Method for predicting residual service life of aero-engine based on fine-grained space-time interaction graph neural network

Abstract

The invention belongs to the technical field of aero-engine residual life prediction, and relates to an aero-engine residual life prediction method based on a fine-grained space-time interaction graph neural network. According to the method, a space-time interaction diagram component is realized through a self-adaptive patch segmentation and space-time interaction diagram construction method, and fine-grained modeling of the dependency relationship is realized. Firstly, adaptive patch segmentation is realized by introducing an offset network and a weight network, patch division matched with gas path characteristics is dynamically generated, mode and semantic features in an aero-engine multi-element sensor sequence can be more comprehensively extracted, and different working conditions are adapted. When the space-time interaction diagram is constructed, values of all sensor time steps are mapped to nodes on the space-time interaction diagram, and a probability model is used for learning a diagram structure, so that capturing of a space-time dependency relationship is facilitated. Through comparison on an open source data set, the result shows that the method effectively improves the prediction quality of the residual life of the aero-engine, and is of great significance for guaranteeing the flight safety.

Description

Technical Field

[0001] The invention belongs to the technical field of aircraft engine remaining service life prediction, and relates to an aircraft engine remaining service life prediction method based on a fine-grained spatiotemporal interactive graph neural network. Background Art

[0002] Aircraft engines are complex thermal machines that operate continuously under harsh conditions of high speed, heavy load and high temperature. The prediction of the remaining service life of aircraft engines is of great significance. It can predict the remaining operating time of equipment or systems in future operations, thereby guiding the maintenance and decision-making process. Accurate remaining service life prediction can significantly reduce system maintenance costs and reduce unplanned downtime. Deep learning methods have been widely used in the remaining service life prediction of aircraft engines. Patent CN118936883A discloses a remaining service life prediction method, which first uses a multi-head attention mechanism to capture important key features, and then extracts multi-scale features from the data reconstructed by the multi-head attention mechanism using a multi-scale temporal convolutional network. Patent CN115952724A adds the position encoding information vector encoding information of sensor data to the Transformer model, so that the model has the ability to model and generalize time information. And introduce multi-head probabilistic sparse self-attention to improve the feature extraction ability. However, the operating status of aircraft engines is usually monitored by multiple sensors. The data captured by these sensors not only have temporal dependence, but also have complex correlations in space, such as the interaction between different physical quantities such as temperature and pressure. The above methods do not consider the interaction between sensors, resulting in incomplete modeling and missing information.

[0003] Thanks to the powerful relationship capture function, graph neural network-based methods have become a popular solution for modeling spatiotemporal dependencies in multivariate time series data. Graph neural networks are used as spatial encoders and temporal encoders such as long short-term memory networks to capture spatiotemporal dependencies. For example, patent CN202410470446.4 discloses a remaining life prediction method using graph convolution and attention mechanism, which uses GCN to extract the spatial features of sensor sequences and adopts self-attention mechanism network to extract its temporal degradation trend. The method following the above paradigm takes into account both temporal and spatial dependencies, but potential limitations still exist:

[0004] 1) The influence pattern of sensors is usually regarded as an inherent characteristic of the system, and the node connections on the time snapshot level graph are often predetermined and fixed based on prior knowledge. However, in the operation of aircraft engines, the influence pattern of sensors is affected by the dynamic changes of multiple factors (such as working environment, load conditions, etc.), and it is difficult to fully describe its complex correlation with a fixed graph structure.

[0005] 2) Current methods mostly model data relatively independently at the time snapshot level, which makes it difficult to capture dependencies across time steps. For example, the temperature and pressure sensors in an engine are related to each other, and because the engine system has inertia and dynamic response time, historical temperature sensor readings may have a significant impact on future changes in the pressure sensor. This cross-time step and cross-sensor interaction is difficult to fully represent using existing modeling methods.

[0006] In summary, it is difficult for existing methods to fully capture the spatiotemporal dependencies in aeroengine sensor sequences. To address these limitations, the method disclosed in the present invention can not only cope with the dynamic changes of sensor data in aeroengines, but also effectively handle dependencies across time steps, ensuring more accurate remaining life prediction. Summary of the invention

[0007] In order to solve the above problems, the present invention provides a method for predicting the remaining useful life of an aircraft engine based on a fine-grained spatiotemporal interactive graph neural network.

