Aero-engine remaining life prediction method based on multi-level graph feature fusion

Abstract

The present invention discloses a method for predicting the remaining life of an aircraft engine based on multi-level graph feature fusion, which relates to the technical field of health management and prediction of aircraft engines. The method comprises the following steps: obtaining an aircraft engine status data set; dividing the aircraft engine status data set into a training set and a test set; preprocessing the operating condition signals, multi-sensor time series signals and RUL labels in the training set and the test set; building a multi-level graph feature fusion model; using the training set to train the multi-level graph feature fusion model for a set number of rounds to obtain a trained multi-level graph feature fusion model; inputting the test set into the multi-level graph feature fusion model to obtain an aircraft engine RUL prediction sequence; smoothing the RUL prediction sequence predicted by the multi-level graph feature fusion model to obtain a RUL smoothed prediction sequence in real time. The method solves the problems of highly regularized feature extraction of multi-sensor time series signals and low credibility of prediction results.

Description

Technical Field

[0001] The present invention relates to the technical field of health management and prediction of aircraft engines, and in particular to a method for predicting the remaining life of an aircraft engine based on multi-level graph feature fusion. Background Art

[0002] Currently, with the increasing complexity and sophistication of modern industrial systems, higher requirements are being placed on the safety and reliability of systems, especially for critical systems such as aircraft engines. Once a failure occurs during operation, it may cause significant economic losses and threaten life safety. Therefore, the development of equipment fault prognosis and health management (PHM) technology is extremely important for ensuring system safety, improving system reliability, and reducing maintenance costs. Remaining useful life (RUL) prediction is an important part of the PHM system. Its main purpose is to analyze and process the sensor signals inside the aircraft engine, model and predict the remaining useful life of the aircraft engine, provide decision support for predictive maintenance of the equipment, and achieve the goals of improving equipment operation safety, reducing equipment maintenance costs, and increasing equipment service life.

[0003] Generally speaking, there are three main research approaches for remaining useful life prediction: physics-based models, data-driven methods, and hybrid methods. Due to the extremely complex physical mechanisms of equipment, it is difficult to establish precise mathematical models, which means that physics-based models and hybrid methods have limitations. However, with the continuous development of big data technology, data-driven methods have become a hot research topic. Unlike the other two approaches, this method does not require prior domain knowledge and can directly learn features that reflect the health status of the equipment from massive amounts of operating data. Based on the multivariate condition monitoring sequences recorded by the equipment sensor network, the degradation characteristics of the aircraft engine are modeled and mapped to obtain the current RUL prediction value of the aircraft engine. However, since the aircraft engine condition monitoring sequence is not one-dimensional but rather a multivariate time series signal collected by the sensor network, and the system's operating conditions and degradation patterns are not single, feature extraction is also difficult.

[0004] Deep learning technology has become the hottest topic in current data-driven methods due to its powerful feature extraction and data fitting capabilities. Examples include convolutional neural networks (CNNs) and long short-term memory networks (LSTMs). In 2022, Cheng et al. proposed a multidimensional recurrent neural network based on Bi-LSTM and Bi-GRU in their paper "Multi-dimensional recurrent neural network for remaining useful life prediction under variable operating conditions and multiple fault modes." This multidimensional recurrent neural network is used to predict the RUL under variable operating conditions and multiple faults, and can extract features from multiple sensor data and operating conditions through different channels. In their paper "Spatiotemporally multi-differential processing deep neural network and its application to equipment remaining useful life prediction," Xiang et al. proposed a spatiotemporal multi-differential deep neural network based on LSTM to predict the RUL of equipment, leveraging multi-trend and multi-stage information to improve feature extraction capabilities.

[0005] Currently, data-driven aircraft engine RUL prediction technology has two main defects:

[0006] First, data processing is highly regularized. Most existing methods treat the time series signals collected by sensors as regular Euclidean data similar to text, voice, and images, and use traditional, highly regularized deep learning technologies such as CNN and LSTM for processing. This ignores the topological characteristics contained in the sensor network. For example, in an aircraft engine, the fluid pressure and temperature at different monitoring locations have certain physical correlations and are not independent of each other. It is actually unreasonable to treat them as Euclidean data. In addition, this regularized data poses a security risk. Once the input order of the sensor data is exchanged during the input process of the model, completely wrong prediction results will appear, which leads to low robustness of the prediction model.

[0007] Second, the smoothness of the prediction results is ignored. The existing RUL prediction technology based on deep learning often has the problem of severe jitter in the prediction curve. For example, the difference in the prediction values ​​within several consecutive days is too large, which reduces the credibility of the prediction results and greatly affects the practical application of the technology. Summary of the Invention

[0008] The problem to be solved by the present invention is to address the deficiencies of the above-mentioned existing technologies and provide an aircraft engine remaining life prediction method based on multi-level graph feature fusion to solve the problems of highly regularized feature extraction of existing multi-sensor time series signal features and low credibility of prediction results.

[0009] The present invention provides a method for predicting the remaining life of an aircraft engine based on multi-level graph feature fusion, comprising the following steps:

[0010] Step 1: Acquire an aircraft engine status dataset; the aircraft engine status dataset includes several operating condition signals, several multi-sensor timing signals, several virtual sensor signals, engine health parameters, RUL labels, engine unit numbers, flight cycles, flight categories, and health status labels;

[0011] Step 2: Divide the aircraft engine status dataset into a training set and a test set;

[0012] Step 3: Preprocess the working condition signals, multi-sensor time series signals and RUL labels in the training set and test set;

[0013] Step 3.1: Perform maximum and minimum value normalization operations on the working condition signals and the multi-sensor time series signals to obtain a number of normalized working condition signals and multi-sensor time series signals;

[0014] Step 3.2: Sort the values ​​in each normalized multi-sensor time series signal to obtain a ranking value of each value, and form a ranked sequence based on the ranking values ​​of each value, thereby obtaining several ranked sequences from the multiple multi-sensor time series signals;

[0015] Step 3.3: Based on the rank sequence, use the Spearman correlation analysis method to calculate the normalized multivariate sensor time series signal x s The Spearman correlation coefficient between any two multivariate sensor time series signals is calculated, and a sensor network topology graph is constructed according to the Spearman correlation coefficient; the sensor network topology graph is composed of a plurality of sensor nodes representing sensors and edges connecting them;

[0016] The method for constructing a sensor network topology graph is specifically as follows: if the Spearman correlation coefficient between two multi-sensor time series signals is greater than or equal to a set threshold, then there exists an edge between the sensor nodes corresponding to the two multi-sensor time series signals, otherwise there does not exist an edge;

[0017] The Spearman correlation coefficient calculation formula is:

[0018]

[0019] Where n represents the sample size in the rank sequence; S′ 1,index is the index-th value in the rank sequence S'1; S' 2,index is the index-th value in the rank sequence S2; ρ s is the Spearman correlation coefficient;

[0020] Step 3.4: Perform degradation enhancement processing on the multivariate sensor time series signal to obtain a multivariate residual time series signal;

[0021] Step 3.4.1: Use the healthy operating condition signal in the operating condition signal as input and the healthy multi-sensor time series signal in the multi-sensor time series signal as output, and use a three-layer fully connected neural network (FNN) to fit the aircraft engine input-output model in the healthy state;

[0022] Step 3.4.2: Input any operating condition signal into the aircraft engine input-output model in a healthy state to obtain the corresponding ideal multi-sensor time series signal in a healthy state;

[0023] Step 3.4.3: Calculate the residual between the ideal multi-sensor timing signal in the healthy state and the real multi-sensor timing signal in the aircraft engine status dataset, and perform SG smoothing and denoising processing on a flight cycle basis to obtain the multi-sensor residual timing signal.

