Physical structure guidance-based multi-scale lightweight space-time network and aviation engine residual service life prediction method

Abstract

The invention belongs to the technical field of artificial intelligence, and discloses a multi-scale lightweight space-time network based on physical structure guidance and an aero-engine remaining service life prediction method. The multi-scale lightweight space-time network comprises a multi-scale time sequence feature extraction module, a physical deployment feature classification module, a spatial feature extraction module and an output module which are connected in sequence. The multi-scale feature extraction module performs down-sampling on time sequence data acquired by each sensor for several times, and performs fusion output on obtained time sequence features of different coarsening scales according to a down-sampling inverted sequence; the physical deployment feature classification module embeds the multi-scale time sequence features of each layer into corresponding deployment components; the spatial feature extraction module further extracts the spatial relationship between the components; the output module obtains the remaining service life of the aero-engine. According to the method, different scale features are mutually fused, and the aero-engine service life can be effectively and accurately predicted by introducing the aero-engine key component relationship.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence technology, relates to a multi-scale lightweight space-time network guided by physical structure, and also relates to a method for predicting the remaining service life of an aircraft engine based on the multi-scale lightweight space-time network. Background Art

[0002] Aircraft engines are the core components of aircraft power systems, and their health status is directly related to flight safety. With the continuous development of aircraft engine design technology, the internal structure of the engine has become more sophisticated and complex. Under extreme operating conditions such as high temperature, high speed, and high pressure, the internal subcomponents are prone to performance degradation, and as the operating time increases, the degradation rate accelerates, which may eventually lead to serious failures. If a severely degraded aircraft engine cannot be maintained in time, it may cause a major safety accident. In the field of civil aviation, engine failure not only affects the normal operation of the aircraft, but also leads to high maintenance and overhaul costs. Unplanned maintenance and flight cancellations caused by engine failure will result in additional economic losses, further increasing operating costs. In the military field, the reliability of the engine directly affects the combat readiness of the Air Force and the success rate of the mission. Engine failure has a significant impact on the safety, maintenance cost, and combat readiness of the aircraft, and often requires high operation and maintenance costs.

[0003] In order to improve the maintenance efficiency of aircraft engines, reduce the waste of component resources, reduce maintenance costs and ensure the operating reliability of engines, it is crucial to realize the fault prognostics and health management (PHM) of aircraft engines. Among them, the remaining service life prediction technology, as the core technology in the field of aircraft engine PHM, aims to analyze the sensor data of aircraft engines to predict the remaining operating time of the engine and judge the engine performance degradation state, so as to provide a scientific basis for operation and maintenance decisions.

[0004] In the PHM system, the remaining useful life (RUL) prediction technology is a core component and is particularly critical. The goal of RUL prediction technology is to predict the remaining operating time of the engine by analyzing the sensor data of the aircraft engine during operation, determine its performance degradation state, and provide corresponding maintenance suggestions to avoid serious failures. The accuracy and reliability of this technology are of great significance to extending the service life of the engine, reducing operating costs and ensuring flight safety.

[0005] Existing research on the prediction of the remaining useful life of aircraft engines mainly relies on data-driven deep learning algorithms. Although such algorithms have powerful nonlinear feature extraction capabilities and can improve prediction accuracy, these models usually only focus on the mining and integration of sensor data, lacking an in-depth understanding of the complex internal structure and operating conditions of aircraft engines, which easily leads to the capture of redundant data by model learning, and thus it is difficult to fully and accurately describe the degradation trend of aircraft engines under variable operating conditions, which ultimately limits its effectiveness and reliability in practical applications. In addition, in order to pursue higher prediction accuracy, current deep learning models often increase the complexity and number of parameters of the model, resulting in a significant increase in computational overhead. This not only reduces the real-time computing efficiency of the model, but also increases the difficulty of deployment and application, especially in resource-constrained environments, which may cause the model to fail to meet the requirements of practical applications. Summary of the invention

[0006] In view of the deficiencies in the above-mentioned prior art, the present invention proposes a multi-scale lightweight spatiotemporal network guided by physical structure to solve the shortcomings of the prior art, such as low real-time computing efficiency, high difficulty in deployment and application, and inability to meet actual application requirements in resource-constrained environments.

[0007] Another object of the present invention is to provide a method for predicting the remaining useful life of an aircraft engine.

[0008] In order to achieve the first invention object of the present invention, the present invention provides a multi-scale lightweight spatiotemporal network guided by physical structure, which comprises: a multi-scale temporal feature extraction module, a physical deployment feature classification module, a spatial feature extraction module and an output module connected in sequence;

[0009] The multi-scale feature extraction module is used to perform several downsampling on the time series data collected by each sensor to obtain several layers of time series features with different coarsening scales; and fuse and output the time series features with different coarsening scales in the reverse order of downsampling to obtain several layers of multi-scale time series features;

[0010] The physical deployment feature classification module is a debottleneck convolution layer constructed based on the physical structural connection of each sensor in the aircraft engine component, which is used to embed the multi-scale temporal features of each layer into the corresponding deployment component respectively and convert them into the corresponding first aircraft engine component spatiotemporal features;

[0011] The spatial feature extraction module is used to further extract the spatial relationship between components from the spatiotemporal features of the first aircraft engine components of each layer to obtain the corresponding spatiotemporal features of the second aircraft engine components;

[0012] The output module is used to obtain the remaining service life of the aircraft engine based on the splicing results of the temporal and spatial characteristics of the second aircraft engine components of each layer.