[0008] The technical solution of the present invention is as follows:

[0009] A method for predicting the remaining useful life of an aircraft engine based on a fine-grained spatiotemporal interactive graph neural network, the steps are as follows:

[0010] The first step is to preprocess the aircraft engine sensor data.

[0011] In the second step, pattern recognition is performed on the multivariate time series of aircraft engine sensors and adaptively segmented into semantically complete patches.

[0012] The third step is to build a spatiotemporal interaction graph based on the adaptively segmented patches to better capture the spatiotemporal dependencies.

[0013] The fourth step is to fuse the embeddings of multiple spatiotemporal interaction graphs and realize the remaining life prediction of aircraft engines.

[0014] More specifically, the second step process is as follows:

[0015] When segmenting a multi-sensor sequence of aero-engine gas path, the interaction between sensors needs to be considered simultaneously to ensure that the segmented patches can reflect global information. Adaptive patch segmentation can flexibly adjust the size and position of patches according to the pattern changes in the sensor sequence by dynamically adjusting the center position and boundaries of each patch. Each patch not only has a center point, but also contains offsets to the center, left boundary, and right boundary to more accurately capture the local pattern. In this way, the size of the patch can be dynamically adjusted as the time series changes to ensure that each patch can cover the most important spatiotemporal information. Adaptive patch segmentation is achieved through two key components: the offset network and the weight network. The offset network is responsible for determining the optimal position of each patch to capture the complete pattern and semantic information in the sequence. This process is achieved by analyzing the internal structure and pattern of the sequence to infer the optimal location of the patch. The weight network further optimizes the importance distribution of each sensor dimension at different time points, so that the generated patches can fully reflect the important features under different working conditions. This dynamic segmentation strategy based on global information helps to more efficiently extract key patterns reflecting performance degradation from aero-engine data.

[0016] The third step process is as follows:

[0017] First, the multivariate time series is mapped to nodes on the spatiotemporal interaction graph through the joint action of spatiotemporal and patch embedding. Through graph structure learning, the model can identify and model the spatial dependencies between different sensors and the temporal dependencies of the same sensor at different time points. In this way, the model can not only learn the dynamic evolution of sequence data, but also capture the potential spatiotemporal dependencies between different time steps. In order to improve the flexibility of the model, the edge weights of the graph are not directly learned, but the interactions between nodes are regarded as the results of sampling from the probability distribution. This probabilistic modeling method can better control the characteristics of the graph structure. The connection between nodes is achieved by learning the probability of their interaction. Specifically, it is assumed that the connection probability between each pair of nodes is calculated based on their features, and a probabilistic network is used to quantify these interactions. The probability-based edge modeling can capture the dynamic impact of sensors under different working conditions, load states and other environmental changes. At the same time, the optimization process of the model is guided by the graph structure loss, so that it can more accurately capture the spatiotemporal dependencies in the sequence.

[0018] The fourth step process is as follows:

[0019] After constructing the spatiotemporal interaction graph, the message passing neural network is used to embed features on the graph. In the gas path prediction task, the status of each key component of the engine is collected by multiple sensors (pressure, temperature, flow, etc.). The data of each sensor has dynamic changes in the time dimension and is also affected by the interaction of adjacent sensors in the space dimension. Sensor nodes can absorb the status information of surrounding neighbor sensors through the message passing mechanism, which not only helps to capture the time delay effect, but also reveals the direct or indirect interaction between sensors. By comprehensively learning spatiotemporal features and realizing efficient fusion of information in different time and sensor dimensions. Finally, by fusing the features of multiple spatiotemporal patches, a more accurate prediction of the remaining service life of aircraft engines can be achieved.

[0020] The technical concept of the present invention is: in view of the fact that the traditional fixed-size patch division method often cannot accurately capture the dynamic changes and local features in the multi-sensor sequence, the present invention proposes an adaptive patch segmentation method to ensure the semantic integrity of key modes such as degradation. The method flexibly determines the center position and boundary of each patch through the dynamic adjustment of the offset network and the weight network, and can adaptively adjust the size and position of the patch according to the actual working conditions, so that the patch division can accurately reflect the key fault characteristics and trend changes during the operation of the aircraft engine, and improve the ability to predict the remaining life of the aircraft engine.

