Aero-engine gas path performance anomaly detection method fusing spatial and temporal features

Abstract

This invention discloses a method for detecting anomalous performance of aero-engine gas path by fusing spatial and temporal features, comprising the following steps: S1. Data acquisition; S2. Modeling for the problem of detecting anomalous performance of aero-engine gas path; S3. Establishing an exhaust gas temperature (EGT) anomaly detection model that fuses spatial and temporal features; S4. Performing spatial feature extraction based on GCN; S5. Performing temporal feature extraction based on gated temporal convolutional neural network; S6. Improving the Informer prediction model. This invention introduces a BiLSTM module into the Informer encoder to enhance the local dependency of the degenerate feature map. The Probesparse self-attention mechanism realizes adaptive differentiation and selection of the main features in the feature matrix, improving the prediction accuracy. The proposed method is trained and validated using a large amount of real-world flight data. The proposed model solves the problems of unrealistic aero-engine simulation data and incomplete experimental data.

Description

Technical Field

[0001] This invention relates to the field of aero-engine gas path performance anomaly detection technology, specifically to an aero-engine gas path performance anomaly detection method that integrates spatial and temporal characteristics. Background Technology

[0002] Currently, methods for detecting gas path anomalies in aero-engines can be categorized into two types: model-based methods and data-driven methods. Model-based methods require establishing accurate physical and mathematical models of the engine to describe its operating characteristics. Since a real aero-engine is an extremely complex nonlinear system, model-based anomaly detection methods demand high precision in model design. Data-driven methods, on the other hand, only require sufficient judgment experience and historical data, exhibiting greater generalization and adaptability. Therefore, data-driven gas path anomaly detection methods have been extensively explored.

[0003] In practical engineering applications, the performance of aero-engine gas path components gradually deteriorates with increasing flight time, leading to an increase in fuel injection to meet thrust requirements during actual operation. This incremental fuel combustion results in an increase in exhaust gas temperature (EGT). Therefore, exhaust gas temperature is a crucial indicator of aero-engine gas path performance and a reference for determining disassembly deadlines, optimizing maintenance plans, and developing operational schedules. Many studies use EGT as a predictive indicator for monitoring and detecting engine gas path performance anomalies. Li et al. employed reinforcement learning to predict EGT as an engine anomaly monitoring indicator and revised engine maintenance plans accordingly. Zhong et al. combined convolutional neural networks (CNNs) with support vector machines (SVMs) to use EGT to assess the operating status of aero-engines and then used other parameters to assist in diagnosing engine faults.

[0004] However, during flight, when the gas path parameters of an aero-engine change rapidly and dynamically, improving the prediction accuracy of gas path parameters such as exhaust temperature to enhance the accuracy of anomaly monitoring is one of the most challenging issues.

[0005] In summary, a method for detecting anomalous aero-engine gas path performance that integrates spatial and temporal characteristics needs to be proposed to address the aforementioned issues. Summary of the Invention

[0006] The purpose of this invention is to provide a method for detecting anomalous performance of aero-engine air path by integrating spatial and temporal characteristics, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A method for detecting anomalous performance of aero-engine gas path by integrating spatial and temporal characteristics includes the following steps:

[0009] S1. Data Acquisition;

[0010] S2. Modeling is performed to address the problem of detecting abnormal gas path performance in aero-engines;

[0011] S3. Establish an exhaust gas temperature (EGT) anomaly detection model that integrates spatial and temporal characteristics;

[0012] S4. Perform spatial feature extraction based on GCN;

[0013] S5. Perform temporal feature extraction based on a gated temporal convolutional neural network;

[0014] S6. Improve the Informer prediction model;

[0015] S7. Set the abnormal threshold;

[0016] S8. Perform data processing and experimental setup;

[0017] S9. Analyze the experiment and draw conclusions.

[0018] Preferably, step S1 uses actual flight data of a B737 from a certain airline. The engine model used in the aircraft is CFM56_7B, and the data comes from the B737's fast access recorder.

[0019] Preferably, step S2 is implemented as follows:

[0020] S2-1. Define the exhaust temperature of an aircraft engine as Y∈R T The historical time series matrix, Y, has values ​​at T time steps within a fixed input window. 1:T The covariates related to Y are X∈R N×T Where N is the number of covariates, and each submatrix X i These are all time-varying sequences, representing a covariate matrix related to sequence Y, namely, multidimensional information on airspeed, flight altitude, fuel flow rate, and rotor speed;

[0021] S2-2. Goal Setting: The goal is to learn a model F(·) based on the covariate X. 1:T+1 Predicting the exhaust temperature one time step in the future θ are the parameters of the model;

[0022] S2-3. Predicted exhaust temperature Error calculations were performed between the actual exhaust temperature Y and the actual exhaust temperature, and anomaly detection of aero-engine exhaust temperature was performed based on the proposed monitoring strategy.

[0023] Preferably, step S3 is implemented as follows:

[0024] S3-1. Use the preprocessing module to preprocess historical data in the QAR of aero-engines and filter relevant features;

[0025] S3-2. Feed the processed relevant parameter sequence into the GCN and gated TCN modules to extract the corresponding features, and then predict the normal EGT based on the improved Informer;

[0026] S3-3. Compare the predicted normal EGT with the actual EGT, and determine the alarm level based on the proposed fault prediction strategy;

[0027] S3-4. Locate potential faults and issue early warnings based on alarm levels;

[0028] S3-5. During aircraft operation, repeat steps S3-1 to S3-4 for newly input data.

[0029] Preferably, step S4 is implemented as follows:

[0030] S4-1. GCN is used to extract spatial features from the spatial state parameters of the aircraft's attitude change, airspeed, and flight altitude. The spatial state parameters of the aero-engine are used as nodes in the graph, including airspeed, flight altitude, and pitch angle. The relationships between nodes are established by the interaction and dependency between the state parameters.