[0013] In one possible implementation, the downsampling used by the multi-scale feature extraction module is average pooling, and the formula is as follows:

[0014]

[0015] Among them, y u,v Represents the value of the position (u,v) in the output time series feature graph, x u·s+a,v·s+b It represents the pixel value of the input time series data feature map at the pooling window (a, b), D×D represents the pooling window size, and s represents the pooling step size.

[0016] In one achievable manner, the multi-scale feature extraction module fuses and outputs temporal features of different coarsening scales in a reverse order of downsampling through a linear layer.

[0017] In an achievable manner, in the physical deployment feature classification module, the aircraft engine components C1, C2, ..., C n And the sensor s deployment on the component, we get:

[0018] C i =[s i1 ,s i2 ,…,s id ];

[0019] Where d represents the number of sensors associated with the i-th component;

[0020] C i The temporal feature vector data of the sensors in the figure are concatenated along the dimension where the sensors are located and input into the anti-bottleneck convolution layer to obtain the first spatiotemporal features of the corresponding components:

[0021]

[0022] First spatiotemporal characterization of all aero-engine components The spatiotemporal characteristics of the first aero-engine component are obtained by splicing.

[0023] The specific parameter settings of the anti-bottleneck convolution layer can be: the number of input channels is the sensor dimension after splicing, the number of hidden layer channels is the number of input channels × expansion factor, and the number of output channels is 1.

[0024] In one implementable method, the spatial feature extraction module is a two-layer graph attention network, including a first graph attention network and a second graph attention network arranged in sequence, and the graph attention network adopts a multi-head attention mechanism; for the output of the first graph attention network, the features of the neighborhood nodes are aggregated by splicing and used as the input of the second graph attention network; for the output of the second graph attention network, the spatiotemporal features of the second aircraft engine component are obtained by averaging the features of the aggregated neighborhood nodes.

[0025] For any layer of first aircraft engine component spatiotemporal features, the first spatiotemporal features of each aircraft engine component represent a node.

[0026] For the first graph attention network, it is necessary to first perform a linear transformation on the node features, then calculate the attention node coefficient and normalize it, and then use K independent attention head matrices to calculate the hidden layer and concatenate the aggregated neighbor node features to obtain the first feature representation of each node;

[0027] For the second graph attention network, it is necessary to first perform a linear transformation of the node features, then calculate the attention node coefficient and normalize it, and then use K independent attention head matrices to calculate the hidden layer and average the neighbor node features to obtain the second feature representation of each node. The second feature representation of all nodes is used as the spatiotemporal feature of the second aircraft engine component.

[0028] In one possible implementation, the output module uses a linear layer. Specifically, the output module includes a first linear layer and a second linear layer; the first linear layer is used to perform linear regression on the spatiotemporal characteristics of the second aircraft engine components of each layer to obtain the regression results of each layer; the second linear layer is used to further perform linear regression on the regression results of all layers to obtain the prediction results of the remaining service life of the aircraft engine.

[0029] In order to achieve the second invention purpose of the present invention, the present invention first analyzes the aircraft engine data, screens out high-value sensor information that can effectively reflect the engine degradation trend, and performs clustering and normalization for different working conditions; then the sensor data is input into the multi-scale linear layer to learn the time series information, and the sensor data is further classified using the physical deployment feature classification module of the multi-scale lightweight spatiotemporal network, and the classification structure relationship is learned through the spatial feature extraction module, and finally the trained multi-scale lightweight spatiotemporal network is used to realize the prediction of the aircraft engine life.

[0030] Specifically, the method for predicting the remaining useful life of an aviation engine provided by the present invention comprises the following steps:

[0031] Step 1: Analyze the historical data obtained by the aircraft engine sensor and construct a training sample data set; this step includes the following sub-steps:

[0032] Step 1.1: Analyze the historical data obtained by the aircraft engine sensors and select the sensor time series data related to the aircraft engine life;

[0033] Step 1.2: Standardize the sensor time series data;

[0034] Step 1.3: Perform a sliding window operation on the standardized sensor time series data to divide it into several short sequence samples of equal length, and build a training sample data set in combination with the aircraft engine remaining useful life RUL label;

[0035] Step 2: Using a training sample data set to train the multi-scale lightweight spatiotemporal network guided by physical structure;

[0036] Step 3: The real-time data collected by sensors related to the life of aircraft engines are input into the trained multi-scale lightweight spatiotemporal network after standardized processing to complete the remaining life prediction of aircraft engines.

[0037] The above step 1 is to construct a training sample dataset for a multi-scale lightweight spatiotemporal network.

[0038] In step 1.1, the raw data obtained by the sensor is analyzed, and the sensor time series data related to the engine life is selected as the input feature according to the parameter conditions and failure modes. The screening condition is sensor data with a good degradation trend; if the sensor changes within the entire operating cycle and the data remains stable, it cannot reflect the degradation and needs to be eliminated. In addition, in order to train the network, the aircraft engine service life RUL label is also required. Here, according to the engine degradation, the original aircraft engine operating time is set in reverse order as the aircraft engine service life RUL label, and a fixed early remaining life is set, which is divided into a segmented linear RUL label.

[0039] In step 1.2, if the sensor time series data is obtained under a single working condition, the sensor time series data is standardized;

[0040] If the sensor time series data is obtained under multiple working conditions, the DBSCAN algorithm is used to perform clustering and normalization processing on the sensor time series data.