[0021] Secondly, in response to the challenge of modeling complex spatiotemporal dependencies in multivariate sensor sequences, the present invention introduces a spatiotemporal and patch embedding mechanism to map multivariate aircraft engine sensor sequences into nodes of a spatiotemporal interaction graph. In order to overcome the shortcomings of traditional graph structure learning methods in dealing with complex dependencies, the present invention adopts a graph structure learning method based on probability distribution to calculate the probability of interaction between nodes through a probabilistic network. This method can maintain the sparsity of the graph structure while flexibly modeling the complex interaction relationships between nodes, effectively capturing the temporal dependencies between different time points in the same sensor, and the spatial dependencies between different sensors, thereby improving the accurate prediction of the remaining life of aircraft engines.

[0022] Beneficial effects of the present invention:

[0023] This method realizes the spatiotemporal interaction graph component through adaptive patch segmentation and spatiotemporal interaction graph construction methods, and realizes fine-grained modeling of dependencies. First, adaptive patch segmentation is realized by introducing offset networks and weight networks, and patch partitions that match the gas path characteristics are dynamically generated, which can more comprehensively extract patterns and semantic features in the multi-sensor sequence of aircraft engines and adapt to changes under different working conditions. When constructing the spatiotemporal interaction graph, the values ​​of all sensor time steps are mapped to nodes on the spatiotemporal interaction graph, and a probabilistic model is used to learn the graph structure, which helps to capture spatiotemporal dependencies. Through comprehensive comparison on open source datasets, the results show that the proposed method effectively improves the quality of aircraft engine remaining life prediction, which is of great significance to ensuring flight safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a block diagram of fine-grained spatiotemporal interaction network for aircraft engine remaining life prediction;

[0025] Figure 2 It is a schematic diagram of the fine-grained spatiotemporal interaction network used for the prediction of the remaining life of aircraft engines;

[0026] Figure 3 It is a flow chart of fine-grained spatiotemporal interaction network for aircraft engine remaining life prediction;

[0027] Figure 4 is a schematic diagram of the adaptive patch segmentation module;

[0028] Figure 5 It is a schematic diagram of the transformation of multivariate time series to spatiotemporal interaction diagram;

[0029] Figure 6 It is a visualization diagram of the present invention on a public dataset. DETAILED DESCRIPTION

[0030] The present invention is further described below in conjunction with the accompanying drawings and implementation modes.

[0031] like Figure 1 and Figure 2 , a fine-grained spatiotemporal interactive graph neural network is used for aircraft engine remaining useful life prediction, including the following:

[0032] For the multivariate time series of aircraft engine sensors, data preprocessing is first performed, including data cleaning and feature selection, to ensure the quality and validity of the input data. Then, adaptive block partitioning is given based on the patch segmentation network to ensure that the captured pattern is more continuous and complete. Subsequently, the spatiotemporal interaction graph construction is performed for each patch, which includes two steps. First, the spatiotemporal embedding of nodes and patch embedding are performed to prevent the loss of key spatiotemporal information in the process from sensor sequence to spatiotemporal interaction graph. Secondly, the spatiotemporal dependency of sensors is modeled through Bernoulli distribution to effectively capture the spatiotemporal dependency. After the spatiotemporal interaction graph is constructed, message passing is performed on the graph to generate a vector representation of the multivariate time series, and then the prediction of the remaining useful life of the aircraft engine is obtained through the fully connected layer. The entire framework combines the difference between the predicted and actual RUL through the Huber loss function, and introduces the graph structure loss to optimize the spatiotemporal relationship between nodes in the graph, further improving the learning effect of the model. Finally, based on the RUL prediction results, optimization suggestions for maintenance plans, maintenance cycles, and operation strategies are provided to help airlines improve the scientificity of decision-making and the safety and reliability of aircraft engines.

[0033] like Figure 3 As shown, it can be implemented in the following steps:

[0034] S101: Sensor Data Collection

[0035] Multi-sensor data collection covers operation data in different time and space dimensions. For aircraft engine health status monitoring, the collected sensor data includes but is not limited to temperature, pressure, speed, etc.

[0036] S102: Data preprocessing

[0037] Preprocess the raw data, select important features, and normalize the data.

[0038] S103: Adaptive Patch Segmentation

[0039] Adaptive patch segmentation algorithm is used to adaptively divide the multi-sensor sequence through offset and weight network. This module can flexibly capture the changing trend of local patterns by dynamically adjusting the size and weight of the patch. Adaptive segmentation not only adapts to the changes in the time dimension, but also optimizes the learning process of local patterns by combining spatial information.

