nodes. For any node in the graph convolutional graph neural network, its feature vector is regarded as an aggregated mapping of the initial inputs of surrounding nodes. When the graph convolutional graph neural network processes information of the physical information graph, it uses the adjacency matrix A∈{0,1} N×N Determine the adjacency set of a node, where A is the set representing whether there is an edge between two nodes, i.e., the adjacency matrix, and N represents the number of nodes. The adjacency matrix represents the effect of a node, which is calculated by the number of its adjacency nodes.

[0047] In a preferred embodiment of the method, in step 4, the graph convolutional graph neural network matrix is ​​defined as follows: in, A is an adjacent matrix, I N It is an ordered identifier matrix, D is composed of... The calculated diagonal matrix, f() is the activation function used in the neural network, x represents the input, and ω represents the learnable parameter matrix.

[0048] In a preferred embodiment of the method, in step 4, for time series data, a recursive operation is constructed to extract time information from the time series data. in, A is an adjacent matrix, I N It is an ordered identifier matrix, D is composed of... The calculated diagonal matrix, f() is the activation function used in the neural network, x represents the input, ω represents the learnable parameter matrix, and h t-1 This is the result of the previous learning session.

[0049] In a preferred embodiment of the method, in step 4, a dense layer is used for prediction, and the formula for the dense layer is expressed as: Among them, h t The result is the processing result of the recurrent convolutional graph neural network. θ is the learnable parameter matrix, b is the learnable bias, and ReLU() is the activation function used. By learning to change the parameters of the dense layer, the processing result of the recurrent convolutional graph neural network in step 4 is processed using the learned and trained dense layer to obtain the prediction result.

[0050] In one embodiment, the engine is a liquid rocket engine.

[0051] In one embodiment, such as Figure 1 As shown, the example of liquid rocket engine fault diagnosis includes the following steps:

[0052] (1) Use 10 different types of sensors to collect liquid rocket engine operation signals.

[0053] Liquid rocket engine structure as follows Figure 1 As shown, hundreds of sensors are deployed to monitor the system's operational status, but 21 of these sensors are used to determine the threshold for normal operation. Analysis of the system's dynamics reveals a relationship between these sensors, meaning that measurements are influenced by other sensory signals. Based on this physical prior, several sensors are selected to construct a physical information graph. Figure 2 The image shows the selected sensors used to construct the influence map.

[0054] (2) A physical information graph is constructed based on 10 selected sensors, which is used as the input to a physical information recursive graph neural network.

[0055] In this embodiment, the physical information graph is constructed as follows: Figure 3 As shown, it can be represented as:

[0056]

[0057] If a ij If the value is 1, then in the physical information graph, there is an edge between the nodes representing the two sensors; otherwise, there is no edge.

[0058] (3) Based on the physical information graph obtained in step (2), a recursive graph neural network is used for processing. The processing procedure is as follows:

[0059] The basic idea of ​​a graph convolutional neural network (Graph Convolutional Graph Neural Network) is to generate node embeddings based on the local neighborhood of each node. For any node in a Graph Convolutional Graph Neural Network, its feature vector is considered as an aggregated mapping of the initial inputs of its surrounding nodes. When the graph network processes physical information graph information, it uses the adjacency matrix A∈{0,1} N×N Determine the adjacency set of a node, where A is the set representing whether there is an edge between two nodes, i.e., the adjacency matrix, and N represents the number of nodes. The adjacency matrix represents the effect of a node, which is calculated by the number of its adjacency nodes.

[0060] The matrix of a graph convolutional neural network is defined as follows: in, A is an adjacent matrix, I N It is an ordered identifier matrix, D is composed of... The calculated diagonal matrix, f(), is the activation function used in the neural network, x represents the input, and ω represents the learnable parameter matrix. However, for time series data, sequence information may be lost through graph neural networks. To address this issue, a recursive operation is constructed to extract the time information from the sequence data. The operation process is represented as:

[0061] in, A is an adjacent matrix, I N It is an ordered identifier matrix, D is composed of... The calculated diagonal matrix, f(), is the activation function used in the neural network, x represents the input, ω represents the learnable parameter matrix, and h t-1 This is the result of the previous learning session.

[0062] Finally, a dense layer is used for prediction. The formula for the dense layer is as follows: Among them, h t The result is the processing result of the recurrent convolutional graph neural network. θ is the learnable parameter matrix, b is the learnable bias, and ReLU() is the activation function used. By learning to change the parameters of the dense layer, the processing result of the recurrent convolutional graph neural network in step 4 is processed using the learned and trained dense layer to obtain the prediction result.

[0063] (4) Based on the processing results obtained in step (3), plot the parameter simulation results under leakage fault conditions, such as... Figure 5 As shown.

[0064] (5) Observe and compare the parameter simulation results under normal conditions, such as Figure 4 The parameter simulation results under the conditions shown are as follows: Figure 5 As shown, the fault diagnosis is completed.

[0065] Although embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of the present invention, and all of these are within the scope of protection of the present invention.

Claims

1. A method for engine fault diagnosis using a physical-driven recursive graph neural network, characterized in that, The method includes the following steps: Step 1: Install multiple sensors in the engine; Step 2: Multiple sensors measure sensor monitoring data of the engine in normal and fault states in real time; Step 3: Based on the sensor monitoring data and the prior physical knowledge that there is mutual interference between multiple sensors, construct an adjacency matrix to form a physical information graph. The physical information graph is constructed as follows: Based on the acquired monitoring data from multiple sensors, a physical information map is constructed according to the following rules: ， like In the physical information graph, there is an edge between the nodes representing the two sensors; otherwise, there is no edge. Step 4: Using a graph convolutional graph neural network to process the physical information graphs of normal and abnormal states, compare the processing results of normal and abnormal states to obtain the fault diagnosis result. The graph convolutional graph neural network used is a recurrent graph neural network. When the graph network processes the physical information graph information, it uses the adjacency matrix... Determine the adjacent set of a node, where, The set representing the relationship between two nodes, i.e., the adjacency matrix, is a set where there is an edge between them. The number of nodes is represented by an adjacency matrix, which represents the effect of a node and is calculated by the number of its neighboring nodes. For time series data, a recursive operation is constructed to extract the time information from the time series data. ,in, , It is an adjacent matrix. It is an ordered identifier matrix. It is by The calculated diagonal matrix, It is the activation function used in neural networks. Indicates input, Represents the learnable parameter matrix, This is the result of the previous learning session.

2. The method according to claim 1, characterized in that, In step 4, the graph convolutional neural network generates node embeddings based on the local neighborhood of each node. For any node in the graph convolutional neural network, Its feature vectors are all regarded as aggregated mappings of the initial inputs of surrounding nodes. When a graph convolutional graph neural network processes information from a physical information graph, it does so through the adjacency matrix. Determine the adjacent set of a node, where, The set representing the relationship between two nodes, i.e., the adjacency matrix, is a set where there is an edge between them. The adjacency matrix represents the effect of a node, which is calculated by the number of its adjacent nodes.

3. The method according to claim 1, characterized in that, In step 4, a dense layer is used for prediction. The formula for the dense layer is as follows: ,in, This is the result of processing by a recurrent convolutional graph neural network. For learnable parameter matrix, For learnable bias, The activation function is used, and the parameters of the dense layer are changed by learning. The processed results of the recurrent graph neural network are processed by the dense layer after learning and training to obtain the prediction result.

Images