[0024] Step 3.5: Based on the sensor network topology, a graph signal is constructed for the multivariate residual time series signal to obtain a graph sequence of the entire life cycle of the aircraft engine; the graph signal includes three elements: topology, node features, and edge features;

[0025] The method for obtaining a graph sequence specifically comprises: using a time sliding window method to intercept a multivariate residual time series signal within a flight cycle according to a set window size, using the intercepted time series signal as a node feature, using the edge weights in the adjacency matrix corresponding to the sensor network topology map as edge features, constructing a graph signal from the sensor network topology map, the node features, and the edge features to obtain a subgraph, then placing multiple subgraphs within a flight cycle into a set to obtain a subgraph sequence of a flight cycle, and further obtaining a graph sequence of the entire life cycle of an aircraft engine composed of subgraph sequences of multiple cycles;

[0026] Step 3.6: Perform nonlinear processing on the original RUL label to define the RUL label;

[0027] The RUL label t Defined as:

[0028]

[0029] Among them, y t is the RUL label value of flight period t; t EOL Indicates the time when the engine fails and cannot be used; FPT indicates the time when the engine begins to degrade; RUL t Represents the original RUL label value of the current flight cycle t in the aircraft engine status dataset.

[0030] Step 4: Build a multi-level graph feature fusion model; the multi-level graph feature fusion model includes a node-level feature extraction model, a subgraph-level feature extraction model, a cycle-level feature extraction model, a time-level feature extraction model and an output layer in sequence;

[0031] Step 4.1: Build a node-level feature extraction model;

[0032] The node-level feature extraction model inputs a graph sequence including a subgraph sequence of m flight cycles of an aircraft engine, performs node-level feature extraction processing, and outputs a graph sequence including a subgraph sequence of m flight cycles of an aircraft engine that has undergone node-level feature extraction processing;

[0033] The node-level feature extraction model is composed of a number of GNN-ResNet blocks that are set in parallel and share parameters. The number of the GNN-ResNet blocks is equal to the number m of flight cycles contained in the input graph sequence. The GNN-ResNet block includes a first GATv2 layer, a first LayerNorm layer, a first nonlinear activation layer, a second GATv2 layer, a second LayerNorm layer, a first FNN layer and a second nonlinear activation layer. Each GNN-ResNet block aggregates the node features of the subgraph sequence of a flight cycle in units of subgraphs through the first GATv2 layer, and uses the first LayerNorm layer to perform node feature aggregation on the first GATv2 layer. The aggregated node feature vector group is normalized, and then the normalized node feature vector group is nonlinearly mapped through the first nonlinear activation layer. The node features are further aggregated using the second GATv2 layer, and the aggregated node feature vector group is normalized using the second LayerNorm layer. The first FNN layer is used to transform the dimension of the input subgraph sequence of the flight cycle and perform feature fusion with the normalized node feature vector group output by the second LayerNorm layer. The fused node feature vector group is then nonlinearly activated using the second nonlinear activation layer to obtain a subgraph sequence of a path that has been feature extracted by the GNN-ResNet block.

[0034] The first nonlinear activation layer is a ReLu activation function; the second nonlinear activation layer is a Tanh function;

[0035] Step 4.2: Build a sub-graph-level feature extraction model;

[0036] The subgraph-level feature extraction model inputs a graph sequence of a subgraph sequence of m flight cycles of an aircraft engine that has undergone node-level feature extraction processing, performs subgraph-level feature extraction processing, and outputs a subgraph-level feature sequence set consisting of subgraph-level feature sequences of m flight cycles; the subgraph-level feature extraction model is composed of a plurality of parallel and parameter-shared self-attention aggregation layers with temperature, the number of parallel paths of the self-attention aggregation layer being equal to the number m of flight cycles contained in the input graph sequence; the self-attention aggregation layer performs attention calculation on the input subgraph sequence through a self-attention mechanism with temperature to obtain a subgraph-level feature sequence; the self-attention mechanism with temperature performs similarity calculation on the nodes in each subgraph in the subgraph sequence output by the node-level feature extraction model, calculates the similarity between any node and other nodes, and normalizes these similarities into attention weights, wherein a temperature parameter is added to adjust the calculated value of the attention weight, increasing the value of the attention weight. Then, by weighted summing each node feature with the corresponding attention weight, the subgraph-level feature h output by the self-attention mechanism can be obtained. k,t , integrating sub-graph level features into sub-graph level feature sequences;

[0037] Step 4.3: Build a cycle-level feature extraction model;

[0038] The cycle-level feature extraction model inputs a sub-graph-level feature sequence set consisting of sub-graph-level feature sequences of m flight cycles, performs cycle-level feature extraction processing, and outputs a cycle-level feature sequence integrated from the m cycle-level features; the cycle-level feature extraction model is composed of a plurality of parallel and parameter-shared self-attention aggregation layers with temperature, the number of parallel self-attention aggregation layers being equal to the number m of flight cycles contained in the input graph sequence, the input of any self-attention aggregation layer being a sub-graph-level feature sequence of a flight cycle output by the sub-graph-level feature extraction model, and the output is the cycle-level features of the sub-graph-level feature sequence;

[0039] Step 4.4: Build a time-level feature extraction model;

[0040] The time-level feature extraction model includes a first Bi-LSTM layer, a second Bi-LSTM layer, and a third nonlinear activation layer in sequence. The input is a period-level feature sequence, and the output is a time-level degradation feature extracted from the time-level feature.

[0041] The third nonlinear activation layer is a Tanh function;

[0042] Step 4.5: Design the output layer;

[0043] The output layer includes a feature expansion layer, a second FNN layer, a third FNN layer, a fourth FNN layer, a fourth nonlinear activation layer, and a fifth nonlinear activation layer in sequence. The input is the time-level degradation feature output by the time-level feature extraction model, and the output is the RUL prediction value of the current m periods and the RUL prediction value of the next m periods; the fourth nonlinear activation layer is a Tanh function, and the fifth nonlinear activation layer is a ReLu function.

[0044] Step 5: Use the training set to train the multi-level graph feature fusion model for a set number of rounds to obtain a trained multi-level graph feature fusion model;

[0045] Step 6: Input the test set into the multi-level graph feature fusion model to obtain the aircraft engine RUL prediction sequence;

[0046] Step 6.1: Input the graph sequence into the node-level feature extraction model to extract node-level features, and obtain a graph sequence after node-level feature extraction;

[0047] The node-level feature extraction expression is:

[0048]

[0049]

[0050] Among them, α i,j represents the attention coefficient between node i and node j, Represents the node features of the current layer of node i, represents the node features of the next layer of node i, Represents the node features of the current layer of node j, and represents the learnable parameters, represents the set of neighbor nodes of node i, j represents the node number, and the value of j can be equal to i; [x i ||x j ] represents node features and The splicing combination of , LeakyReLU(·) represents the activation function;

[0051] Step 6.2: Input the graph sequence after node-level feature extraction into the sub-graph-level feature extraction model for sub-graph-level feature extraction to obtain a sub-graph-level feature sequence set;

[0052] The expression for sub-graph level feature extraction is:

[0053]

[0054]

[0055]

[0056] Among them, h k,t represents the subgraph-level features of the kth subgraph in flight cycle t; represents the attention coefficient of the i-th node in the k-th subgraph in flight cycle t; represents the features of the i-th node of the k-th subgraph in flight cycle t; represents the features of the jth node of the kth subgraph in flight cycle t; represents the similarity measure between the i-th node and the j-th node of the k-th subgraph in flight cycle t; d represents the node feature The dimension of ; N represents the number of nodes, i, j represent the node numbers; τ represents the temperature coefficient;

[0057] Step 6.3: Input the subgraph-level feature sequence set into the period-level feature extraction model to extract period-level features and obtain a period-level feature sequence;

[0058] The period-level feature extraction expression is:

[0059]

[0060]

[0061]

[0062] Among them, h t represents the periodic characteristics of the flight period t; α k,t represents the attention coefficient of the kth sub-graph in flight cycle t. The larger the coefficient, the higher the attention of the corresponding sub-graph, that is, the stronger the relationship with the degradation feature; represents the similarity measure between the kth subgraph and the lth subgraph in period t; d′ represents the period feature h k,t The dimension of ; M represents the number of subgraphs in the flight cycle t, k and l represent the subgraph numbers in a flight cycle; τ represents the temperature coefficient;

[0063] Step 6.4: Input the period-level feature sequence into the time-level feature extraction model for time-level feature extraction, and then input it into the output layer to obtain the aircraft engine RUL prediction sequence.