[0041] The DBSCAN algorithm is a density-based clustering algorithm that defines clusters by density and connects density-connected points into clusters in a density-reachable manner. The DBSCAN algorithm is used to perform clustering and normalization processing on sensor time series data under various working conditions. The operation is as follows: first, define the two key parameters ∈-domain and core point required by the DBSCAN algorithm;

[0042] The ∈-neighborhood is used to specify the neighborhood range of a point, which is defined as the neighbors of the point within the range; for a point p, its ∈-neighborhood is defined as the set of all points within the distance ∈:

[0043] N ∈ (p)={q∈D|dist(p,q)≤∈};

[0044] Where dist(p,q) represents the distance between points p and q; Euclidean distance can be used;

[0045] Core point: If a point p's ∈-neighborhood contains at least the minimum number of neighbors (MinPts) points required, then p is a core point:

[0046] ∣N ∈ (p) |≥MinPts;

[0047] Based on the above parameters, clusters are calculated according to the following concepts:

[0048] 1) Density direct access: If p is a core point and q is in the ∈-neighborhood of p, then q is considered to be density-directly accessible from p;

[0049] 2) Density reachable: If there exists a series of points p1, p2, ..., p m Make each All from Direct density reachable (that is, density directly reachable), then p m It is density-reachable from p1;

[0050] 3) Density connected: If there exists a point o such that both points p and q are density reachable from o, then p and q are density connected;

[0051] 4) The largest set of density-connected points is defined as a cluster, thus completing clustering;

[0052] Then, each cluster that has completed clustering is standardized.

[0053] The standardization process uses the Z-score standardization process, and the formula is:

[0054]

[0055] Among them, x is the original sensor time series data, μ is the mean of the original sensor time series data, σ is the standard deviation of the original sensor data, and z is the standardized sensor time series data.

[0056] Then, the normalized sensor time series data is divided by sliding window operation to obtain several short sequence samples of equal length. Several short sequence samples of equal length are combined with the corresponding aircraft engine remaining useful life RUL labels to construct a training sample data set.

[0057] The above step 2 is to use the training sample data set to effectively train the multi-scale lightweight spatiotemporal network. This step 2 includes the following sub-steps:

[0058] Step 2.1: The training samples obtained in step 1 are downsampled several times through the multi-scale time series feature extraction module to obtain several layers of time series features with different coarsening scales. Then, the time series features with different coarsening scales are multi-scale fused from coarse to fine by the linear layer to obtain several layers of multi-scale time series features with trend fusion.

[0059] Step 2.2: embed the multi-scale temporal features of several layers obtained in step 2.1 into the corresponding components through the physical deployment feature classification module, and convert them into the spatiotemporal features of the first aircraft engine components corresponding to each layer;

[0060] Step 2.3: Processing the spatiotemporal features of the first aircraft engine components corresponding to each layer through a spatial feature extraction module to obtain the spatiotemporal features of the second aircraft engine components corresponding to each layer;

[0061] Step 2.4: The spatiotemporal characteristics of the second aircraft engine components corresponding to each layer are output to obtain the remaining service life of the aircraft engine through the output module;

[0062] Step 2.5: Calculate the loss value based on the predicted remaining useful life of the aircraft engine and the aircraft engine remaining useful life RUL label;

[0063] Step 2.6: Optimize the parameters of the multi-scale lightweight spatiotemporal network based on the loss value;

[0064] Repeat the above steps 2.1-2.6 until the multi-scale lightweight spatiotemporal network converges and the training is completed.

[0065] In the above step 2.1, the training samples (i.e., short sequence samples) obtained in step 1 are downsampled M times to obtain M+1 layers of time series feature maps with different coarsening scales. Then, the M+1 layers of time series features are fused from coarse to fine multi-scales by the linear layer to construct M+1 layers of multi-scale time series features with trend fusion; that is, in the reverse order of downsampling, the fused output time series features of the next layer are combined with the output time series features of the current layer through the linear layer to obtain the fused output time series features of the current layer.

[0066] In the above step 2.2, based on the aircraft engine components and the sensor s deployment on the components, a list of sensors related to each component is sorted out, namely C i =[s i1 ,s i2 ,…,s id ], and then through the anti-bottleneck convolution layer, the multi-scale temporal features are embedded into the corresponding components and converted into the first spatiotemporal features of the corresponding engine components First spatiotemporal characterization of all aero-engine components The spatiotemporal characteristics of the first aero-engine component are obtained by splicing.

[0067] In the above step 2.3, the spatiotemporal features of the first aircraft engine components corresponding to each layer are input into the spatial feature extraction module, and the spatial relationship between the components is further extracted through the double-layer graph attention network to obtain the spatiotemporal features of the second aircraft engine components corresponding to each layer.

[0068] In the above step 2.4, the spatiotemporal characteristics of the second aircraft engine components corresponding to each layer are subjected to linear layer regression to obtain the prediction result of the remaining service life of the aircraft engine.

[0069] In the above step 2.5, the loss function used to calculate the loss value is the mean square error, and the calculation formula is as follows:

[0070]

[0071] Among them, y l Indicates the real RUL value, Represents the predicted RUL value, and L represents the total number of samples.