[0031] S4-2. Construct a graph structure G = (V, E), where V is the set of all spatial state parameter nodes and E represents the set of all edges. Use GCN to learn the features of the graph structure G to obtain the feature representation of each time point, thereby extracting the spatial relationships of the spatial state parameters. The graph convolutional layer uses the adjacency matrix A of the graph structure as prior knowledge and simulates the spatial dependence during flight by extracting the pairwise relationships between each state parameter. The adjacency matrix A is defined as in formula (1), where a represents the different spatial state parameters v at each time point. i ,v j The magnitude of the correlation, where τ represents the correlation threshold:

[0032]

[0033] S4-3. The Pearson correlation coefficient is used to characterize the correlation between different spatial state parameters, as shown in formula (2):

[0034]

[0035] In the formula, the closer the value of a is to 1, the stronger the correlation between the two sets of data. The correlation between the two parameters is evaluated by the absolute correlation coefficient value, which ranges from 0.8 to 1.0, 0.6 to 0.8, 0.4 to 0.6, 0.2 to 0.4 and 0 to 0.2, representing extremely strong, strong, moderate, weak and extremely weak correlations. In this step, the total correlation threshold τ is set to 0.4.

[0036] S4-4. The state parameters at each time point on the graph are considered as graph signals. The graph convolution of the graph signal x and the graph convolution kernel Θ is defined by formula (3):

[0037]

[0038] In the formula, For the graph convolution operator, L = DA is the Laplacian matrix of the graph, using symmetric normalization as shown in formula (4):

[0039]

[0040] In the formula, I N Let A be the identity matrix, D be the adjacency matrix, and D be the diagonal matrix.

[0041] In formula (3), U is the Fourier transform basis obtained by the Laplace eigenvalue decomposition L = UΛU, where Λ = diag([λ0,…,λ)). N-1 ])∈R N×N Given an eigenvalue diagonal matrix, a graph convolution kernel is constructed using Chebyshev polynomials to simplify the calculation process, as shown in formula (5):

[0042]

[0043] In the formula, It is a k-th order Chebyshev polynomial. λ max is the largest eigenvalue of the Laplacian matrix, and K is the kernel size;

[0044] S4-5. If only the first-order Chebyshev polynomial is considered, it is further simplified to a first-order approximation and a self-loop is added. At the same time, the nonlinearity of the convolution result is enhanced by the activation function. The characteristics of the flight state parameters after graph convolution are given by formula (6):

[0045]

[0046] In the formula, for The degree matrix, H (l+1) Let θ be the feature matrix of the (l+1)th layer after convolution. l Let be the parameters of the l-th layer, and σ(·) be the activation function.

[0047] Preferably, step S5 is implemented as follows:

[0048] S5-1. One-dimensional convolution is improved by using causal convolution, dilated convolution, and residual connections;

[0049] S5-2. For n-dimensional gas path input parameters X = [x1, x2, ... x T The dilated convolution F is defined by formula (7):

[0050]

[0051] In the formula, d is the dilation rate, k is the kernel size, and f = {ω0, ω1, ..., ω k-1} represents a filter.

[0052] Meanwhile, residual connections improve the efficiency of information propagation through cross-layer direct connections. TCN enhances the network's learning ability by introducing residual connections, making the structure more stable.

[0053] S5-3. A gating mechanism is added between the dilated causal convolutions of the TCN. The gated temporal convolution unit mainly uses a gated activation unit to dynamically adjust the input signal to determine the probability of information from this layer being passed to the next convolutional layer. At the same time, the convolutional structure does not require a forget gate; only the output gate needs to be considered to control the propagation of signals between layers. The gated temporal convolution unit consists of two parts, as shown in formula (8):

[0054] H l =(W * X+b)⊙σ(V * X+c) (8);

[0055] In the formula, W and V are parameter matrices, b and c are learned model parameters; ⊙ represents the pointwise multiplication operation of the matrix, and σ(·) is the sigmoid activation function, which filters out useless information.

[0056] S5-4. For multidimensional gas path time series data X = [X (t-T+1) ,...,X (t) ]∈R T×P Where T is the historical time window and P is the feature dimension, according to formula (7), the causal convolution with the i-th parameter can be defined as formula (9):

[0057]

[0058] In the formula, h i,tThis represents the convolution result of the i-th parameter at time t, where d is the dilation rate and ω is the lattice depth. k,p For each element of the convolution kernel, ω in the formula constitutes the convolution kernel W∈R. K×F×P In the formula, K is the kernel size and F is the number of output channels;

[0059] S5-5. Combining formula (8), the time parameter features extracted by multi-layer gated TCN can be expressed as formula (10):

[0060]

[0061] In the formula, H l For the output characteristics of the l-th gated TCN, W l The kernel of the l-th layer is... * dl represents a tensor causal convolution with an expansion rate of d in the l-th layer, and σ(·) is the sigmoid activation function used to control the propagation ratio between layers.

[0062] Preferably, step S6 is implemented as follows:

[0063] S6-1. Based on Informer, introduce a BiLSTM module in the encoder section;

[0064] S6-2. The Probesparse self-attention mechanism in Informer proposes a query sparsity measure based on Kullback-Leibler divergence, allowing the Key to focus only on the top u most important queries, as shown in formula (12):

[0065]

[0066] In the formula, It is a sparse matrix containing only the first u queries;

[0067] S6-3. Replace the traditional self-attention mechanism with a probable sparse self-attention mechanism. Only a portion of representative samples are selected from the input sequence for computation. The self-attention distillation method is used to reduce the network dimension, thereby further reducing the computational complexity.

[0068] S6-4. The self-attention distillation layer consists of a one-dimensional convolution, using ELU as the activation function. After convolution, it passes through a pooling layer, concatenates the feature maps of all multi-head self-attention layers, and then inputs them into the Informer decoder.

[0069] S6-5. Construct a decoder consisting of a multi-head probabilistic sparse self-attention layer, a full attention layer, and a fully connected layer;

[0070] S6-6. Introduce a BiLSTM layer in the Informer encoder section. At time t, the calculation formulas for the hidden layer and the output layer in both directions are as shown in formulas (13), (14) and (15):

[0071]

[0072]

[0073]

[0074] Output value y t By combining two LSTM units, a weighted fully connected structure is used to further extract the temporal information of the Informer encoder output features.