[0040] S104: Construction of spatiotemporal interaction graph

[0041] Based on multivariate sensor data, a spatiotemporal interaction graph is constructed. The nodes in the graph represent the states of different sensors at each time step. The edge weights between nodes are dynamically calculated through a probabilistic neural network to quantitatively characterize the spatiotemporal correlation strength between different sensor nodes. Furthermore, the edges are sampled from a probabilistic perspective, and the graph structure is optimized by selecting edges with high correlation.

[0042] S105: Model Training

[0043] During the model training phase, graph structure loss is introduced to maintain the sparsity of the graph. By optimizing the graph structure loss, the model can not only better learn the spatiotemporal dependencies between sensor nodes, but also improve the rationality of node connections and reduce the interference of invalid connections.

[0044] S106: Remaining life prediction

[0045] Based on the trained spatiotemporal interaction graph model, the remaining life of aircraft engines is predicted. By extracting the spatiotemporal dependencies in multivariate sensor data, the model can accurately predict the remaining life of the engine under multiple operating conditions.

[0046] S107: Decision Support

[0047] Based on the remaining life prediction results, decision support is provided for the maintenance and operation of aircraft engines, maintenance plans are optimized, and flight safety is improved. By combining the model prediction results with the existing maintenance strategy, maintenance timing and tasks can be dynamically optimized.

[0048] Specifically, Figure 4-5 As shown, a method for predicting the remaining useful life of an aircraft engine based on a fine-grained spatiotemporal interactive graph neural network of the present invention comprises the following steps:

[0049] The first step is to preprocess the aircraft engine sensor data. The process is as follows:

[0050] The original aircraft engine sensor data is subjected to feature selection based on mutual information, and the features that best reflect the engine health status are selected to reduce the data dimension and improve the learning efficiency of the model. The input data is then normalized to stabilize the training.

[0051] In the second step, the aircraft engine sensor sequence is segmented into semantically complete patches to achieve pattern recognition. The process is as follows:

[0052] In the prediction of the remaining life of aircraft engines, the multivariate time series data collected by sensors are characterized by high dimensionality, complexity, and non-stationarity. The value of a single time step is difficult to convey complete semantic information, especially when capturing complex dependencies. Therefore, dividing the time series into patches and extracting features from them is an effective strategy to improve the model's ability to capture key patterns. Through adaptive segmentation, not only is the local pattern capture in the sequence optimized, but the interaction between sensors is also considered, realizing pattern recognition of multivariate sensor sequences.

[0053] The adaptive patch segmentation module first divides the sequence into an initialization patch sequence, then calculates the patch offset position through the patch segmentation network, adjusts the position of the initialization patch based on the offset position, and finally obtains the adaptive patch segmentation. Specifically, X=[x1,x2,…,x T ] is the aircraft engine sensor sequence, where T is the input sequence length, N is the number of sensors. The sequence is first divided into a number of patches P = [P 1 ,P 2 ,...,P u ],in D is the patch radius, S is the step size of patch division. Adaptive patch P u You can use the quintuple (c u ,D,δc u ,δl u ,δr u ) indicates that, where c u is the center position of the patch, δc u is the center offset, δl u is the left boundary offset, δr u is the right boundary offset. Patch P u The starting position l u and the end position r u It can be expressed by the following formula:

[0054] l u =(c u +δc u )-(D+δl u )

[0055] r u =(c u +δc u )+(D+δr u )

[0056] It should be noted that the offset of the center point, left boundary and right boundary will cause the patch size to change. In order to keep the size of each patch as 2D, the initialization patch should be sampled and bilinearly interpolated based on the new position boundary to obtain the final adaptive patch sequence Implements pattern recognition, where U is the number of patches.

[0057] like Figure 4 As shown in the figure, the patch segmentation network consists of an offset network and a weight network, which is used to learn the offset δ = (δc, δl, δr). The offset network is responsible for inferring the position of each patch in the sequence with a more complete pattern and semantics. It is achieved by determining the offset of each patch through the structure and pattern in the multi-sensor sequence. The weight network is responsible for adjusting the importance of different positions in the sequence. The weight network further optimizes the importance distribution of each sensor dimension at different time points, so that the generated patches can fully reflect the important features under different working conditions. Given an input First, perform a feature embedding operation on the input and convert it into a matrix Where K represents the embedding dimension. Next, the specific steps for calculating the global offset are as follows:

[0058] gate=sigmoid(W gt ·f(X em ))

[0059] offset = W of (W pj ·f(X em ))

[0060]