[0072] In the above step 2.6, the multi-scale lightweight spatiotemporal network parameters are optimized by the Adam optimization algorithm based on the obtained loss value. Repeat the above steps 2.1-2.6 until the set upper limit of training iterations is reached, and the multi-scale lightweight spatiotemporal network is considered to have converged. The early stopping strategy can also be executed based on the trend of the loss value change (when the loss value tends to be stable or less than the set threshold) or the network evaluation index.

[0073] The root mean square error and S-score are used as evaluation indicators. The specific formulas are as follows:

[0074]

[0075] in,

[0076] In the above step 3, the operating condition of the aircraft engine to be tested is determined, and the collected real-time data can be standardized by the Z-score standardization method based on the historical data of the sensor tested under the operating condition. Then the processed data is input into the trained multi-scale lightweight spatiotemporal network to complete the remaining life prediction of the aircraft engine.

[0077] Compared with the prior art, the present invention has the following technical effects:

[0078] 1) The multi-scale lightweight spatiotemporal network provided by the present invention introduces the relationship between key components of aircraft engines by physically deploying feature classification modules, thereby removing redundant relationships, accurately describing the engine degradation law under multiple working conditions, improving the accuracy of remaining life prediction, reducing operation and maintenance costs, and being more adaptable to different models of aircraft engines, which has important engineering application value;

[0079] 2) The multi-scale lightweight spatiotemporal network provided by the present invention first obtains the time series features of different coarsening scales through the multi-scale feature extraction module, and then fuses the time series features of different coarsening scales, thereby constructing multiple multi-scale time series features with trend fusion, enhancing the time series extraction capability, which can improve the accuracy of the remaining service life of the aircraft engine; at the same time, it also reduces the deployment cost;

[0080] 3) The multi-scale lightweight spatiotemporal network provided by the present invention further processes the multi-scale temporal features of trend fusion through a spatial feature extraction module, thereby effectively extracting the spatial relationship between different aircraft engine components, further improving the accuracy of the remaining service life of the aircraft engine;

[0081] 4) In the process of analyzing the historical data obtained by the aircraft engine sensor, the present invention effectively divides the data under different working conditions through the DBSCAN algorithm, thereby realizing the effective training of the multi-scale lightweight spatiotemporal network. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 A schematic diagram of the structure of a multi-scale lightweight spatiotemporal network based on physical structure guidance provided by the present invention;

[0083] Figure 2 The flowchart of the method for predicting the remaining useful life of an aircraft engine is shown in FIG.

[0084] Figure 3 Schematic diagram of the process of analyzing historical data obtained from aircraft engine sensors;

[0085] Figure 4 Schematic diagram of the training process for a multi-scale lightweight spatiotemporal network;

[0086] Figure 5 is the box plot of sensors in the FD001 subset;

[0087] Figure 6 This is a schematic diagram of the non-clustered normalization of the engine 1 data in the FD002 subset;

[0088] Figure 7 This is a schematic diagram of clustering and normalization of engine 1 data in the FD002 subset;

[0089] Figure 8 Schematic diagram of the layout of key components of an aircraft engine. DETAILED DESCRIPTION

[0090] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0091] Example 1

[0092] This embodiment provides a multi-scale lightweight spatiotemporal network guided by physical structure, such as Figure 1 As shown, it includes a multi-scale temporal feature extraction module, a physical deployment feature classification module, a spatial feature extraction module and an output module which are connected in sequence.

[0093] 1. Multi-scale feature extraction module

[0094] The multi-scale feature extraction module is used to downsample the time series data collected by each sensor several times to obtain several layers of time series features with different coarsening scales; and fuse and output the time series features with different coarsening scales in the reverse order of downsampling to obtain several layers of multi-scale time series features.

[0095] The downsampling used by the multi-scale feature extraction module is average pooling, and its formula is as follows:

[0096]

[0097] Among them, y u,v Represents the value of the position (u,v) in the output time series feature graph, x u·s+a,v·s+b It represents the pixel value of the input time series data feature map at the pooling window (a, b), D×D represents the pooling window size, and s represents the pooling step size.

[0098] Moreover, the multi-scale feature extraction module fuses and outputs the temporal features of different coarsening scales in the reverse order of downsampling through a linear layer.

[0099] In this embodiment, the multi-scale feature extraction module performs M (M=2) downsampling on the input raw data to obtain three layers of time series feature graphs with different coarsening scales, namely, the raw data layer, the coarsening layer 1, and the coarsening layer 2. The output features of the coarsening layer 2 and the output features of the coarsening layer 1 are subjected to a linear layer to obtain a fusion feature a, and the fusion feature a and the output features of the raw data layer are subjected to a linear layer to obtain a fusion feature b. Therefore, the output features of the coarsening layer 2, the fusion feature a, and the fusion feature b are the obtained three layers of multi-scale time series features.

[0100] 2. Physical deployment feature classification module

[0101] The physical deployment feature classification module is a debottlenecking convolution layer constructed based on the physical structural connection of each sensor in the aircraft engine component. It is used to embed the multi-scale temporal features of each layer into the corresponding deployment components and convert them into the corresponding spatiotemporal features of the first aircraft engine component.

[0102] In the physical deployment feature classification module, according to the aircraft engine components C1, C2, ..., C n And the sensor s deployment on the component, we get:

[0103] C i=[s i1 ,s i2 ,…,s id ];

[0104] Where d represents the number of sensors associated with the i-th component;

[0105] C i The temporal feature vector data of the sensors in the figure are concatenated along the dimension where the sensors are located and input into the anti-bottleneck convolution layer to obtain the first spatiotemporal features of the corresponding components:

[0106]

[0107] First spatiotemporal characterization of all aero-engine components The spatiotemporal characteristics of the first aero-engine component are obtained by splicing.