[0075] Preferably, the implementation steps of step S7 are as follows: select a Gaussian distribution to map the accuracy of the prediction error, define the error X as following a normal distribution with expected value u and variance σ, denoted as N(u,σ2), and the probability density function p(x) of the Gaussian distribution is given by formula (16):

[0076]

[0077] The threshold is calculated based on a 99% confidence level of the error using the validation dataset, which means that the false alarm rate of the predicted exhaust temperature under normal operating conditions is 1%.

[0078] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention fully considers the characteristics of aero-engine data in actual flights and proposes a new gas path performance anomaly detection model. Unlike other system monitoring methods, the method in this paper realizes full-process monitoring of flight operations. It designs a feature extraction module based on graph convolutional neural networks and gated temporal convolutional neural networks, which effectively captures the spatial variation characteristics of aero-engine data in different flight stages, as well as the short-term proximity and long-term dependence in temporal variation. The BiLSTM module is introduced into the Informer encoder to enhance the local dependence of the degenerate feature map. The Probesparse self-attention mechanism realizes the adaptive differentiation and selection of the main features in the feature matrix, which improves the prediction accuracy. A large amount of real-world flight data is used to train and verify the proposed method. The proposed model solves the problems of unrealistic aero-engine simulation data and incomplete experimental data. Attached Figure Description

[0079] Figure 1 Flowchart for detecting abnormal engine exhaust temperature;

[0080] Figure 2 Flowchart of a method for modeling graph convolutional networks;

[0081] Figure 3 (a) Gated TCN module; (b) Gated TCN module;

[0082] Figure 4 Here is a diagram of the Informer encoder-decoder architecture;

[0083] Figure 5 This is a schematic diagram of a BiLSTM layer.

[0084] Figure 6 A structural diagram of the improved Informer encoder;

[0085] Figure 7 This is a graph showing the results of outlier detection and filling.

[0086] Figure 8 (a) Pearson results for correlation analysis; (b) MIC results;

[0087] Figure 9 The image shows the prediction results for engines A and B based on the model of this invention.

[0088] Figure 10 A graph showing the variation in prediction accuracy with different numbers of training samples;

[0089] Figure 11 The prediction results for engines A and B under different models are shown in the figure.

[0090] Figure 12 A comparison chart of evaluation metrics for different models;

[0091] Figure 13 Ablation experiment diagrams for each module;

[0092] Figure 14 This is a diagram showing the diagnostic results for engine A malfunction.

[0093] Figure 15 This is a diagram showing the diagnostic results for engine B.

[0094] Figure 16 The images show the results of detecting abnormalities on normal flights in Case 1 and Case 2. Detailed Implementation

[0095] Example 1

[0096] A method for detecting anomalous performance of aero-engine gas path by integrating spatial and temporal characteristics includes the following steps:

[0097] S1. Data Acquisition:

[0098] The actual flight data of a certain airline's B737 was used. The engine model used in the aircraft was CFM56_7B. The data all came from the B737's fast access recorder.

[0099] S2. Modeling the problem of detecting abnormal performance of aero-engine air passages:

[0100] S2-1. Define the exhaust temperature of an aircraft engine as Y∈R T The historical time series matrix, Y, has values ​​at T time steps within a fixed input window. 1:T The covariates related to Y are X∈R N×T Where N is the number of covariates, and each submatrix X i These are all time-varying sequences, representing a covariate matrix related to sequence Y, namely, multidimensional information on airspeed, flight altitude, fuel flow rate, and rotor speed;

[0101] S2-2. Goal Setting: The goal is to learn a model F(·) based on the covariate X. 1:T+1 Predicting the exhaust temperature one time step in the future θ are the parameters of the model;

[0102] S2-3. Predicted exhaust temperature Error calculations were performed between the actual exhaust temperature Y and the actual exhaust temperature Y, and anomaly detection of aero-engine exhaust temperature was performed based on the proposed monitoring strategy.

[0103] S3. Establish an exhaust gas temperature (EGT) anomaly detection model that integrates spatial and temporal characteristics:

[0104] S3-1. Use the preprocessing module to preprocess historical data in the QAR of aero-engines and filter relevant features;

[0105] S3-2. Feed the processed relevant parameter sequence into the GCN and gated TCN modules to extract the corresponding features, and then predict the normal EGT based on the improved Informer;

[0106] S3-3. Compare the predicted normal EGT with the actual EGT, and determine the alarm level based on the proposed fault prediction strategy;

[0107] S3-4. Locate potential faults and issue early warnings based on alarm levels;

[0108] S3-5. During aircraft operation, repeat steps S3-1 to S3-4 for newly input data.

[0109] S4. Perform spatial feature extraction based on GCN:

[0110] S4-1. GCN is used to extract spatial features from the spatial state parameters of the aircraft's attitude change, airspeed, and flight altitude. The spatial state parameters of the aero-engine are used as nodes in the graph, including airspeed, flight altitude, and pitch angle. The relationships between nodes are established by the interaction and dependency between the state parameters.

[0111] S4-2. Construct a graph structure G = (V, E), where V is the set of all spatial state parameter nodes and E represents the set of all edges. Use GCN to learn the features of the graph structure G to obtain the feature representation of each time point, thereby extracting the spatial relationships of the spatial state parameters. The graph convolutional layer uses the adjacency matrix A of the graph structure as prior knowledge and simulates the spatial dependence during flight by extracting the pairwise relationships between each state parameter. The adjacency matrix A is defined as in formula (1), where a represents the different spatial state parameters v at each time point. i ,v j The magnitude of the correlation, where τ represents the correlation threshold:

[0112]

[0113] S4-3. The Pearson correlation coefficient is used to characterize the correlation between different spatial state parameters, as shown in formula (2):

[0114]

[0115] In the formula, the closer the value of a is to 1, the stronger the correlation between the two sets of data. The correlation between the two parameters is evaluated by the absolute correlation coefficient value, which ranges from 0.8 to 1.0, 0.6 to 0.8, 0.4 to 0.6, 0.2 to 0.4 and 0 to 0.2, representing extremely strong, strong, moderate, weak and extremely weak correlations. In this step, the total correlation threshold τ is set to 0.4.