[0061] Where gate represents the gate value calculated by the weight network, offset represents the offset calculated by the offset network, f(·) represents the feature extractor, which is implemented by 1D-CNN, and W gt The weights used to learn different sensor sequences, W pj and W of Used to learn the offset of each patch. δc u Indicates the center point offset. Selecting the Tanh activation function ensures that the displacement is bidirectional. Make sure the offset is within the range , A represents the scaling factor of the center offset. u and δr u ReLU activation function and Ensure that the offset range is within , B represents the scaling factor of the boundary offset. Adaptive patch segmentation effectively captures complex patterns in multivariate sensor data by dynamically optimizing the location and range of patches. For example, when an aircraft engine is in a non-steady-state condition, patch segmentation can be flexibly adjusted to ensure that the changing trends of key patterns (such as temperature and pressure signals) are captured.

[0062] The third step is to construct the spatiotemporal interaction graph. The process is as follows:

[0063] Based on the second step For each patch Construct a spatiotemporal interaction graph. Specifically, map the data from each sensor at each time step in the patch to the nodes on the graph, that is, there are 2*N*D nodes in the graph, and the feature dimension of each node is K, such as Figure 5 This graph is called a space-time interaction graph, and the space-time interaction graph is represented as in, represents the set of nodes on the graph, ε u represents the edge set on the graph. However, due to the unordered nature of the graph, placing all nodes in a patch on the graph may result in the loss of the original sequence information. To solve this problem, the present invention introduces a learnable temporal embedding matrix in the nodes of the same sensor at different time steps on the spatiotemporal interaction graph. Such embedding can give each node the temporal information of the sensor sequence, which helps to preserve the temporal structure characteristics of the original sequence. In addition, considering that different sensor data have different spatial characteristics, a learnable spatial embedding matrix is ​​further introduced for nodes of different sensors at the same time step. In this way, the data characteristics of different sensors are fully utilized. In addition, considering that the multivariate time series is cut into patches, a learnable patch position embedding matrix is ​​introduced for the nodes in the same patch. To preserve the sequence information of the patch. At the same time, self-attention is calculated to integrate different embedding information to obtain the node feature X tsp For subsequent calculations, as follows:

[0064] X tsp =Att(X org +E tem +E spl +E patch )

[0065] X-based tsp The node set of each spatiotemporal interaction graph can be obtained r = 2*N*D. After mapping the aircraft engine sensor data sequence to the nodes of the spatiotemporal interaction graph, the interaction relationship between the nodes, i.e., the edge set ε, is inferred through structural learning. u The goal of structural learning is to effectively capture the temporal dependencies between the same sensor at different time steps and the spatial dependencies between different sensors at the same time step. By modeling node interactions, the dynamic evolution of sequence data can be learned, and the fine-grained spatiotemporal dependencies between different sensors at different time steps can be captured. Using probability distributions to model the interactions between nodes provides greater flexibility for structural learning and allows for better control over the properties of the graph. Edge Set ε u The probability distribution is based on the node characteristics Learned. Assume that the edge between node pairs (i, j) follows a Bernoulli distribution with parameter θ ij ∈(0,1), where i,j∈(1,r). For the adjacency matrix A u ~Ber(θ ij ), the Gumbel-softmax method is used to sample the discrete adjacency matrix. By applying the Gumbel reparameterization method, the differentiability of the sampling process is achieved, as shown in the following formula:

[0066]

[0067]

[0068] in, represents the adjacency matrix of the u-th spatiotemporal interaction graph, and is an iid sample vector drawn from the Gumbel distribution, σ represents the S-type activation function, and s is a parameter that controls the smoothness of the sample. At this point, the problem is transformed into finding the probability θ that there is an edge between nodes (i, j) ij .

[0069] Node set based on the u-th spatiotemporal interaction graph Starting from the beginning, the interaction between node features is quantified using a probabilistic network g(·), which is used to capture the mutual influence of sensor data at different time points and spatial locations. g(·) is implemented by a single-layer network, and its parameters can be expressed as Node interactions are considered as connection probabilities, which can be expressed as follows:

[0070]

[0071] where |·| represents the absolute value of the difference between corresponding elements, and σ is an S-type activation function, ensuring that θ ij The range is between [0,1]. Through the graph structure loss To ensure Update in the right direction, is used as part of the overall loss function as shown below:

[0072]

[0073] If the node v i ,v j If they are far apart in the feature space, then v i ,v jare not too similar, and encourage θ ij Smaller.