[0064] Step 7: Smooth the RUL prediction sequence predicted by the multi-level graph feature fusion model to obtain the RUL smooth prediction sequence in real time.

[0065] The smoothing method is:

[0066]

[0067] in, is the smoothed RUL prediction value of flight cycle t, is the RUL prediction value of the flight cycle t at the i′th position in the current RUL prediction sequence, β i′ is the weight value corresponding to the i′th position of the weight vector.

[0068] The aircraft engine remaining life prediction method based on multi-level graph feature fusion proposed in this invention has the following beneficial effects compared with the existing technology:

[0069] (1) Modeling the multivariate sensor time series signals into more general graph data is conducive to analyzing and extracting the degradation characteristics of the system from the non-Euclidean space, solving the problem of highly regularized feature extraction of the multivariate sensor time series signals and improving the prediction robustness of the model. Moreover, this method is developed based on a data-driven approach and does not require complex prior knowledge, thus avoiding the problem that the internal structure of the complex system is difficult to deeply analyze, and is more scalable and generalizable.

[0070] (2) A multi-level graph feature fusion method is proposed to extract the health features of aircraft engines from four levels: node level, subgraph level, cycle level, and time level. This method not only solves the problem of feature fusion difficulty caused by the variable length of data in different flight cycles, but also increases the information utilization rate of multi-sensor time series signals and improves the RUL prediction accuracy.

[0071] (3) In view of the problem that most common RUL prediction schemes based on deep learning have large fluctuations in the prediction curve, a simple and universal weighted line segment prediction smoothing method is proposed to replace the traditional point prediction. It not only improves the smoothness and reliability of the prediction, but also expands the scope of RUL prediction. It is worth noting that the proposed smoothing technique can realize real-time optimization of the RUL prediction value, which is essentially different from other methods that first calculate all RUL values ​​and then smooth. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 This is a flow chart of a method for predicting remaining life of an aircraft engine based on multi-level graph feature fusion in an embodiment of the present invention;

[0073] Figure 2 A sensor network topology diagram in an embodiment of the present invention;

[0074] Figure 3 1 is a comparison diagram of the multi-sensor time series signals before and after degradation enhancement processing in an embodiment of the present invention;

[0075] Among them, a) is the unprocessed multi-sensor time series signal; b) is the processed multi-sensor time series signal with degradation trend;

[0076] Figure 4 4 is a structural diagram of the feature extraction model at each level in an embodiment of the present invention;

[0077] Among them, a) is the structure diagram of the node-level feature extraction model based on the graph neural network; b) is the structure diagram of the subgraph-level feature extraction model based on the self-attention mechanism; c) is the structure diagram of the cycle-level feature extraction model based on the self-attention mechanism; d) is the structure diagram of the time-level feature extraction model based on Bi-LSTM and the structure diagram of the model output layer;

[0078] Figure 5 Schematic diagram of a Bi-LSTM network in an embodiment of the present invention;

[0079] Among them, a) is a specific network diagram; b) is a computing unit of LSTM in the network;

[0080] Figure 6 This is a comparison chart of RUL prediction values ​​in an embodiment of the present invention;

[0081] Figure 7 This is an attribution analysis diagram in an embodiment of the present invention;

[0082] Among them, a) is the RUL prediction curve of an aircraft engine; b) is the proportion of the attention coefficient of each sensor data in this prediction process. DETAILED DESCRIPTION

[0083] The specific implementation of the present invention is further described in detail below with reference to the accompanying drawings and examples.

[0084] The aircraft engine remaining life prediction method based on multi-level graph feature fusion in this embodiment is as follows: Figure 1 As shown, the method includes the following steps:

[0085] Step 1: Acquire an aircraft engine status dataset; the aircraft engine status dataset includes operating condition signals, multi-sensor timing signals, virtual sensor signals, engine health parameters, RUL labels, engine unit numbers, flight cycles, flight categories, and health status labels;

[0086] In this implementation, the NASA open-source aircraft engine status dataset N-CMAPSS is used as an example. This dataset contains a total of 9 available sub-datasets. Each sub-dataset contains the status monitoring data of multiple engines from healthy to failure throughout their life cycle, including 4 operating condition signals ω, 14 multi-sensor time series signals x, and 14 multi-sensor time series signals x. s , 14 virtual sensor signals x v, 10 engine health parameters θ, 1 RUL tag y, as well as engine unit number, flight cycle, flight category and health status tag. The method proposed in the present invention uses the operating condition signal ω and 14 multi-sensor time series signals x s To predict the current RUL value of the aircraft engine and the RUL changes in the future.

[0087] Step 2: Divide the aircraft engine status dataset into a training set and a test set;

[0088] The N-CMAPSS dataset is divided into 9 sub-datasets. Each sub-dataset contains different degradation patterns. In a sub-dataset, there are full life cycle data of multiple engines. The N-CMAPSS dataset is divided into training set and test set. The specific division is shown in the following table:

[0089] Table 1. Division of training set and test set

[0090]

[0091] Step 3: Preprocess the working condition signals, multi-sensor time series signals and RUL labels in the training set and test set;

[0092] The dataset reveals that in aircraft engine RUL prediction, the coexistence of multiple flight states and the varying lengths of data within different flight cycles in the raw data easily conceal degradation trends in the engine's health, making it difficult to directly observe health degradation information from the time-series signals collected by the sensors. This paper designs operations to enhance flight data degradation trends and construct a topology map of the sensor network.

[0093] Step 3.1: Perform maximum and minimum value normalization operations on the working condition signals and the multi-sensor time series signals to obtain a number of normalized working condition signals and multi-sensor time series signals;

[0094] In this embodiment, the maximum and minimum value normalization operation is performed on the working condition signal and the multi-sensor time series signal. The normalization formula is as follows:

[0095]

[0096] Among them, x′ is the normalized time series signal, x is the original time series signal, and x min is the minimum value in the original time series signal, x max is the maximum value in the original time series signal.

[0097] Step 3.2: Sort the values ​​in each normalized multi-sensor time series signal to obtain a ranking value of each value, and form a ranked sequence based on the ranking values ​​of each value, thereby obtaining several ranked sequences from the multiple multi-sensor time series signals;

[0098] In this embodiment, it is assumed that any two time series of the 14 multi-sensor time series signals are S1 and S2, and the values ​​of these two series are sorted to obtain ranked sequences S'1 and S'2.

[0099] Step 3.3: Based on the rank sequence, use the Spearman correlation analysis method to calculate the normalized multivariate sensor time series signal x s The Spearman correlation coefficient between any two multivariate sensor time series signals is calculated, and a sensor network topology graph is constructed according to the Spearman correlation coefficient; the sensor network topology graph is composed of a plurality of sensor nodes representing sensors and edges connecting them;

[0100] The method for constructing the sensor network topology graph is as follows: if the Spearman correlation coefficient between two multi-sensor time series signals is greater than or equal to a set threshold, then there is an edge between the sensor nodes corresponding to the two multi-sensor time series signals, otherwise there is no edge; in this embodiment, the multi-sensor time series signal x is analyzed by using the Spearman correlation analysis method. s The monotonic relationship between them is characterized, and the connection relationship between sensor nodes is constructed based on this. The Spearman correlation coefficient of the two multivariate sensor time series signals can be calculated as:

[0101]

[0102] Where n represents the sample size in the rank sequence; S′ 1,index is the index-th value in the rank sequence S'1; S' 2,index is the index-th value in the rank sequence S2; ρ s is the Spearman correlation coefficient, which indicates the rank correlation strength between variables. The larger the value, the more similar the monotonic changes between the two variables are. Generally speaking, if ρ s ∈[0.8,1.0], it can be considered that the two sequences are very strongly correlated. Therefore, the method proposed in this invention selects 0.8 as the threshold, that is, when ρ s When ≥0.8, it is considered that there is an edge relationship between the sensor nodes corresponding to the multi-sensor time series signals. The sensor network topology diagram finally calculated is as follows: Figure 2 shown.