[0116] S4-4. The state parameters at each time point on the graph are considered as graph signals. The graph convolution of the graph signal x and the graph convolution kernel Θ is defined by formula (3):

[0117]

[0118] In the formula, For the graph convolution operator, L = DA is the Laplacian matrix of the graph, using symmetric normalization as shown in formula (4):

[0119]

[0120] In the formula, I N Let A be the identity matrix, D be the adjacency matrix, and D be the diagonal matrix.

[0121] In formula (3), U is the Fourier transform basis obtained by the Laplace eigenvalue decomposition L = UΛU, where Λ = diag([λ0,…,λ)). N-1 ])∈R N×N Given an eigenvalue diagonal matrix, a graph convolution kernel is constructed using Chebyshev polynomials to simplify the calculation process, as shown in formula (5):

[0122]

[0123] In the formula, It is a k-th order Chebyshev polynomial. λ max is the largest eigenvalue of the Laplacian matrix, and K is the kernel size;

[0124] S4-5. If only the first-order Chebyshev polynomial is considered, it is further simplified to a first-order approximation and a self-loop is added. At the same time, the nonlinearity of the convolution result is enhanced by the activation function. The characteristics of the flight state parameters after graph convolution are given by formula (6):

[0125]

[0126] In the formula, for The degree matrix, H (l+1) Let θ be the feature matrix of the (l+1)th layer after convolution. l Let σ(·) be the parameter of the l-th layer, and σ(·) be the activation function;

[0127] S5. Perform temporal feature extraction based on a gated temporal convolutional neural network:

[0128] S5-1. One-dimensional convolution is improved by using causal convolution, dilated convolution, and residual connections;

[0129] S5-2. For n-dimensional gas path input parameters X = [x1, x2, ... x T The dilated convolution F is defined by formula (7):

[0130]

[0131] In the formula, d is the dilation rate, k is the kernel size, and f = {ω0, ω1, ..., ω k-1} represents a filter.

[0132] Meanwhile, residual connections improve the efficiency of information propagation through cross-layer direct connections. TCN enhances the network's learning ability by introducing residual connections, making the structure more stable.

[0133] S5-3. A gating mechanism is added between the dilated causal convolutions of the TCN. The gated temporal convolution unit mainly uses a gated activation unit to dynamically adjust the input signal to determine the probability of information from this layer being passed to the next convolutional layer. At the same time, the convolutional structure does not require a forget gate; only the output gate needs to be considered to control the propagation of signals between layers. The gated temporal convolution unit consists of two parts, as shown in formula (8):

[0134] H l =(W * X+b)⊙σ(V * X+c) (8);

[0135] In the formula, W and V are parameter matrices, b and c are learned model parameters; ⊙ represents the pointwise multiplication operation of the matrix, and σ(·) is the sigmoid activation function, which filters out useless information.

[0136] S5-4. For multidimensional gas path time series data X = [X (t-T+1) ,...,X (t) ]∈R T×P Where T is the historical time window and P is the feature dimension, according to formula (7), the causal convolution with the i-th parameter can be defined as formula (9):

[0137]

[0138] In the formula, h i,t This represents the convolution result of the i-th parameter at time t, where d is the dilation rate and ω is the lattice depth. k,p For each element of the convolution kernel, ω in the formula constitutes the convolution kernel W∈R. K×F×P In the formula, K is the kernel size and F is the number of output channels;

[0139] S5-5. Combining formula (8), the time parameter features extracted by multi-layer gated TCN can be expressed as formula (10):

[0140]

[0141] In the formula, H l For the output characteristics of the l-th gated TCN, W l The kernel of the l-th layer is... * dl represents a tensor causal convolution with dilation rate d in the l-th layer, and σ(·) is the sigmoid activation function used to control the propagation ratio between layers;

[0142] S6. Improved Informer prediction model:

[0143] S6-1. Based on Informer, introduce a BiLSTM module in the encoder section;

[0144] S6-2. The Probesparse self-attention mechanism in Informer proposes a query sparsity measure based on Kullback-Leibler divergence, allowing the Key to focus only on the top u most important queries, as shown in formula (12):

[0145]

[0146] In the formula, It is a sparse matrix containing only the first u queries;

[0147] S6-3. Replace the traditional self-attention mechanism with a probable sparse self-attention mechanism. Only a portion of representative samples are selected from the input sequence for computation. The self-attention distillation method is used to reduce the network dimension, thereby further reducing the computational complexity.

[0148] S6-4. The self-attention distillation layer consists of a one-dimensional convolution, using ELU as the activation function. After convolution, it passes through a pooling layer, concatenates the feature maps of all multi-head self-attention layers, and then inputs them into the Informer decoder.

[0149] S6-5. Construct a decoder consisting of a multi-head probabilistic sparse self-attention layer, a full attention layer, and a fully connected layer;

[0150] S6-6. Introduce a BiLSTM layer in the Informer encoder section. At time t, the calculation formulas for the hidden layer and the output layer in both directions are as shown in formulas (13), (14) and (15):

[0151]

[0152]

[0153]

[0154] Output value y t By combining two LSTM units, a weighted fully connected structure is used to further extract the temporal information of the Informer encoder output features.

[0155] S7. Set an anomaly threshold:

[0156] We choose a Gaussian distribution to map the accuracy of the prediction error. We define the error X as following a normal distribution with expected value u and variance σ, denoted as N(u,σ2). The probability density function p(x) of the Gaussian distribution is given by formula (16):

[0157]

[0158] The threshold is calculated based on the 99% confidence level of the error using the validation dataset, which means that the false alarm rate of the predicted exhaust temperature under normal operating conditions is 1%.

[0159] Example 2

[0160] I. Data Processing and Experiment Setup

[0161] Experimental setup:

[0162] The computer platform used in this invention is a Win10 system, a 12th Gen Intel@Core™ i7-12700H, 16GB of RAM, and an RTX2080Ti graphics card. The construction of the anomaly model proposed in this invention is based on Python 3.7.