[0074] Step 4: Feature embedding and remaining life prediction of spatiotemporal interaction graph. The process is as follows

[0075] 4.1 Feature Embedding of Spatiotemporal Interaction Graphs

[0076] After constructing the spatiotemporal interaction graph, the message passing network is used to implement The feature embedding process is as follows:

[0077]

[0078] in, represents the aggregated features of the neighboring node information of node i in the l+1th layer, Agg represents the mean operation, represents the features of node j in layer l, is the set of neighbors around node i in one hop. The aggregated messages of the surrounding neighbors are then connected to the current node.

[0079]

[0080] in, represents the characteristics of node i in the l+1th layer after message transmission, W l+1 represents the trainable weight of the l+1th layer, and ∥ represents concatenation along the dimension. Then, the complete graph feature of the last layer can be represented as H u =(h1,h2,...,h i ), where h i Represents the feature of the i-th node in the last layer, i∈V. By performing the readout operation, the entire graph feature vector can be expressed as Perform graph embedding on all patches to obtain the feature representation Z = (Z 1 ,Z 2 ,...,Z u ). Subsequently, the feature vectors of Z are connected and input into the fully connected layer to obtain the predicted remaining life value of the aircraft engine;

[0081]

[0082] in, is the predicted remaining life value of the aircraft engine, W op are the trainable parameters of the output fully connected layer.

[0083] 4.2 Remaining life prediction

[0084] Based on 4.1 The difference between the predicted and actual RUL is combined through the Huber loss function:

[0085]

[0086] The actual Y is the actual RUL, and δ is used to control the smoothness of the loss function and the size of the transition region. Based on the Huber loss, combined with the graph structure loss Optimize the model parameters as shown below:

[0087]

[0088] Where θ represents the trainable parameters of the model and λ is the graph structure loss The weight coefficient of .

[0089] 4.3 Experimental Results Analysis

[0090] The C-MAPSS dataset describes the degradation of aircraft engines and is a standard benchmark for evaluating remaining life prediction methods. The dataset contains four sub-datasets, namely FD001, FD002, FD003, and FD004, each of which is collected under different operating conditions and failure modes.

[0091] Each data trace contains monitoring signals from 21 sensors, capturing parameters such as temperature, pressure, fan speed, and three operating condition parameters including altitude, Mach number, and throttle stick angle. It is worth noting that some sensors show constant values, especially sensor IDs 1, 5, 6, 10, 16, 18, and 19, which do not contribute significantly to the remaining life prediction task. Therefore, these sensors with constant values ​​are discarded to simplify the dataset. The training and testing datasets in C-MAPSS are pre-grouped. Given that different sensors measure different types of signals, all data must be normalized to account for the variation in data amplitude between different channels. Max-min normalization is used to scale the data of each sensor value to the range [0,1].

[0092] To verify the performance of the proposed method, it is compared with the following methods. Traditional methods, LSTM and CNN-SBULSTM. Transformer-based methods, including Informer and Autoformer. Methods under the GNN framework include GCN, STGCN, ASTGNN and GDAGDL. The results of the comparative experiments are listed in the following table.

[0093]

[0094]

[0095] Consider the experimental results of each model on RMSE. The network model of the present invention (ASINet) performs well on FD002, FD003 and FD004, outperforming the second best method by 2.13%, 2.18% and 2.82% respectively. This significant improvement can be attributed to the fine-grained feature embedding of ASINet. Adaptive patch segmentation takes into account local similarities in multivariate time series, ensuring more continuous spatiotemporal patterns and semantic information within patches. In addition, spatiotemporal interaction graph structure learning based on probabilistic models can more accurately capture the spatiotemporal dependencies in sensor sequences, reduce noisy connections and improve model performance. However, ASINet did not achieve the best performance on FD001, which can be attributed to the relative simplicity of the dataset. Unlike FD002 and FD004, which contain multiple operating conditions, or FD003, which contains multiple failure modes, FD001 has only one operating condition and failure mode. Although ASINet scores lower than the best result on FD001, it is still competitive, with a difference of only 0.4% between them. Overall, the proposed method outperforms other methods in remaining lifetime prediction, highlighting the importance of modeling spatiotemporal dependencies in a fine-grained manner.