[0103] Step 3.4: Perform degradation enhancement processing on the multivariate sensor time series signal to obtain a multivariate residual time series signal;

[0104] Step 3.4.1: Use the healthy operating condition signal in the operating condition signal as input and the healthy multi-sensor time series signal in the multi-sensor time series signal as output, and use a three-layer fully connected neural network (FNN) to fit the aircraft engine input-output model in the healthy state;

[0105] In order to make the hidden degradation trend visible, a degradation enhancement operation of the multi-sensor timing signal is designed. The working condition signal ω in the healthy state is used as input, and the 14 multi-sensor timing signals x s The healthy multi-sensor time series signals in the model are used as outputs, and a three-layer fully connected neural network (FNN) is used to fit the healthy aircraft engine input-output model. The specific parameters are: the first-layer FNN has 128 hidden neurons and uses the ReLU function as a nonlinear activation function; the second-layer FNN has 64 hidden neurons and uses the ReLU function as a nonlinear activation function, with a dropout layer with a parameter of 0.2 to avoid overfitting; the third layer is a linear output layer. From then on, simply inputting the operating conditions into the healthy aircraft engine input-output model yields the ideal multi-sensor time series signals for the equipment's health.

[0106] Step 3.4.2: Input any operating condition signal into the aircraft engine input-output model in a healthy state to obtain the corresponding ideal multi-sensor time series signal in a healthy state;

[0107] Step 3.4.3: Calculate the residual between the ideal multi-sensor time series signal in the healthy state and the real multi-sensor time series signal in the aircraft engine status dataset, and perform SG smoothing and denoising processing on a per-flight cycle basis to obtain the multi-sensor residual time series signal.

[0108] The comparison of the multi-sensor time series signal before and after degradation enhancement in this embodiment is as follows: Figure 3 shown.

[0109] Step 3.5: Based on the sensor network topology, a graph signal is constructed for the multivariate residual time series signal to obtain a graph sequence of the entire life cycle of the aircraft engine; the graph signal includes three elements: topology, node features, and edge features;

[0110] The method for obtaining a graph sequence specifically comprises: using a time sliding window method to intercept a multivariate residual time series signal within a flight cycle according to a set window size, using the intercepted time series signal as a node feature, using the edge weights in the adjacency matrix corresponding to the sensor network topology map as edge features, constructing a graph signal from the sensor network topology map, the node features, and the edge features to obtain a subgraph, then placing multiple subgraphs within a flight cycle into a set to obtain a subgraph sequence of a flight cycle, and further obtaining a graph sequence of the entire life cycle of an aircraft engine composed of subgraph sequences of multiple cycles;

[0111] In this embodiment, the window length of the time sliding window is set to D, and the multivariate residual time series signal within a flight cycle is segmented. The signal obtained by segmenting the i-th multivariate residual time series signal can be expressed as And use it as the node feature of the corresponding sensor node in the graph. About the edge feature between sensor node i and sensor node j in the graph The adjacency matrix corresponding to the sensor network topology graph is The graph signal constructed by the sensor network topology, node features and edge features is defined as an undirected subgraph G k,t =(V,E), since the length of data in different flight cycles is not fixed, the number of sub-graphs segmented is also not fixed, where G k,t represents the kth subgraph with a flight period of t; represents a sensor node, Represents the edge relationship between sensor node i and sensor node j. Then the subgraph sequence of a flight cycle t can be described as Where M is the number of subgraphs cut out in the flight cycle t. Therefore, we can further obtain the graph sequence of the entire life cycle of the aircraft engine Q represents the total number of flight cycles in the entire life cycle of the aircraft engine.

[0112] Step 3.6: Perform nonlinear processing on the original RUL label to define the RUL label;

[0113] Regarding the processing of RUL tags, considering that the degradation of the equipment is not purely linear, it is optimized as a percentage of piecewise linear degradation. The RUL tag y of the flight cycle t t Defined as:

[0114]

[0115] Among them, y t is the RUL label value of flight period t; t EOLIndicates the time when the engine fails and cannot be used; FPT indicates the time when the engine begins to degrade; RUL t Represents the original RUL label value of the current flight cycle t in the aircraft engine status dataset.

[0116] In this embodiment, given a graph sequence of several periods of aircraft engine life cycle, a RUL sequence of several periods is predicted. Therefore, the RUL sequence is defined as The present invention takes the graph sequence of the first 8 flight cycles as an example to predict the RUL values ​​of the current 8 cycles and the next 8 cycles, that is:

[0117]

[0118] where Y t is the RUL tag sequence of flight period t, is the RUL prediction sequence, y t The predicted value of .

[0119] Step 4: Build a multi-level graph feature fusion model; the multi-level graph feature fusion model includes a node-level feature extraction model, a subgraph-level feature extraction model, a cycle-level feature extraction model, a time-level feature extraction model and an output layer in sequence;

[0120] Step 4.1: Build a node-level feature extraction model;

[0121] The input of the node-level feature extraction model is a graph sequence including a subgraph sequence of several flight cycles of an aircraft engine. The output is a graph sequence that has been processed with node-level feature extraction, including a subgraph sequence of several flight cycles of an aircraft engine. One of the input and output subgraph sequences is and The node-level feature extraction model is composed of several parallel and parameter-shared GNN-ResNet blocks. The number of GNN-ResNet blocks is equal to the number of flight cycles m contained in the input graph sequence. For different numbers of subgraphs contained in different cycles, they are treated as variable batch_size for batch processing. The specific model structure is as follows: Figure 4As shown in a), the GNN-ResNet block includes a first GATv2 layer, a first LayerNorm layer, a first nonlinear activation layer, a second GATv2 layer, a second LayerNorm layer, a first FNN layer, and a second nonlinear activation layer; the node-level feature extraction model inputs the data of m flight cycles into m GNN-ResNet blocks respectively. Taking one of the paths as an example: the subgraph sequence of a flight cycle is subjected to node feature aggregation by the first GATv2 layer in units of subgraphs, and the node feature vector group after aggregation by the first GATv2 layer is normalized by the first LayerNorm layer, and then the normalized node feature vector is subjected to the first nonlinear activation layer. The node feature vector group is nonlinearly mapped, and then the second GATv2 layer is used to further aggregate node features, and the second LayerNorm layer is used to normalize the aggregated node feature vector group. Then, the first FNN layer is used to transform the dimension of the subgraph sequence of the input flight cycle and perform feature fusion with the normalized node feature vector group output by the second LayerNorm layer. Then, the second nonlinear activation layer is used to perform nonlinear activation on the fused node feature vector group to obtain a subgraph sequence of a path extracted by the GNN-ResNet block feature. Finally, the subgraph sequences of m paths extracted by the GNN-ResNet block feature are integrated into a graph sequence extracted by node-level features. And send the graph sequence to the sub-graph level feature extraction model;

[0122] In this implementation, the number of parallel paths is 8. The specific parameters of the first GATv2 layer are an input feature dimension of 64, an output feature dimension of 128, a dropout value of 0.2, and a number of attention heads of 4. The first LayerNorm layer normalizes the node feature vector group aggregated by GATv2 to stabilize the gradient of backpropagation and accelerate the convergence of the network. The specific parameters are an input and output dimension of 128. The first nonlinear activation layer is the ReLu activation function. The specific parameters of the second GATv2 layer are an input feature dimension of 128, an output feature dimension of 256, a dropout value of 0.2, and a number of attention heads of 4. The specific parameters of the second LayerNorm layer are an input and output dimension of 256. The outputs of the first FNN layer and the second LayerNorm layer are feature fused. This part of the design draws on the idea of ​​residual networks, allowing inputs to propagate forward faster through cross-layer circuits and improve fitting ability. The specific parameters are an input feature dimension of 64 and an output feature dimension of 256. The second nonlinear activation layer is a tanh layer, which outputs the fused features to the next level of tasks through the Tanh nonlinear function.