[0108] The specific parameter settings of the anti-bottleneck convolution layer can be: the number of input channels is the sensor dimension after splicing, the number of hidden layer channels is the number of input channels × expansion factor, and the number of output channels is 1. In this embodiment, the expansion factor is 2, and the convolution kernel parameter is [1,1].

[0109] 3. Spatial feature extraction module

[0110] The spatial feature extraction module is used to further extract the spatial relationship between components from the spatiotemporal features of the first aircraft engine components of each layer to obtain the corresponding spatiotemporal features of the second aircraft engine components.

[0111] The spatial feature extraction module is a two-layer graph attention network, including a first graph attention network and a second graph attention network arranged in sequence. The graph attention network adopts a multi-head attention mechanism; for the output of the first graph attention network, the features of the neighborhood nodes are spliced ​​and aggregated as the input of the second graph attention network; for the output of the second graph attention network, the spatiotemporal features of the second aircraft engine component are obtained by averaging the features of the aggregated neighborhood nodes.

[0112] For the first graph attention network, it is necessary to first perform a linear transformation on the node features, then calculate the attention node coefficient and normalize it, and then use K independent attention head matrices to calculate the hidden layer and concatenate the aggregated neighbor node features to obtain the first feature representation of each node; the specific process is as follows:

[0113] The linear transformation formula of node features is as follows:

[0114] h′ i =Wh i ;

[0115] Among them, h i represents the input feature vector of node i, W represents the linear transformation matrix, h′ iRepresents the transformed input feature vector, with a dimension of 16;

[0116] The calculation formula of attention node coefficient and normalization is as follows:

[0117] e ij =LeakyReLU(a T [h′ i ||h′j]);

[0118]

[0119] Among them, a represents a trainable weight vector; || represents a vector connection operation; e ij represents the unnormalized attention coefficient between nodes i and j; Represents the set of neighbor nodes of node i; α ij represents the normalized attention weight;

[0120] Use K (K = 8) independent attention head matrices to calculate the hidden layer and then concatenate the aggregated neighbor node features. The calculation formula is as follows:

[0121]

[0122] Among them, σ represents the nonlinear activation function ReLU, W h represents the kth attention head matrix, h″ i Represents the first feature representation of node i.

[0123] For the second graph attention network, it is necessary to first perform linear transformation on the node features, then calculate and normalize the attention node coefficients, and then use K independent attention head matrices to calculate the hidden layer and average the neighbor node features to obtain the second feature representation of each node. The second feature representation of all nodes is used as the second aircraft engine component spatiotemporal features; the specific process is as follows:

[0124] The linear transformation formula of node features is as follows:

[0125]

[0126] in, represents the input feature vector of node i, represents the linear transformation matrix, is the transformed input feature vector with dimension 1;

[0127] The calculation formula of attention node coefficient and normalization is as follows:

[0128]

[0129] in, Represents a trainable weight vector; || represents a vector connection operation; represents the unnormalized attention coefficient between nodes i and j; Represents the set of neighbor nodes of node i; represents the normalized attention weight;

[0130] Use K (K = 8) independent attention head matrices to calculate the hidden layer and average the aggregated neighbor node features. The calculation formula is as follows:

[0131]

[0132] Among them, σ represents the nonlinear activation function ReLU, represents the kth attention head matrix, Represents the second feature representation of node i.

[0133] 4. Output module

[0134] The output module is used to obtain the remaining service life of the aircraft engine based on the splicing results of the spatiotemporal features of the second aircraft engine components of each layer. The output module uses a linear layer. Specifically, the output module includes a first linear layer and a second linear layer; the first linear layer is used to perform linear regression on the spatiotemporal features of the second aircraft engine components of each layer to obtain the regression results of each layer; the second linear layer is used to further linearly regress the regression results of all layers to obtain the prediction results of the remaining service life of the aircraft engine.

[0135] In this embodiment, the input channel parameter of the first linear layer is the number of nodes n=8, and the input channel parameter of the second linear layer is the number of multi-scale layers after corresponding downsampling, M+1=3.

[0136] Example 2

[0137] This embodiment provides a method for predicting the remaining useful life of an aircraft engine based on the multi-scale lightweight spatiotemporal network provided in Embodiment 1. This embodiment uses the C-MAPSS large commercial turbofan engine degradation data set published by NASA to verify the effectiveness of the method for predicting the remaining useful life of an aircraft engine provided in this embodiment.

[0138] This dataset simulates the working process of a turbofan engine from normal to degraded to failure through physical modeling and simulation, and monitors and records multiple working state parameters during the process. The dataset format contains 1 column of engine number, 1 column of engine operation cycle, 3 columns of operation setting parameters, and 21 columns of sensor data. The sensor data is shown in Table 1 below.

[0139] Table 1. Description of sensors in the C-MAPSS dataset

[0140]

[0141]

[0142] According to the operating parameter conditions and failure modes, the CMAPSS dataset is divided into four subsets, as shown in Table 2. Next, for each subset, the corresponding multi-scale lightweight spatiotemporal network is trained.

[0143] Table 2 Description of C-MAPSS dataset subsets

[0144]

[0145] The method for predicting the remaining useful life of an aircraft engine provided in this embodiment is as follows: Figures 2 to 4 As shown, it includes the following steps:

[0146] Step 1: Analyze the historical data obtained by aircraft engine sensors and construct a training sample dataset.