[0096] Overall, the proposed method outperforms other methods in RLS prediction, highlighting the importance of modeling spatiotemporal dependencies in fine-grained manner. Figure 6 As shown, by visualizing the predicted RUL and the actual RUL on the C-MAPSS dataset, it can be found that there is a high degree of consistency between them, which proves the effectiveness of the present invention.

[0097] In summary, the present invention proposes a spatiotemporal interactive network method based on adaptive patch segmentation and message passing neural network, aiming to solve the problem of difficulty in capturing spatiotemporal dependencies in aircraft engine remaining life prediction. This method uses adaptive patch segmentation technology to flexibly adjust the size and position of the patch according to the dynamic changes of the time series to better capture multi-dimensional spatiotemporal information.

[0098] The present invention uses adaptive patch segmentation technology to divide the multivariate sensor sequence into dynamically adjusted spatiotemporal patches, providing richer information expression for subsequent spatiotemporal feature extraction. Then, the spatiotemporal interactive graph network modeling is adopted to integrate the node and its neighborhood information layer by layer through the message passing mechanism between nodes to capture the complex spatiotemporal dependencies in the sequence. The network dynamically learns the graph structure through the probabilistic modeling method, ensuring the flexibility of the graph and enhancing the ability to capture spatiotemporal features in the sequence data.

[0099] The advantage of this invention is that it solves the problem that traditional methods cannot effectively handle complex spatiotemporal dependencies in multivariate time series based on the perspective of graph neural networks, and significantly improves the accuracy and stability of aircraft engine remaining life prediction. Experimental results show that the proposed method has significant advantages in the field of aircraft engine remaining life prediction, provides important technical support for the reliable operation of aircraft engines, and is of great significance to ensuring flight safety.