[0123] Step 4.2: Build a sub-graph-level feature extraction model;

[0124] The subgraph-level feature extraction model is composed of several parallel and parameter-shared self-attention aggregation layers with temperature. The number of parallel paths of the self-attention aggregation layer is equal to the number of flight cycles m contained in the input graph sequence. The graph sequence after node feature extraction is input into m self-attention aggregation layers according to the flight cycle. That is, the input of any self-attention aggregation layer is the subgraph sequence after node feature extraction of one flight cycle output by the node-level feature extraction model. The output is the subgraph-level feature sequence {h 1,t ,h 2,t ,…,h M,t The self-attention aggregation layer performs attention calculation on the input subgraph sequence through the self-attention mechanism with temperature to obtain the subgraph level feature sequence. The self-attention mechanism with temperature calculates the attention of each subgraph in the subgraph sequence output by the node-level feature extraction model. The similarity calculation is performed on the nodes in the graph, and the similarity between any node and other nodes is calculated. These similarities are normalized into attention weights, where a temperature parameter is added to adjust the calculated value of the attention weight, increasing the value of the attention weight. Then, by weighted summing each node feature with the corresponding attention weight, the subgraph-level feature h output by the self-attention mechanism can be obtained. k,t , integrate the sub-graph level features into a sub-graph level feature sequence {h 1,t ,h 2,t ,…,h M,t}, and finally the subgraph-level feature sequences of the m paths are integrated into a subgraph-level feature sequence set and sent to the cycle-level feature extraction model, where the subgraph-level feature sequence set is represented as {{h 1,t ,h 2,t ,…,h M,t},…,{h 1,t+m-1 ,h 2,t+m-1 ,…,h M′,t+m-1}}.

[0125] In this embodiment, the subgraph-level feature extraction model includes a self-attention aggregation layer with temperature and 8 parallel paths. This network layer is improved based on the attention mechanism. The attention mechanism is a method that imitates the human visual and cognitive system, which allows the neural network to assign different weights to different positions in the input sequence. By introducing the attention mechanism, the neural network can automatically learn and selectively focus on important information in the input, improving the performance and generalization ability of the model. Specifically, Figure 4As shown in b), for each node in the graph signal, the self-attention mechanism calculates the similarity between it and other nodes, and normalizes these similarities into attention weights. Then, by weighted summing each node feature with the corresponding attention weight, the subgraph feature output by the self-attention mechanism can be obtained. The present invention introduces a temperature parameter to expand the value of the attention coefficient, so that nodes with large similarities occupy the dominant factor of the subgraph features, focusing on those sensor nodes that are strongly correlated with the degradation of health parameters, and also laying the foundation for the attribution analysis of subsequent RUL prediction results. The specific network layer parameters are 4 attention heads and a temperature coefficient of 0.1.

[0126] Step 4.3: Build a cycle-level feature extraction model;

[0127] The cycle-level feature extraction model is composed of several parallel and parameter-sharing self-attention aggregation layers with temperature. The number of parallel self-attention aggregation layers is equal to the number of flight cycles m contained in the input graph sequence. The sub-graph-level feature sequence set is divided into m sub-graph-level feature sequences according to the flight cycle, and each is input into m self-attention aggregation layers. That is, the input of any self-attention aggregation layer is the sub-graph-level feature sequence of a flight cycle output by the sub-graph-level feature extraction model, and the output is the cycle-level feature of the sub-graph-level feature sequence. Finally, the cycle-level features of the m paths are integrated into the cycle-level feature sequence {h t ,h t+1 ,…,h t+m-1} is sent to the time-level feature extraction model, where h t It is the cycle-level feature fused from all sub-graph-level features in flight cycle t.

[0128] In this embodiment, the cycle-level feature extraction model includes a self-attention aggregation layer with temperature, and the number of parallel paths is 8, such as Figure 4 As shown in c), all sub-graph-level features within a flight cycle are weighted and fused into a cycle-level feature through a temperature-based self-attention aggregation layer. All cycle-level features of the parallel paths are then integrated to obtain a cycle-level feature sequence, preparing for subsequent time series prediction. The specific network layer parameters are 4 attention heads and a temperature coefficient of 0.1.

[0129] Step 4.4: Build a time-level feature extraction model;

[0130] The temporal feature extraction model includes the first Bi-LSTM layer, the second Bi-LSTM layer and the third nonlinear activation layer in sequence. The input is a periodic feature sequence {h t ,h t+1 ,…,h t+m-1}, the output is the time-level degradation feature after time-level feature extraction;

[0131] The purpose of the temporal feature extraction model in this embodiment is to extract degradation features from the periodic feature sequence obtained above and associate them with RUL labels through the output layer. The specific network layer parameters are: input feature dimension 256, number of hidden layer units 32, number of network layers 2, and dropout value 0.2.

[0132] Step 4.5: Design the output layer;

[0133] The output layer is a feature space mapping network using a fully connected neural network, which includes a feature expansion layer, a second FNN layer, a third FNN layer, a fourth FNN layer, a fourth nonlinear activation layer, and a fifth nonlinear activation layer. The input is the time-level degradation feature output by the time-level feature extraction model, and the output is the RUL prediction value of the current several periods and the RUL prediction value of the next several periods.

[0134] In this embodiment, the second FNN layer has an input feature dimension of 512, an output dimension of 64, and a dropout value of 0.2; the third FNN layer has an input feature dimension of 64, an output feature dimension of 32, and a dropout value of 0.2; the fourth FNN layer has an input feature dimension of 32, an output feature dimension of 16, the fourth nonlinear activation layer is a Tanh function, and the fifth nonlinear activation layer is a ReLu function.

[0135] In this embodiment, the time-level degradation features of 8 cycles are expanded into a column of feature vectors (2*8*32=512), and then the above-obtained column of feature vectors is subjected to dimensionality reduction mapping through the second, third and fourth FNN layers, and is mapped to be consistent with the dimension of the RUL of several flight cycles to be predicted. Then, the fourth and fifth nonlinear activation layers are used to perform nonlinear activation processing on the feature vectors output by the above FNN layers, and the predicted values ​​are constrained within the range of [0,1]. The obtained output vector is the RUL predicted value of several flight cycles to be predicted. Here, the 16 values ​​of the output vector correspond to the RUL predicted values ​​of the current 8 cycles and the predicted values ​​of the next 8 cycles, respectively.

[0136] Step 5: Use the training set to train the multi-level graph feature fusion model for a set number of rounds to obtain a trained multi-level graph feature fusion model;

[0137] Generally speaking, for RUL prediction, we hope that the trained model can more accurately fit the degradation stage, which has a more important guiding significance for predictive maintenance. Therefore, referring to the design ideas of the reward function in reinforcement learning, the flight cycle time t is added to the traditional MSE loss function, and its expression is:

[0138]

[0139] Where loss represents the loss value, Indicates the RUL prediction value corresponding to the flight cycle t, y t Indicates the RUL label corresponding to the flight cycle t, where t represents the flight cycle of the aircraft.