[0147] This step includes the following sub-steps:

[0148] Step 1.1: Analyze the historical data obtained by aircraft engine sensors and filter out sensor time series data related to aircraft engine life.

[0149] This step analyzes the sensor raw data based on the sensor trend and selects the sensor data that can well show the degradation trend. Figure 5 In the FD001 box plot data shown, the antenna length of some sensors is 0 during the entire operation cycle, and the data remains stable, so it cannot reflect the degradation situation. Therefore, data records of 14 sensors are obtained by removing and screening [2, 3, 4, 7, 8, 9, 11, 12, 13, 14, 15, 17, 20, 21].

[0150] In addition, in order to train the network, the aircraft engine service life RUL label is also required. The filtered data is set as the aircraft engine service life RUL label in reverse order of the original aircraft engine operation time according to the engine degradation; and RUL = 125 is set as the fixed early remaining life, which is divided into a piecewise linear RUL label; after RUL = 125, the engine service life decreases linearly.

[0151] Step 1.2: Standardize the sensor time series data.

[0152] The above data subsets FD001 and FD003 are single working conditions, and the sensor time series data of all engines in the two data subsets are subjected to Z-score standardization.

[0153] For the data subsets FD002 and FD004, each subset includes 6 groups of operating conditions. Therefore, the DBSCAN algorithm (as described above) is first used, with parameters eps = 10 and min_sample = 2, to perform cluster analysis on the multi-condition data to generate 6 groups of conditions, which is consistent with the total number of actual conditions in the data set; further, the operating condition labels are annotated for each condition, and the sensor time series data under the same operating conditions (i.e., the same condition) are normalized by Z-Score. Figure 6 and Figure 7 The normalized graphs of the engine 1 data in FD002 before and after clustering are given respectively. It can be seen from the graph that clustering normalization can ensure the validity of the data.

[0154] Step 1.3: Perform a sliding window operation on the standardized sensor time series data to divide it into several short sequence samples of equal length, and construct a training sample dataset in combination with the aircraft engine remaining useful life RUL label.

[0155] In this step, the sliding window sizes [30, 20, 30, 15] are used for FD001 to FD004 respectively, and the normalized sensor time series data are subjected to sliding window operations to be divided into short sequence samples of equal length.

[0156] The divided short sequence samples are combined with the corresponding aircraft engine remaining useful life RUL labels to construct a training sample dataset.

[0157] As shown in Table 2, in this embodiment, the original training data set provided by CMAPSS is used and processed according to the above steps 1.1-1.3 to obtain a training sample data set, which is randomly divided into a training set and a validation set in a ratio of 8:2.

[0158] At the same time, the original test data set provided by CMAPSS is used as the test set of this embodiment, and is also processed according to the above steps 1.1-1.3.

[0159] Step 2: Use the training sample data set to train the multi-scale lightweight spatiotemporal network provided in Example 1.

[0160] This step includes the following sub-steps:

[0161] Step 2.1: The training samples obtained in step 1 are downsampled several times through the multi-scale time series feature extraction module to obtain several layers of time series features with different coarsening scales. Then, the time series features with different coarsening scales are multi-scale fused from coarse to fine by the linear layer to obtain several layers of multi-scale time series features with trend fusion.

[0162] In this embodiment, the training samples (i.e., short sequence samples) obtained in step 1 are downsampled M times (M=2) through a multi-scale time series feature extraction module to obtain three layers of time series feature graphs with different coarsening scales. Then, the three layers of time series features are multi-scale fused from coarse to fine by the linear layer to construct three layers of multi-scale time series features with trend fusion.

[0163] Step 2.2: The multi-scale time series features of several layers obtained in step 2.1 are embedded into the corresponding components through the physical deployment feature classification module and converted into the corresponding spatiotemporal features of the first aircraft engine components of each layer.

[0164] According to the prior knowledge of the physical deployment structure of sensors in aircraft engine components, the key components of aircraft engines such as Figure 8 As shown in the figure, the key components of the aircraft engine C1, C2, ..., C8 and the sensor s deployment on the key components are sorted out to obtain the sensor list, namely C i =[s i1 ,s i2 ,…,s id ], as shown in Table 3.

[0165] Table 3 C-MAPSS dataset sensor grouping description

[0166]

[0167] In this embodiment, the multi-scale temporal features are embedded into the corresponding components through the anti-bottleneck convolution layer and converted into the first spatiotemporal features of the corresponding engine components. First spatiotemporal characterization of all aero-engine components The spatiotemporal characteristics of the first aero-engine component are obtained by splicing.

[0168] For the three-layer multi-scale time series features with trend fusion, the spatiotemporal features of the first aircraft engine component of the corresponding three layers are obtained through the physical deployment feature classification module.

[0169] Step 2.3: The spatiotemporal features of the first aircraft engine components corresponding to each layer are processed by a spatial feature extraction module to obtain the spatiotemporal features of the second aircraft engine components corresponding to each layer.

[0170] In this embodiment, the spatiotemporal features of the first aircraft engine components of the three layers are input into the spatial feature extraction module, and the spatial relationship between the components is further extracted through the two-layer graph attention network to obtain the corresponding spatiotemporal features of the second aircraft engine components of each layer.