Claims

1. A method for predicting the remaining useful life of an aircraft engine based on a fine-grained spatiotemporal interactive graph neural network, characterized in that: Here are the steps: The first step is to pre-process the aircraft engine sensor data; Perform feature selection based on mutual information on the original aircraft engine sensor data to select the features that best reflect the health status of the engine; Then the data is max-min normalized to stabilize the training; The second step is to perform pattern recognition on the multivariate time series of aircraft engine sensors and adaptively segment them into semantically complete patches; The adaptive patch segmentation module first divides the sequence into an initialization patch sequence, then calculates the patch offset position through the patch segmentation network, adjusts the position of the initialization patch based on the offset position, and finally obtains the adaptive patch segmentation; specifically, X = [x1, x2, …, x T ] is the aircraft engine sensor sequence, where T is the input sequence length, N is the number of sensors; the sequence is first divided into a number of patches P = [P 1 ,P 2 ,...,P u ],in D is the patch radius, S is the step size of patch division; adaptive patch P u Use quintuple (c u ,D,δc u ,δl u ,δr u ) indicates that, where c u is the center position of the patch, δc u is the center offset, δl u is the left boundary offset, δr u is the right boundary offset; patch P u The starting position l u and the end position r u It is expressed by the following formula: l u =(c u +δc u )-(D+δl u ) r u =(c u +δc u )+(D+δr u ) In order to keep the size of each patch as 2D, the initialized patches are sampled and bilinearly interpolated based on the new position boundary to obtain the final adaptive patch sequence Implement pattern recognition, where U is the number of patches; The patch segmentation network consists of an offset network and a weight network, which is used to learn the offset δ = (δc, δl, δr); the offset network is responsible for inferring the position of each patch in the sequence with a more complete pattern and semantics, which is achieved by determining the offset of each patch through the structure and pattern in the multi-sensor sequence; the weight network is responsible for adjusting the importance of different positions in the sequence; given an input First, perform a feature embedding operation on the input and convert it into a matrix Where K represents the embedding dimension; Next, the specific steps for calculating the global offset are as follows: gate=sigmoid(W gt ·f(X em )) offset=W of (W pj ·f(X em )) Where gate represents the gate value calculated by the weight network, offset represents the offset calculated by the offset network, f(·) represents the feature extractor, which is implemented by 1D-CNN, and W gt The weights used to learn different sensor sequences, W pj and W of Used to learn the offset of each patch; δc u Indicates the center point offset; the Tanh activation function is selected to ensure that the displacement is bidirectional. Make sure the offset is within the range A represents the scaling factor of the center offset; δl u and δr u Represents the left and right boundary offsets; ReLU activation function and Ensure that the offset range is within , B represents the scaling factor of the boundary offset; The third step is to build a spatiotemporal interaction graph based on adaptively segmented patches to better capture spatiotemporal dependencies; Based on the second step For each patch Construct a spatiotemporal interaction graph; specifically, map the data from each time step of each sensor in the patch to the nodes on the graph, that is, there are 2*N*D nodes in the graph, and the feature dimension of each node is K. This graph is called the spatiotemporal interaction graph, and the spatiotemporal interaction graph is represented as Among them, v u represents the set of nodes on the graph, ε u represents the edge set on the graph; A learnable temporal embedding matrix is ​​introduced between nodes of the same sensor at different time steps on the spatiotemporal interaction graph. Each node is given the temporal information of the sensor sequence; different sensor data have different spatial characteristics, and a learnable spatial embedding matrix is ​​further introduced for nodes of different sensors at the same time step In this way, the data characteristics of different sensors are fully utilized; since the multivariate time series is cut into patches, a learnable patch position embedding matrix is ​​introduced for the nodes in the same patch To preserve the sequence information of the patch; at the same time, self-attention is calculated to integrate different embedding information to obtain the node feature X tsp For subsequent calculations, as follows: X tsp =That(X org +E tem +E spl +E patch ) X-based tsp Get the node set v of each spatiotemporal interaction graph u =(v1,v2,...,v r ), r = 2*N*D; after mapping the aircraft engine sensor data sequence to the nodes of the spatiotemporal interaction graph, the interaction relationship between the nodes, i.e., the edge set ε, is inferred through structural learning. u ; Use probability distribution to model the interactions between nodes; For Edge Set ε u The probability distribution is based on the node feature v u Learned; assume that the edge between node pairs (i, j) follows a Bernoulli distribution with parameter θ ij ∈(0,1), where i,j∈(1,r); for the adjacency matrix A u ~Ber(θ ij ); the Gumbel-softmax method is used to sample the discrete adjacency matrix; by applying the Gumbel reparameterization method, the differentiability of the sampling process is achieved, as shown in the following formula: in, represents the adjacency matrix of the u-th spatiotemporal interaction graph, and is an iid sample vector drawn from the Gumbel distribution, σ represents the S-type activation function, and s is a parameter that controls the smoothness of the sample; at this point, the problem is transformed into finding the probability θ that there is an edge between nodes (i, j) ij ; The node set V based on the u-th spatiotemporal interaction graph u =(v1,v2,...,v r ), the interaction between node features is quantified using a probabilistic network g(·), which is used to capture the mutual influence of sensor data at different time points and spatial locations; g(·) is implemented by a single-layer network, and its parameters are expressed as Node interactions are considered as connection probabilities, which are expressed as follows: where |·| represents the absolute value of the difference between corresponding elements, and σ is an S-type activation function, ensuring that θ ij The range is between [0,1]; through the graph structure loss To ensure Update in the right direction, is used as part of the overall loss function as shown below: Step 4: Fusion the embeddings of multiple spatiotemporal interaction graphs to achieve the remaining life prediction of aircraft engines; 4.1 Feature Embedding of Spatiotemporal Interaction Graphs After constructing the spatiotemporal interaction graph, the message passing network is used to implement The feature embedding process is as follows: in, represents the aggregated features of the neighboring node information of node i in the l+1th layer, Agg represents the mean operation, represents the feature of node j in layer l, V i (1) is the set of neighbors around node i in one hop; the aggregated messages of the surrounding neighbors are then connected to the current node, in, represents the characteristics of node i in the l+1th layer after message transmission, W l+1 represents the trainable weight of the l+1th layer, ∥ represents concatenation along the dimension; then, the complete graph feature of the last layer is represented as H u =(h1,h2,...,h i ), where h i Represents the feature of the i-th node in the last layer, i∈V; by performing the readout operation, the feature vector of the entire graph is represented as Perform graph embedding on all patches to obtain the feature representation Z = (Z 1 ,Z 2 ,...,Z u ); Then, the feature vectors of Z are connected and input into the fully connected layer to obtain the predicted remaining life value of the aircraft engine; in, is the predicted remaining life value of the aircraft engine, W op is the trainable parameter of the output fully connected layer; 4.2 Remaining life prediction Based on 4.1 The difference between the predicted and actual RUL is combined through the Huber loss function: The actual Y is the actual RUL, and δ is used to control the smoothness of the loss function and the size of the transition region; Based on Huber loss, combined with graph structure loss Optimize the model parameters as shown below: Where θ represents the trainable parameters of the model and λ is the graph structure loss The weight coefficient of .

Images