[0140] Step 6: Input the test set into the multi-level graph feature fusion model to obtain the aircraft engine RUL prediction sequence;

[0141] In this embodiment, a graph sequence including an 8-cycle subgraph sequence is input to obtain a 16-cycle aircraft engine RUL prediction value;

[0142] Step 6.1: Input the graph sequence into the node-level feature extraction model to extract node-level features, and obtain a graph sequence after node-level feature extraction;

[0143] In the node-level feature extraction task, the GNN network and the residual neural network are mainly used. On the basis of the two, a GNN-ResNet block is constructed. This block contains two layers of GATv2. Its calculation formula is shown in Equations (6) and (7). This module can not only extract the features of the sensor node itself, but also aggregate the features of adjacent nodes, and learn the degenerate features contained in each node more richly. Specifically, the data is first input into a GATv2 layer for the first feature extraction. Then, it passes through a LayerNorm layer to normalize the node features aggregated by the GATv2 layer, stabilize the gradient of back propagation, and accelerate the convergence of the network. Then, it passes through another GATv2 layer for the second feature extraction, and the node normalization operation mentioned above is also performed. Then, it is fused with the features processed by the FNN of the jump link. This part of the design draws on the idea of ​​the residual network, allowing the input to propagate forward faster through the cross-layer circuit, improving the fitting ability. Finally, the fused features are output to the next level of task through the Tanh nonlinear function.

[0144]

[0145]

[0146] where α i,j represents the attention coefficient between node i and node j, Represents the node features of the current layer of node i, represents the node features of the next layer of node i, Represents the node features of the current layer of node j, and represents the learnable parameters, represents the set of neighbor nodes of node i, j represents the node number, and the value of j can be equal to i; [xi ||x j ] represents node features and LeakyReLU(·) represents the activation function.

[0147] Step 6.2: Input the graph sequence after node-level feature extraction into the sub-graph-level feature extraction model for sub-graph-level feature extraction to obtain a sub-graph-level feature sequence set;

[0148] In the subgraph-level feature extraction task, a temperature-based self-attention aggregation layer is designed using the attention mechanism to fuse the features of all nodes in the subgraph, focusing on sensor nodes that are strongly correlated with health parameter degradation, and laying the foundation for subsequent attribution analysis of RUL prediction results. The specific expression is as follows:

[0149]

[0150]

[0151]

[0152] Among them, h k,t represents the subgraph-level features of the kth subgraph in flight cycle t; represents the attention coefficient of the i-th node in the k-th subgraph in flight cycle t. The larger the coefficient, the higher the attention of the corresponding node, that is, the stronger the relationship with the degradation feature; represents the features of the i-th node of the k-th subgraph in flight cycle t; represents the features of the jth node of the kth subgraph in flight cycle t; represents the similarity measure between the i-th node and the j-th node of the k-th subgraph in flight cycle t; d represents the node feature dimension; N represents the number of nodes, i, j represent the node numbers; τ represents the temperature coefficient, which can control the concentration of attention. If τ ≤ 1, the original large value of the attention coefficient can be amplified, making the feature fusion more targeted, guiding the model to learn features that are more relevant to RUL changes, and enhancing the accuracy of prediction.

[0153] Step 6.3: Input the subgraph-level feature sequence set into the period-level feature extraction model to extract period-level features and obtain a period-level feature sequence;

[0154] In the cycle-level feature extraction task, the same processing strategy as the sub-graph-level task is adopted, as shown in Equations (11) to (13), except that the purpose of this task is to fuse all sub-graph features within a flight cycle to lay the foundation for subsequent time-level feature extraction. This operation can also solve the problem of different lengths of flight cycle data in the original data, that is, all data within a flight cycle are highly aggregated into a high-dimensional feature that is strongly correlated with the engine health status, namely the cycle-level feature h t .

[0155]

[0156]

[0157]

[0158] Among them, h t represents the periodic characteristics of the flight period t; α k,t represents the attention coefficient of the kth sub-graph in flight cycle t. The larger the coefficient, the higher the attention of the corresponding sub-graph, that is, the stronger the relationship with the degradation feature; represents the similarity measure between the kth subgraph and the lth subgraph in period t; d′ represents the period feature h k,t dimension; M represents the number of subgraphs in the flight cycle t, k and l represent the subgraph numbers in a flight cycle; τ represents the temperature coefficient.

[0159] Step 6.4: Input the period-level feature sequence into the time-level feature extraction model for time-level feature extraction, and then input it into the output layer to obtain the aircraft engine RUL prediction sequence;

[0160] In the time-level feature extraction task, the bidirectional long short-term memory network (Bi-LSTM) is used to extract time-series features. The high-dimensional features that can characterize the health status of the aircraft engine, namely the period-level feature sequence {h t ,h t+1 ,…,h t+7} is input into the network to further extract the characteristics of the changes in the health status of the aircraft engine during different flight cycles. The structure of Bi-LSTM is as follows Figure 5 As shown in a), each LSTM computing unit structure is as follows Figure 5 b), the specific feature calculation formula is as follows:

[0161] f t =σ(W xf h t +b xf +W hf p t-1 +b hf )

[0162] i t =σ(W xi h t +b xi +W hi p t-1 +b hi )

[0163] g t =tanh(W xg h t +b xg +W hg p t-1 +b hg )

[0164] o t =σ(W xo h t +b xo +W ho p t-1 +b ho )

[0165] c t =f t ⊙c t-1 +i t ⊙g t

[0166] h′ t =o t ⊙tanh(c t ) (14)

[0167] Among them, h t represents the periodic characteristics of the flight period t; p t-1 Indicates the hidden state value of the previous moment; c t Represents the state value of the memory unit, f t Represents the output value of the forget gate, which is used to control the memory state value c at the previous moment t-1 The proportion of information preserved; i t Represents the value of the input gate, selectively memorizes the input state h t The state value c of the memory unit at the previous moment t-1 The impact of g t Represents the candidate memory cell value, when updating c t The candidate memory unit value g will be calculated first t , used to update the memory unit state value; o t Indicates the value of the output gate, which controls the current memory cell value c t The output of h′ t Represents the output value of the LSTM basic unit, which is passed to the next layer; W xf, W hf , W xi , W hi , W xg , W hg , W xo , W ho and b xf , b hf , b xi , b hi , b xg , b hg , b xo , b ho σ represents the learnable parameters of each unit, and σ represents the Sigmoid activation function. The gating mechanism enables LSTM networks to incorporate long-range dependency information while preventing exploding and vanishing gradients during training. Bi-LSTM can extract features from the input time series in both forward and reverse order, allowing the feature data to incorporate both past and future information, resulting in more accurate predictions.

[0168] Finally, the time-level degradation features processed by the time-level feature extraction model are mapped to the RUL space through the output layer to obtain the RUL prediction value sequence of the aircraft engine. The complete framework of the proposed method is shown in the figure. Figure 1 shown.

[0169] Step 7: Smooth the RUL prediction sequence predicted by the multi-level graph feature fusion model to obtain the RUL smooth prediction sequence in real time;

[0170] Taking into account that the common RUL prediction schemes based on deep learning currently almost all have the problem of large fluctuations in the prediction curve. The present invention proposes a universal RUL prediction smoothing technique, which improves the point prediction in the traditional RUL prediction method to a line segment prediction. The improved prediction method can not only increase the prediction range of RUL, but also optimize the prediction results by using multiple prediction values ​​at the same time, making the prediction results more accurate and the prediction curve smoother. It is worth noting that the smoothing technique proposed in the present invention can achieve real-time optimization of the RUL prediction value. The model can predict the RUL prediction values ​​corresponding to the current and future cycles from the data of several cycles input each time, which is essentially different from the method of first calculating all RUL values ​​and then performing smoothing. In order to adjust the prediction performance, a confidence weight β is added here to more reasonably optimize the RUL prediction value. The final output RUL prediction value is shown in the following formula:

[0171]

[0172] in, is the smoothed RUL prediction value of flight cycle t, is the RUL prediction value of the flight cycle t at the i′th position in the current RUL prediction sequence, β i′ is the weight value corresponding to the i′th position of the weight vector.

[0173] In this implementation, the input graph sequence contains data from 8 flight cycles, and the corresponding RUL predicted values ​​for the current 8 cycles and the next 8 cycles are output. When making predictions, we assume that the RUL predicted values ​​for the first 8 cycles predicted by the model are accurate because there are real multivariate sensor time series signals corresponding to them, while the RUL predicted values ​​for the next 8 cycles are uncertain. Therefore, it is advisable to design a high confidence level for the first 8 cycles and a linearly weakened confidence level for the last 8 cycles. The confidence weight vector β = [1, 1, 1, 1, 1, 1, 1, 1, 0.9, 0.8, 0.7, 0.6, 0.5, 0.4, 0.3] can be designed to optimize the RUL predicted value.