[0163] Data preprocessing

[0164] Due to communication problems, sensor failures or signal loss, and the influence of other signals during transmission, QAR data contains a certain number of outliers and missing values. Therefore, it is necessary to perform outlier monitoring and data imputation to provide a high-quality data foundation for subsequent research.

[0165] This invention employs box plots for outlier removal. The box length of the box plot is the interquartile range (IQR), and its expression is:

[0166] J IQR =Q3-Q1 (16)

[0167] F u and F l These represent the upper and lower edges of the box plot, respectively. Values ​​exceeding these edges are marked as outliers. The expressions for the upper and lower edges are as follows:

[0168]

[0169] In the formula, Q1 and Q3 are the lower and upper quartiles, respectively, and k is the outlier coefficient. When k = 1.5, values ​​greater than or less than F... u and F l The parameter values ​​are mild outliers. When k=3, this is an extreme outlier. In this invention, k=3 is used to retain mild outliers in the parameter data and remove extreme outliers.

[0170] This invention uses piecewise linear interpolation to complete missing data, as shown in the following formula.

[0171] X k ={x 1k ,x 2k ,…,x nk} (18)

[0172] In the formula, is the set of data indicators; k is the number of data indicators; and n is the number of elements of the feature indicators.

[0173] x ik =N i-1k x i-1k +N i+1k x i+1k (19)

[0174] In the formula, x ik These are missing values. To verify the effectiveness of outlier handling and data imputation, Figure 7 The results of outlier detection and filling for exhaust temperature of a sample aircraft engine on a certain flight.

[0175] Considering the high-dimensional nature of aero-engine data and the need for continuous updates and iterations, the Z-score normalization method is chosen, as shown in the following formula:

[0176]

[0177] In the formula, x represents the original data, μ is the mean of the data, and σ is the standard deviation of the data.

[0178] Correlation analysis and feature selection:

[0179] Exhaust temperature is influenced by a variety of factors, including fuselage angle, altitude, load, and combustion efficiency during flight, all of which cause significant changes in exhaust temperature. In this invention, we categorize the parameters recorded by QAR (Quick Airflow Animation) and comprehensively analyze the impact of aero-engine performance and flight characteristics on exhaust temperature. Specifically, we include parameters closely related to flight attitude, such as changes in altitude and fuselage angle, within the category of spatial state parameters, fully considering the changes in exhaust temperature under different flight conditions. Simultaneously, we classify other gas path parameters, such as fuel flow rate and rotor speed, as time-state parameters, allowing for a more detailed analysis of the dynamic trends of engine performance over time.

[0180] To reduce the number of model inputs and avoid overfitting, this invention uses the Pearson correlation coefficient to analyze the linear relationship between aircraft engine exhaust temperature and various parameters recorded by QAR. However, the relationship between exhaust temperature and these features is not only significant and direct, but also involves numerous indirect nonlinear connections. Therefore, we further use the maximum mutual information coefficient (MIC) method to conduct an in-depth study of their nonlinear relationships. Equation (21) defines the Pearson coefficient, while the calculation of MIC is based on equations (22) and (23). The results are as follows: Figure 8 As shown.

[0181]

[0182]

[0183]

[0184] The Pearson correlation between two parameters can be evaluated using the absolute correlation coefficient, ranging from 0.8 to 1.0, 0.6 to 0.8, 0.4 to 0.6, 0.2 to 0.4, and 0 to 0.2, representing extremely strong, strong, moderate, weak, and extremely weak correlations, respectively. The closer the maximum mutual information value is to 1, the stronger the correlation between the parameters. This invention selects a Pearson correlation coefficient threshold of 0.4 and a maximum mutual information threshold of 0.6. Based on the combined results of Pearson correlation and maximum mutual information, the input features and corresponding results of the model of this invention are shown in Table 1.

[0185] Table 1. Model Input Features and Correlation Analysis Results

[0186]

[0187]

[0188] Model parameter settings:

[0189] The main parameters of the model include model features, encoder input dimension (enc_in), decoder input dimension (dec_in), output dimension (c_out), model size (d_model), number of multi-head attention nodes (n_heads), batch size (batch_size), encoder input length (seq_len), number of sampling factors (factor), decoder input length (label_len), number of training iterations (train_epochs), BiLSTM hidden layer dimension (hidden_size), number of residual channels in gated TCN, and number of channels in dilated convolutions (dilation_channels). Extensive experiments were conducted to determine the optimal model parameters, using known values ​​for the empirical parameters d_model, n_heads, and factor, as well as experimentally defined values ​​for features, enc_in, dec_in, c_out, and pred_len.

[0190] Through grid search experiments, we experimented with different batch_size values, including {16, 32, 64}; seq_len values, including {10, 30, 60}; label_len values, including {5, 15, 30}; train_epochs values, including {10, 15, 20}; hidden_size values, including {64, 128, 256}; and residual_channels and dilation_channels values, including {16, 32, 64}. Ultimately, through performance evaluation, the optimal parameter configuration was determined as follows: batch_size = 32, seq_len = 30, label_len = 10, train_epochs = 20, hidden_size = {64, 128}, residual_channels and dilation_channels = 32, kernel size in the gated TCN = 2, dropout = 0.2, encoder layers in the Informer = 2, decoder layers = 1, dropout = 0.05.

[0191] In the experiments, a Transformer-type network model was used, and the parameter settings are shown in Table 2. To prevent overfitting, we introduced an early stopping mechanism and an adaptive learning rate, with an initial learning rate of 0.0001, to improve the training efficiency of the model. After the model training was completed, we used the validation set to evaluate the optimal model and then applied the model to the test set for prediction.

[0192] Table 2 Parameter Model Settings

[0193]

[0194] Comparison of model selection and parameter settings:

[0195] To verify the effectiveness of the fitting model proposed in this invention, the proposed model is compared with the following benchmark models:

[0196] (1) Transformer

[0197] This model introduces a self-attention mechanism, which can effectively capture long-distance dependencies in sequences. It also has the ability to perform parallel computation, which improves the training speed of the model and achieves superior prediction performance.