[0171] Step 2.4: The spatiotemporal characteristics of the second aircraft engine components corresponding to each layer are output to the output module to obtain the remaining service life of the aircraft engine.

[0172] In this embodiment, the three layers of corresponding spatiotemporal features of the second aircraft engine components are regressed through the first linear layer and the second linear layer of the output module to obtain the prediction result of the remaining service life of the aircraft engine.

[0173] Step 2.5: Calculate the loss value based on the predicted remaining useful life of the aircraft engine and the remaining useful life RUL label of the aircraft engine.

[0174] In this embodiment, the loss function used to calculate the loss value is the mean square error, and the calculation formula is as follows:

[0175]

[0176] Among them, y l Indicates the real RUL value, Represents the predicted RUL value, and L represents the total number of samples.

[0177] Step 2.6: Optimize the multi-scale lightweight spatiotemporal network parameters based on the loss value.

[0178] In this step, the Adam optimization algorithm is used to optimize the parameters of the multi-scale lightweight spatiotemporal network. The learning rate is an exponentially decayed learning rate, which is initially 0.01 and the decay factor is 0.95.

[0179] Repeat the above steps 2.1-2.6 until the set upper limit of the number of training iterations is reached, and the multi-scale lightweight spatiotemporal network is considered to have converged. In this embodiment, the upper limit of the number of iterations is 30 times, and the batch size of the training samples used in each iteration is 256. After each iterative training, the multi-scale lightweight spatiotemporal network is verified using the validation set data.

[0180] Finally, based on the subset differences, the total parameters of the multi-scale lightweight spatiotemporal network trained in step 2 above are shown in Table 4.

[0181] Table 4 Multi-scale lightweight spatiotemporal network model parameters corresponding to different data subsets

[0182]

[0183] Step 3: The real-time data collected by sensors related to the life of aircraft engines are input into the trained multi-scale lightweight spatiotemporal network after standardized processing to complete the remaining life prediction of aircraft engines.

[0184] In this embodiment, the trained multi-scale lightweight spatiotemporal network is evaluated by a test set. The root mean square error and S-score are used as evaluation indicators, and the specific formula is as follows:

[0185]

[0186] in,

[0187] The prediction effects of the present invention on the remaining life of aircraft engines on the test set are compared with those of other public pure data-driven methods. The results are shown in Tables 5 and 6.

[0188] Table 5 Comparison of RMSE indicators of different methods on C-MAPSS dataset

[0189]

[0190] Table 6 Comparison of S-score indicators of different methods in C-MAPSS dataset

[0191]

[0192] In the comparison of various indicators given in the table, all indicators of the method of the present invention achieve excellent performance, among which the RMSE of FD002 to FD004 are all optimal, and the S-score of FD001, FD002, and FD004 are all optimal.

[0193] In summary, the method of the present invention is applicable to the remaining life prediction of various types of aircraft engines. In practical applications, implementers can use the method of the present invention to introduce the relationship between key components of aircraft engines and integrate different scale features before model construction. The prediction of the remaining service life of aircraft engines using the multi-scale lightweight spatiotemporal network based on physical structure guidance proposed by the method of the present invention can enhance the time series extraction capability and reduce the deployment cost; remove redundant relationships, accurately characterize the engine degradation law under multiple working conditions, improve the accuracy of remaining life prediction, reduce operation and maintenance costs, and be more adaptable to the prediction of different models of aircraft engines and more types of faults.

[0194] Those skilled in the art will appreciate that the embodiments herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A multi-scale lightweight spatiotemporal network guided by physical structure, characterized in that: include: A multi-scale temporal feature extraction module, a physical deployment feature classification module, a spatial feature extraction module and an output module connected in sequence; The multi-scale feature extraction module is used to perform several down-sampling on the time series data collected by each sensor to obtain several layers of time series features with different coarsening scales; The time series features of different coarsening scales are fused and output in the reverse order of downsampling to obtain several layers of multi-scale time series features; The physical deployment feature classification module is a debottlenecking convolution layer constructed based on the physical structural connection of each sensor in the aircraft engine component, which is used to embed the multi-scale temporal features of each layer into the corresponding deployment component respectively and convert them into the corresponding spatiotemporal features of the first aircraft engine component; The spatial feature extraction module is used to further extract the spatial relationship between components from the spatiotemporal features of the first aircraft engine components of each layer to obtain the corresponding spatiotemporal features of the second aircraft engine components; The output module is used to obtain the remaining service life of the aircraft engine based on the splicing results of the temporal and spatial characteristics of the second aircraft engine components of each layer.

2. The multi-scale lightweight spatiotemporal network based on physical structure guidance according to claim 1, characterized in that: The multi-scale feature extraction module fuses and outputs the temporal features of different coarsening scales in the reverse order of downsampling through a linear layer.

3. The multi-scale lightweight spatiotemporal network based on physical structure guidance according to claim 1, characterized in that: In the physical deployment feature classification module, according to the aircraft engine components C1, C2, ..., C n And the sensor s deployment on the component, we get: C i =[s i1 ,s i2 ,…,s id ]; Where d represents the number of sensors associated with the i-th component; C i The temporal feature vector data of the sensors in the figure are concatenated along the dimension where the sensors are located and input into the anti-bottleneck convolution layer to obtain the first spatiotemporal features of the corresponding components: First spatiotemporal characterization of all aero-engine components The splicing results are used to obtain the spatiotemporal characteristics of the first aero-engine component.