[0174] In order to measure the effectiveness of the designed method, performance indicators such as root mean square (RMSE), mean square error (MSE), dynamic time warping (DTW), NASA scoring function (S) and PHM-2021 scoring function (score) are used to verify the performance of the method. The evaluation index formula is as follows:

[0175]

[0176]

[0177]

[0178]

[0179] score=0.5×RMSE+0.5×S (20)

[0180] Where Q is the total number of flight cycles over the life cycle of the aircraft engine in the selected test cell; and y t Represent the predicted value and true value of RUL respectively; and Y represent the predicted and actual RUL sequences of an engine respectively; K represents the warping path length of DTW; w k′ Indicates the k′th step in the deformation path and y t The distance between them; α represents the penalty coefficient, which is 1 / 13 if the RUL is underestimated, and 1 / 10 otherwise.

[0181] To verify the effectiveness of the method, experiments were conducted on all nine datasets and compared with traditional deep learning methods.

[0182] In order to prove the superiority of the method proposed in the present invention, the N-CMAPSS dataset was used as the experimental object, and compared with existing common deep learning methods such as deep convolutional neural network DCNN, residual neural network ResNet, long short-term memory network LSTM, bidirectional long short-term memory network Bi-LSTM, ProgNet and Inception DCNN. In order to ensure the fairness of the comparison, the above methods all performed degradation trend enhancement processing on the multi-sensor time series signal in the data preprocessing stage. The difference is that the above methods process the multi-sensor time series signal into highly regularized data, while the method proposed in the present invention models the data as a non-Euclidean graph structure. The experimental results are as follows: Figure 6 As shown, from left to right, the first column is the prediction result graph of each method under the DS01 sub-dataset unit8 test unit, the second column is the prediction result graph of each method under the DS03 sub-dataset unit10 test unit, and the third column is the prediction result graph of each method under the DS08a sub-dataset unit12 test unit; from top to bottom, the first row is the prediction result graph based on the DCNN method, the second row is the prediction result graph based on the ResNet method, the third row is the prediction result graph based on the LSTM method, the fourth row is the prediction result graph based on the Bi-LSTM method, and the fifth row is the prediction result graph based on the method of the present invention. It can be seen that the RUL curve predicted by the method of the present invention is closer to the true RUL curve and has better smoothness. In particular, under the DS08a sub-dataset unit12 test unit, the comparison method cannot even complete the RUL prediction task, and the later prediction value directly becomes a constant. In contrast, the methods proposed in this article can complete the RUL prediction task and can portray the RUL change trend of the aircraft engine, indicating that the method of the present invention has better prediction ability. Combined with the performance index data in Table 2 below, it can be seen that in the RMSE, MSE, S, score indicators, the present invention has lower values ​​than other methods, indicating that the method of the present invention has higher prediction accuracy; in the DTW indicator, the present invention also has lower values ​​than other methods, indicating that the RUL prediction curve has smoother performance and the prediction results are more robust. In summary, the method proposed by the present invention is superior to other methods, with higher prediction accuracy, better smoothness and robustness. Moreover, Figure 7 As shown in the figure, the proposed method can also perform attribution analysis to understand the main basis for the model to derive the RUL value. If a certain attention coefficient accounts for a large proportion, it means that the corresponding sensor timing signal is the main factor affecting the current RUL prediction value, which is convenient for engineers to conduct further analysis and judgment, and can provide a certain degree of decision support for predictive maintenance, which is more applicable to engineering applications.

[0183] Table 2 Statistics of performance indicators under different prediction methods

[0184]

Claims

1. A method for predicting the remaining life of an aircraft engine based on multi-level graph feature fusion, characterized in that: The steps include: Step 1: Acquire an aircraft engine status dataset; the aircraft engine status dataset includes operating condition signals, multi-sensor timing signals, virtual sensor signals, engine health parameters, RUL labels, engine unit numbers, flight cycles, flight categories, and health status labels; Step 2: Divide the aircraft engine status dataset into a training set and a test set; Step 3: Preprocess the working condition signals, multi-sensor time series signals and RUL labels in the training set and test set; Step 4: Build a multi-level graph feature fusion model; The multi-level graph feature fusion model includes a node-level feature extraction model, a subgraph-level feature extraction model, a cycle-level feature extraction model, a time-level feature extraction model and an output layer in sequence; Step 5: Use the training set to train the multi-level graph feature fusion model for a set number of rounds to obtain a trained multi-level graph feature fusion model; Step 6: Input the test set into the multi-level graph feature fusion model to obtain the aircraft engine RUL prediction sequence; Step 7: Smooth the RUL prediction sequence predicted by the multi-level graph feature fusion model to obtain the RUL smooth prediction sequence in real time.

2. The method for predicting the remaining life of an aircraft engine based on multi-level graph feature fusion according to claim 1, characterized in that: The step 3 includes the following specific steps: Step 3.1: Perform maximum and minimum value normalization operations on the working condition signals and the multi-sensor time series signals to obtain a number of normalized working condition signals and multi-sensor time series signals; Step 3.2: Sort the values ​​in each normalized multi-sensor time series signal to obtain a ranking value of each value, and form a ranked sequence based on the ranking values ​​of each value, thereby obtaining several ranked sequences from the multiple multi-sensor time series signals; Step 3.3: Based on the rank sequence, use the Spearman correlation analysis method to calculate the normalized multivariate sensor time series signal x s The Spearman correlation coefficient between any two multivariate sensor time series signals is calculated, and a sensor network topology graph is constructed according to the Spearman correlation coefficient; the sensor network topology graph is composed of a plurality of sensor nodes representing sensors and edges connecting them; The method for constructing a sensor network topology graph is specifically as follows: if the Spearman correlation coefficient between two multi-sensor time series signals is greater than or equal to a set threshold, then there exists an edge between the sensor nodes corresponding to the two multi-sensor time series signals, otherwise there does not exist an edge; The Spearman correlation coefficient calculation formula is: Where n represents the sample size in the rank sequence; S' 1,index is the index-th value in the rank sequence S'1; S' 2,index is the index-th value in the rank sequence S2; ρ s is the Spearman correlation coefficient; Step 3.4: Perform degradation enhancement processing on the multivariate sensor time series signal to obtain a multivariate residual time series signal; Step 3.5: Based on the sensor network topology, a graph signal is constructed for the multivariate residual time series signal to obtain a graph sequence of the entire life cycle of the aircraft engine; the graph signal includes three elements: topology, node features, and edge features; The method for obtaining a graph sequence specifically comprises: using a time sliding window method to intercept a multivariate residual time series signal within a flight cycle according to a set window size, using the intercepted time series signal as a node feature, using the edge weights in the adjacency matrix corresponding to the sensor network topology map as edge features, constructing a graph signal from the sensor network topology map, the node features, and the edge features to obtain a subgraph, then placing multiple subgraphs within a flight cycle into a set to obtain a subgraph sequence of a flight cycle, and further obtaining a graph sequence of the entire life cycle of an aircraft engine composed of subgraph sequences of multiple cycles; Step 3.6: Perform nonlinear processing on the original RUL label to define the RUL label; The RUL label t Defined as: Among them, y t is the RUL label value of flight period t; t EOL Indicates the time when the engine fails and cannot be used; FPT indicates the time when the engine begins to degrade; RUL t Represents the original RUL label value of the current flight cycle t in the aircraft engine status dataset.