[0198] (2) Autoformer

[0199] This model designs sequence decomposition units to embed into deep models, enabling progressive prediction. It also introduces an autocorrelation mechanism based on stochastic process theory, discarding the self-attention mechanism of point-wise connections and realizing the autocorrelation mechanism of series-wise connections.

[0200] (3)LSTM

[0201] LSTM is an excellent variant of RNN, inheriting most of the characteristics of RNN models, while solving the vanishing gradient problem caused by gradual reduction in gradient backpropagation.

[0202] (4) Reformer

[0203] This model is based on the self-attention mechanism of the Locality Sensitive Hashing (LSH) algorithm. At the same time, it uses reversible residual connections instead of traditional residual connections, which can achieve good prediction results.

[0204] The parameter settings for several comparative models are as follows: In the Transformer model, the basic parameter settings are consistent with those of the model proposed in this invention; in the Autoformer model, Moving_avg is set to 29, d_ff is set to 2048, and other basic parameter settings are consistent with those of the model proposed in this invention; in the LSTM model, a two-layer LSTM model is used, and hidden_size is set to 256. In the Reformer model, bucket_size is set to 4, n_hashes is set to 4, and other parameter settings are consistent with those of the model proposed in this invention.

[0205] Evaluation indicators:

[0206] This invention uses mean squared error (EMSE), mean absolute error (EMAE), and mean absolute percentage error (EMAPE) as indicators to evaluate the prediction accuracy of the model. These evaluation indicators can objectively reflect the degree of fit of the model to the real data, and the calculation formulas are as follows:

[0207]

[0208]

[0209]

[0210] In the formula, y is the predicted value at time t of the model proposed in this invention. i It is the actual value at time t.

[0211] II. Analysis of Experimental Results

[0212] Prediction results and discussion:

[0213] Experimental prediction results:

[0214] To further verify the accuracy of the model proposed in this invention in predicting engine exhaust temperature, the operating data of engines A and B during normal operation were used for verification.

[0215] The training set consists of 100 flights that operated normally with engine A from June 9, 2019 to June 2, 2020.

[0216] The training set consisted of 120 flights that operated normally with engine B from February 7, 2020 to August 20, 2020.

[0217] Four flights were randomly selected to verify the prediction of flight exhaust temperature. The results are as follows: Figure 9 As shown in Table 3;

[0218] As can be seen, the model proposed in this invention exhibits excellent predictive performance. Throughout all phases of the flight's operation, the predicted exhaust temperature can closely approximate the actual exhaust temperature.

[0219] It can maintain good predictive performance for different engines.

[0220] Table 3 Evaluation Indicators of the Model of the Invention

[0221]

[0222] The variation in prediction accuracy with different numbers of training samples is as follows: Figure 10 As shown.

[0223] Figure 11 The study presents a comparison of the predicted exhaust temperature for a flight on January 23, 2020, using engine A, and a flight on April 21, 2020, using engine B, under different models.

[0224] To more intuitively compare prediction accuracy, Table 4 and Figure 12 It shows a comparison of metrics for different models;

[0225] Taking EMAPE as an example, this illustrates the reduction in prediction error of the model of this invention compared to other models for different engines. Compared to the Transformer model, the EMAPE index of the model of this invention is reduced by 0.347% and 0.177%, respectively.

[0226] Compared to the Autoformer model, the EMAPE index decreased by 0.717% and 0.417%, respectively.

[0227] Compared with the LSTM model, the EMAPE index decreased by 0.557% and 0.277%, respectively; compared with the Reformer model, the EMAPE index decreased by 0.257% and 0.327%, respectively.

[0228] Experimental results show that the model proposed in this invention has the highest prediction accuracy and the best performance compared with other models, and the model has universality and accuracy.

[0229] Table 4 Comparison of Evaluation Indicators for Each Model

[0230]

[0231] Meanwhile, in order to verify the effectiveness of each functional module of the model of the present invention, three variant models can be obtained by removing the modules designed in the model. Three ablation experiments were designed for each variant model, and the engine A and B datasets were used as examples for comparison and illustration.

[0232] To verify the importance of feature extraction, the experiment kept other modules of the model unchanged and made predictions under the condition that other input features were the same.

[0233] Where Model* represents the experimental model with the GCN feature extraction module removed;

[0234] Model** represents the experimental model with the feature extraction module of G-TCN removed;

[0235] Model*** indicates an experimental model in which the BiLSTM module was removed from the Informer encoder;

[0236] The comparison results are as follows Figure 13 As shown in Table 5, the evaluation indicators are as follows.

[0237] Table 5 Comparison of evaluation indicators for ablation experiments

[0238]

[0239] As shown in Table 5, the EMAPE of the model of this invention is reduced by 0.137% and 0.077% compared with the Model* model. It can be seen that the improved design using the GCN feature extraction module makes the model more adaptable to flight phase transitions.

[0240] Compared with the Model** model, the EMAPE of the present invention was reduced by 0.287% and 0.137%, respectively. It can be found that the G-TCN feature extraction module can better focus on the short-term proximity and long-term dependence of flight data.

[0241] Compared with the Model*** model, the proposed model reduces EMAPE by 0.107% and 0.027%, respectively. Introducing a BiLSTM module in the encoder part can better capture local correlations in the feature sequence.

[0242] Based on the above results, the prediction model built in this invention has better prediction performance than other methods, and its module design is reasonable, so it can be used as a threshold indicator for judging abnormal exhaust temperature.

[0243] Abnormal threshold:

[0244] This invention uses normal operating data of aero-engines to train the model, outputs predicted exhaust temperature values, calculates the difference between the predicted and actual exhaust temperatures, and uses the Kolmogorov-Smirnov test to verify whether the data follows a normal distribution.

[0245] The results show that the pvalue of EGTpre_dif is 0.283 > 0.05, accepting the null hypothesis that EGT... pre_dif It follows a normal distribution.

[0246] The confidence intervals and maximum values ​​are shown in Table 6.