4. The multi-scale lightweight spatiotemporal network based on physical structure guidance according to claim 1, characterized in that: The spatial feature extraction module is a two-layer graph attention network, including a first graph attention network and a second graph attention network arranged in sequence, and the graph attention network adopts a multi-head attention mechanism; for the output of the first graph attention network, the features of the neighborhood nodes are spliced ​​and aggregated as the input of the second graph attention network; for the output of the second graph attention network, the spatiotemporal features of the second aircraft engine component are obtained by averaging the features of the aggregated neighborhood nodes.

5. The multi-scale lightweight spatiotemporal network based on physical structure guidance according to claim 4, characterized in that: For any layer of the first spatiotemporal features of an aircraft engine component, the first spatiotemporal features of each aircraft engine component represent a node; For the first graph attention network, it is necessary to first perform a linear transformation on the node features, then calculate the attention node coefficient and normalize it, and then use K independent attention head matrices to calculate the hidden layer and concatenate the aggregated neighbor node features to obtain the first feature representation of each node; For the second graph attention network, it is necessary to first perform a linear transformation of the node features, then calculate the attention node coefficient and normalize it, and then use K independent attention head matrices to calculate the hidden layer and average the neighbor node features to obtain the second feature representation of each node. The second feature representation of all nodes is used as the spatiotemporal feature of the second aircraft engine component.

6. The multi-scale lightweight spatiotemporal network based on physical structure guidance according to claim 1, characterized in that: The output module uses a linear layer.

7. A method for predicting the remaining useful life of an aircraft engine, characterized in that: The following steps are involved: Step 1: Analyze the historical data obtained by the aircraft engine sensor and construct a training sample data set; this step includes the following sub-steps: Step 1.1: Analyze the historical data obtained by the aircraft engine sensors and select the sensor time series data related to the aircraft engine life; Step 1.2: Standardize the sensor time series data; Step 1.3: Perform a sliding window operation on the standardized sensor time series data to divide it into several short sequence samples of equal length, and build a training sample data set in combination with the aircraft engine remaining useful life RUL label; Step 2: Using a training sample data set to train the multi-scale lightweight spatiotemporal network guided by physical structure as described in any one of claims 1 to 6; Step 3: The real-time data collected by sensors related to the life of aircraft engines are input into the trained multi-scale lightweight spatiotemporal network after standardized processing to complete the remaining life prediction of aircraft engines.

8. The method for predicting the remaining useful life of an aircraft engine according to claim 7, characterized in that: In step 1.2, if the sensor time series data is obtained under a single working condition, the sensor time series data is standardized; If the sensor time series data is obtained under multiple working conditions, the DBSCAN algorithm is used to perform clustering and normalization processing on the sensor time series data.

9. The method for predicting the remaining useful life of an aircraft engine according to claim 8, characterized in that: The clustering and normalization processing of sensor time series data under various working conditions using DBSCAN algorithm is as follows: first, define the two key parameters ∈ - domain and core point required by DBSCAN algorithm; The ∈-neighborhood is used to specify the neighborhood range of a point, which is defined as the neighbors of the point within the range; for a point p, its ∈-neighborhood is defined as the set of all points within the distance ∈: N ∈ (p)={q∈D∣dist(p,q)≤∈}; Where dist(p,q) represents the distance between points p and q; Core point: If a point p's ∈-neighborhood contains at least the minimum number of neighbors required, then p is a core point: ∣N ∈ (p)∣≥MinPts; Based on the above parameters, clusters are calculated according to the following concepts: 1) Density direct access: If p is a core point and q is in the ∈-neighborhood of p, then q is considered to be density-directly accessible from p; 2) Density reachable: If there exists a series of points p1, p2, ..., p m So that each p w+1 All from p w Direct density can be reached, then p m It is density-reachable from p1; 3) Density connected: If there exists a point o such that both points p and q are density reachable from o, then p and q are density connected; 4) The largest set of density-connected points is defined as a cluster, thus completing clustering; Then, each cluster that has completed clustering is standardized.

10. The method for predicting the remaining service life of aviation engines based on a multi-scale lightweight spatiotemporal network guided by physical structure according to claim 6, characterized in that: The step 2 comprises the following sub-steps: Step 2.1: The training samples obtained in step 1 are downsampled several times through the multi-scale time series feature extraction module to obtain several layers of time series features with different coarsening scales. Then, the time series features with different coarsening scales are multi-scale fused from coarse to fine by the linear layer to obtain several layers of multi-scale time series features with trend fusion. Step 2.2: embed the multi-scale temporal features of several layers obtained in step 2.1 into the corresponding components through the physical deployment feature classification module, and convert them into the spatiotemporal features of the first aircraft engine components corresponding to each layer; Step 2.3: Processing the spatiotemporal features of the first aircraft engine components corresponding to each layer through a spatial feature extraction module to obtain the spatiotemporal features of the second aircraft engine components corresponding to each layer; Step 2.4: The spatiotemporal characteristics of the second aircraft engine components corresponding to each layer are output to obtain the remaining service life of the aircraft engine through the output module; Step 2.5: Calculate the loss value based on the predicted remaining useful life of the aircraft engine and the aircraft engine remaining useful life RUL label; Step 2.6: Optimize the parameters of the multi-scale lightweight spatiotemporal network based on the loss value; Repeat the above steps 2.1-2.6 until the multi-scale lightweight spatiotemporal network converges to obtain a trained model.

Images