3. The method for predicting the remaining life of an aircraft engine based on multi-level graph feature fusion according to claim 1, characterized in that: The step 4 includes the following specific steps: Step 4.1: Build a node-level feature extraction model; The node-level feature extraction model inputs a graph sequence including a subgraph sequence of m flight cycles of an aircraft engine, performs node-level feature extraction processing, and outputs a graph sequence including a subgraph sequence of m flight cycles of an aircraft engine that has undergone node-level feature extraction processing; The node-level feature extraction model is composed of a number of GNN-ResNet blocks that are set in parallel and share parameters. The number of the GNN-ResNet blocks is equal to the number m of flight cycles contained in the input graph sequence. The GNN-ResNet block includes a first GATv2 layer, a first LayerNorm layer, a first nonlinear activation layer, a second GATv2 layer, a second LayerNorm layer, a first FNN layer and a second nonlinear activation layer. Each GNN-ResNet block aggregates the node features of the subgraph sequence of a flight cycle in units of subgraphs through the first GATv2 layer, and uses the first LayerNorm layer to perform node feature aggregation on the first GATv2 layer. The aggregated node feature vector group is normalized, and then the normalized node feature vector group is nonlinearly mapped through the first nonlinear activation layer. The node features are further aggregated using the second GATv2 layer, and the aggregated node feature vector group is normalized using the second LayerNorm layer. The first FNN layer is used to transform the dimension of the input subgraph sequence of the flight cycle and perform feature fusion with the normalized node feature vector group output by the second LayerNorm layer. The fused node feature vector group is then nonlinearly activated using the second nonlinear activation layer to obtain a subgraph sequence of a path that has been feature extracted by the GNN-ResNet block. The first nonlinear activation layer is a ReLu activation function; the second nonlinear activation layer is a Tanh function; Step 4.2: Build a sub-graph-level feature extraction model; The subgraph-level feature extraction model inputs a graph sequence of a subgraph sequence of m flight cycles of an aircraft engine that has undergone node-level feature extraction processing, performs subgraph-level feature extraction processing, and outputs a subgraph-level feature sequence set consisting of subgraph-level feature sequences of m flight cycles; the subgraph-level feature extraction model is composed of a plurality of parallel and parameter-shared self-attention aggregation layers with temperature, the number of parallel paths of the self-attention aggregation layer being equal to the number m of flight cycles contained in the input graph sequence; the self-attention aggregation layer performs attention calculation on the input subgraph sequence through a self-attention mechanism with temperature to obtain a subgraph-level feature sequence; the self-attention mechanism with temperature performs similarity calculation on the nodes in each subgraph in the subgraph sequence output by the node-level feature extraction model, calculates the similarity between any node and other nodes, and normalizes these similarities into attention weights, wherein a temperature parameter is added to adjust the calculated value of the attention weight, increasing the value of the attention weight. Then, by weighted summing each node feature with the corresponding attention weight, the subgraph-level feature h output by the self-attention mechanism can be obtained. k,t , integrating sub-graph level features into sub-graph level feature sequences; Step 4.3: Build a cycle-level feature extraction model; The cycle-level feature extraction model inputs a sub-graph-level feature sequence set consisting of sub-graph-level feature sequences of m flight cycles, performs cycle-level feature extraction processing, and outputs a cycle-level feature sequence integrated from the m cycle-level features; the cycle-level feature extraction model is composed of a plurality of parallel and parameter-shared self-attention aggregation layers with temperature, the number of parallel self-attention aggregation layers being equal to the number m of flight cycles contained in the input graph sequence, the input of any self-attention aggregation layer being a sub-graph-level feature sequence of a flight cycle output by the sub-graph-level feature extraction model, and the output is the cycle-level features of the sub-graph-level feature sequence; Step 4.4: Build a time-level feature extraction model; The time-level feature extraction model includes a first Bi-LSTM layer, a second Bi-LSTM layer, and a third nonlinear activation layer in sequence. The input is a period-level feature sequence, and the output is a time-level degradation feature extracted from the time-level feature. The third nonlinear activation layer is a Tanh function; Step 4.5: Design the output layer; The output layer includes a feature expansion layer, a second FNN layer, a third FNN layer, a fourth FNN layer, a fourth nonlinear activation layer, and a fifth nonlinear activation layer in sequence. The input is the time-level degradation feature output by the time-level feature extraction model, and the output is the RUL prediction value of the current m periods and the RUL prediction value of the next m periods; the fourth nonlinear activation layer is a Tanh function, and the fifth nonlinear activation layer is a ReLu function.

4. The method for predicting the remaining life of an aircraft engine based on multi-level graph feature fusion according to claim 1, characterized in that: The step 6 includes the following specific steps: Step 6.1: Input the graph sequence into the node-level feature extraction model to extract node-level features, and obtain a graph sequence after node-level feature extraction; Step 6.2: Input the graph sequence after node-level feature extraction into the sub-graph-level feature extraction model for sub-graph-level feature extraction to obtain a sub-graph-level feature sequence set; Step 6.3: Input the subgraph-level feature sequence set into the period-level feature extraction model to extract period-level features and obtain a period-level feature sequence; Step 6.4: Input the period-level feature sequence into the time-level feature extraction model for time-level feature extraction, and then input it into the output layer to obtain the aircraft engine RUL prediction sequence.

5. The method for predicting the remaining life of an aircraft engine based on multi-level graph feature fusion according to claim 2, characterized in that: The step 3.4 includes the following specific steps: Step 3.4.1: Use the healthy operating condition signal in the operating condition signal as input and the healthy multi-sensor time series signal in the multi-sensor time series signal as output, and use a three-layer fully connected neural network (FNN) to fit the aircraft engine input-output model in the healthy state; Step 3.4.2: Input any operating condition signal into the aircraft engine input-output model in a healthy state to obtain the corresponding ideal multi-sensor time series signal in a healthy state; Step 3.4.3: Calculate the residual between the ideal multi-sensor timing signal in the healthy state and the real multi-sensor timing signal in the aircraft engine status dataset, and perform SG smoothing and denoising processing on a flight cycle basis to obtain the multi-sensor residual timing signal.

6. The method for predicting remaining life of an aircraft engine based on multi-level graph feature fusion according to claim 1, characterized in that: The smoothing method in step 7 is: in, is the smoothed RUL prediction value of flight cycle t, is the RUL prediction value of the flight cycle t at the i′th position in the current RUL prediction sequence, β i′ is the weight value corresponding to the i′th position of the weight vector.

7. The method for predicting the remaining life of an aircraft engine based on multi-level graph feature fusion according to claim 4, characterized in that: The expression for node-level feature extraction in step 6.1 is: Among them, α i,j represents the attention coefficient between node i and node j, Represents the node features of the current layer of node i, represents the node features of the next layer of node i, Represents the node features of the current layer of node j, and represents the learnable parameters, represents the set of neighbor nodes of node i, j represents the node number, and the value of j can be equal to i; [x i ||x j ] represents node features and LeakyReLU(·) represents the activation function.

8. The method for predicting the remaining life of an aircraft engine based on multi-level graph feature fusion according to claim 4, characterized in that: The expression for sub-graph-level feature extraction in step 6.2 is: Among them, h k,t represents the subgraph-level features of the kth subgraph in flight cycle t; represents the attention coefficient of the i-th node in the k-th subgraph in flight cycle t; represents the features of the i-th node of the k-th subgraph in flight cycle t; represents the features of the jth node of the kth subgraph in flight cycle t; represents the similarity measure between the i-th node and the j-th node of the k-th subgraph in flight cycle t; d represents the node feature dimension; N represents the number of nodes, i, j represent the node numbers; τ represents the temperature coefficient.

9. The method for predicting the remaining life of an aircraft engine based on multi-level graph feature fusion according to claim 4, characterized in that: The cycle-level feature extraction expression in step 6.3 is: Among them, h t represents the periodic characteristics of the flight period t; α k,t represents the attention coefficient of the kth sub-graph in flight cycle t. The larger the coefficient, the higher the attention of the corresponding sub-graph, that is, the stronger the relationship with the degradation feature; represents the similarity measure between the kth subgraph and the lth subgraph in period t; d′ represents the period feature h k,t dimension; M represents the number of subgraphs in the flight cycle t, k and l represent the subgraph numbers in a flight cycle; τ represents the temperature coefficient.

Images