[0247] Table 6. Numerical Analysis of Exhaust Temperature Difference

[0248]

[0249] Based on the normal distribution results and the symmetry of the data distribution, this invention uses the difference between the actual exhaust temperature and the predicted exhaust temperature as a benchmark to set the exhaust temperature prediction difference EGT. minus =4.92℃ is set as the alarm threshold for abnormal exhaust temperature. When the difference between the predicted exhaust temperature and the actual value is greater than the set threshold, it is considered that there is an abnormal exhaust temperature in the engine air circuit.

[0250] Abnormal diagnostic analysis:

[0251] For the abnormal exhaust temperature fault, data from seven flights involving the faulty engine A in Case 1, from March 26 to May 18, 2020, were selected for verification. Figure 14 As shown, the exhaust temperature remained stable during flights from March 26th to March 28th without any abnormal fluctuations. A potential fault period began on March 29th. The method of this invention accurately detected the abnormal fluctuations during the Taxi-out phase of the March 29th flight, identifying early signs of a fault and issuing an early warning. However, in reality, after several subsequent flights experiencing abnormal fluctuations, the airline only discovered the abnormal exhaust temperature of the left engine (A) on May 18th. During flights where the airline did not perform regular maintenance, this posed a significant flight hazard. The anomaly detection model proposed in this invention can promptly detect early signs of a fault during the Taxi-out phase of a flight, providing a basis for airlines to develop maintenance plans.

[0252] The model built using this invention was used to verify the abnormal exhaust temperature of the faulty engine B in Case 2. The data consisted of three flight records from June 22nd to June 24th, 2020. According to maintenance records, this aircraft did not exhibit any abnormalities in flights prior to June 24th, 2020. During flight on June 24th, it experienced a loss of control, and the engine responsible for the accident was identified as engine B. Figure 15 As can be seen, engine B remained stable during the flight on June 22nd without any abnormal fluctuations. The method of this invention detected a significant increase in exhaust temperature on June 23rd. Although the temperature returned to normal during subsequent flights on June 23rd, the exhaust temperature decreased for a period shortly after takeoff on June 24th, and extreme temperature anomalies occurred in subsequent flights, posing a significant flight hazard. This verifies that the proposed model can not only detect potential gas path faults in aero engines but also provide early warning of engine runaway.

[0253] Robustness, reliability, and superiority verification:

[0254] An effective anomaly detection model should have the following three characteristics:

[0255] 1) Robustness: The model can accurately diagnose air circuit faults in other engines.

[0256] As can be observed from the anomaly detection results of Case 1 and Case 2 in Section 4.3, the method proposed in this invention can issue timely and effective early warnings for abnormal fluctuations in engine exhaust temperature of different aircraft, thus verifying the robustness of the method.

[0257] 2) Reliability: The model should not issue false warnings to normal engines.

[0258] Two regular flights were randomly selected from the two cases for verification, and the results are as follows. Figure 16 As shown, no abnormalities were found in the exhaust system of the aircraft engines on flights A (January 23, 2020) and B (April 21, 2020). Using the proposed anomaly detection method, all exhaust temperatures were within normal operating ranges. This demonstrates that the proposed anomaly detection method does not issue false warnings, verifying its reliability.

[0259] 3) Superiority: Compared with other methods, this model has higher prediction accuracy.

[0260] The comparison results of different models show that the model proposed in this invention has the highest prediction accuracy, which verifies the superiority of the proposed method.

[0261] Meanwhile, the ablation experiments of each module of the model proposed in this invention also verified the necessity of spatiotemporal state fusion.

[0262] The insufficient accuracy of other models may be attributed to the fact that the comparison methods do not consider the influence of space state parameters on the prediction of aero-engine exhaust temperature, and rely heavily on the nonlinear mapping relationship and different periodic variation relationships of time series data.

[0263] in conclusion

[0264] This invention proposes a deep learning framework that integrates spatial and temporal features to address the problem of detecting anomalies in the aero-engine's airflow performance. This method fully considers the influence of spatial states such as aircraft attitude changes, airspeed, and altitude during actual flight, as well as the short-term proximity and long-term dependence of other parameters in the temporal dimension. The BiLSTM module introduced in the encoder section enhances the local correlation in the feature sequence. Validation with actual flight data samples demonstrates that the proposed model is a high-precision and high-stability method. Simultaneously, it effectively enables online monitoring of early fault symptoms. This provides a basis for airline maintenance decision-making, reduces unplanned downtime, and has certain engineering application value.

[0265] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting anomalous performance of aero-engine airflow path by integrating spatial and temporal characteristics, characterized in that, Includes the following steps: S1. Data Acquisition; S2. Modeling is performed to address the problem of detecting abnormal gas path performance in aero-engines; S3. Establish an exhaust gas temperature (EGT) anomaly detection model that integrates spatial and temporal characteristics; S4. Perform spatial feature extraction based on GCN; S5. Perform temporal feature extraction based on a gated temporal convolutional neural network; S6. Improve the Informer prediction model; S7. Set the abnormal threshold; S8. Perform data processing and experimental setup; S9. Analyze the experiment and draw conclusions; Step S1 uses actual flight data of a certain airline's B737. The engine model used in the aircraft is CFM56_7B, and the data comes from the B737's fast access recorder. The implementation steps of step S2 are as follows: S2-1. Define the exhaust temperature of an aircraft engine as... Historical time series matrix, The values ​​at T time steps within the fixed input window are ,and The relevant covariates are Where N is the number of covariates, and each submatrix They are all sequences that change over time, representing sequences. The relevant covariate matrix includes multidimensional information on airspeed, flight altitude, fuel flow rate, and rotor speed; S2-2. Goal setting: The goal is to learn a model. According to covariates Predicting the exhaust temperature one time step in the future , These are the parameters of the model; S2-3. Predicted exhaust temperature With actual exhaust temperature Error calculations were performed, and anomaly detection of aero-engine exhaust temperature was conducted based on the proposed monitoring strategy. The implementation steps of step S3 are as follows: S3-1. Use the preprocessing module to preprocess historical data in the QAR of aero-engines and filter relevant features; S3-2. Feed the processed relevant parameter sequence into the GCN and gated TCN modules to extract the corresponding features, and then predict the normal EGT based on the improved Informer; S3-3. Compare the predicted normal EGT with the actual EGT, and determine the alarm level based on the proposed fault prediction strategy; S3-4. Locate potential faults and issue early warnings based on alarm levels; S3-5. During aircraft operation, repeat steps S3-1 to S3-4 for newly input data; The implementation steps of step S6 are as follows: S6-1. Based on Informer, introduce a BiLSTM module in the encoder section; S6-2. The Probesparse self-attention mechanism in Informer proposes a query sparsity measure based on Kullback-Leibler divergence, allowing the Key to focus only on the top u most important queries, as shown in formula (12): （12）； In the formula, It is a sparse matrix containing only the first u queries; S6-3. Replace the traditional self-attention mechanism with a probable sparse self-attention mechanism. Only a portion of representative samples are selected from the input sequence for computation. The self-attention distillation method is used to reduce the network dimension, thereby further reducing the computational complexity. S6-4. The self-attention distillation layer consists of a one-dimensional convolution, using ELU as the activation function. After convolution, it passes through a pooling layer, concatenates the feature maps of all multi-head self-attention layers, and then inputs them into the Informer decoder. S6-5. Construct a decoder consisting of a multi-head probabilistic sparse self-attention layer, a full attention layer, and a fully connected layer; S6-6. Introduce a BiLSTM layer in the Informer encoder section. At time t, the calculation formulas for the hidden layer and the output layer in both directions are as shown in formulas (13), (14), and (15): （13）； （14）； （15）； Output value By combining two LSTM units, a weighted fully connected structure is used to further extract the temporal information of the Informer encoder output features.

2. The method for detecting anomalous performance of aero-engine gas path by integrating spatial and temporal characteristics according to claim 1, characterized in that, The implementation steps of step S4 are as follows: S4-1. GCN is used to extract spatial features from the spatial state parameters of the aircraft's attitude change, airspeed, and flight altitude. The spatial state parameters of the aero-engine are used as nodes in the graph, including airspeed, flight altitude, and pitch angle. The relationships between nodes are established by the interaction and dependency between the state parameters. S4-2. Construct a graph structure ,in The set of all spatial state parameter nodes. Represents the set of all edges, using GCN to process graph structures. Feature learning allows us to obtain feature representations for each time point, thereby extracting the spatial relationships of spatial state parameters. Graph convolutional layers utilize graph-structured adjacency matrices. As prior knowledge, spatial dependencies during flight are simulated by extracting pairwise relationships between state parameters, using the adjacency matrix. The definition is as shown in formula (1), where, Represents different spatial state parameters at each time point. The magnitude of the correlation, Relevance threshold: （1）； S4-3. The Pearson correlation coefficient is used to characterize the correlation between different spatial state parameters, as shown in formula (2): （2）； In the formula, The closer the value is to 1, the stronger the correlation between the two sets of data. The correlation between the two parameters is evaluated by the absolute correlation coefficient, ranging from 0.8 to 1.0, 0.6 to 0.8, 0.4 to 0.6, 0.2 to 0.4, and 0 to 0.2, representing extremely strong, strong, moderate, weak, and extremely weak correlations, respectively. This step involves setting the total correlation threshold. Set to 0.4; S4-4. Treat the state parameters at each time point on the graph as graph signals. graph convolution kernel The graph convolution is defined by formula (3): （3）； In the formula, For graph convolution operators, For the Laplace matrix of the graph, use symmetric normalization, as shown in formula (4): （4）； In the formula, Let A be the identity matrix, A be the adjacency matrix, and D be the diagonal matrix; In formula (3) Through Laplace eigenvalue decomposition The obtained Fourier transform basis, where For the eigenvalue diagonal matrix, a graph convolution kernel is constructed using Chebyshev polynomials to simplify the calculation process, as shown in formula (5): （5）； In the formula, It is a k-th order Chebyshev polynomial. , The largest eigenvalue of the Laplace matrix. The kernel size; S4-5. If only the first-order Chebyshev polynomial is considered, it is further simplified to a first-order approximation and a self-loop is added. At the same time, the nonlinearity of the convolution result is enhanced by the activation function. The characteristics of the flight state parameters after graph convolution are given by formula (6): （6）； In the formula, , for The degree matrix, , The convolutional result is the first... Layer feature matrix, For the first Layer parameters, This is the activation function.

3. The method for detecting anomalous performance of aero-engine air path based on the fusion of spatial and temporal characteristics according to claim 2, characterized in that, The implementation steps of step S5 are as follows: S5-1. One-dimensional convolution is improved by using causal convolution, dilated convolution, and residual connections; S5-2. For Dimensional gas path input parameters dilated convolution Defined as formula (7): （7）； In the formula, For expansion rate, The kernel size is [size]. For filters; Meanwhile, residual connections improve the efficiency of information propagation through cross-layer direct connections. TCN enhances the network's learning ability by introducing residual connections, making the structure more stable. S5-3. A gating mechanism is added between the dilated causal convolutions of TCN. The gated temporal convolution unit mainly uses the gated activation unit to dynamically adjust the input signal to determine the probability of the information of this layer being passed to the next convolutional layer. At the same time, the convolutional structure does not need a forget gate, and only the output gate is considered to control the propagation of signals between layers. The gated temporal convolution unit consists of two parts, as shown in formula (8): （8）； In the formula, and It is a parameter matrix, where b and c are the learned model parameters; This refers to point-by-point multiplication of matrices. The sigmoid activation function filters out useless information. In the formula, and It is a parameter matrix, where b and c are the learned model parameters; This refers to point-by-point multiplication of matrices. Use the sigmoid activation function to filter out useless information; S5-4. For multidimensional gas path time series data Where T is the historical time window and P is the feature dimension, according to formula (7), the causal convolution with the i-th parameter can be defined as formula (9): （9）； In the formula, This represents the convolution result of the i-th parameter at time t, where d is the dilation rate. For the elements of the convolution kernel, all of them in the formula Constructing convolution kernels In the formula, K is the kernel size and F is the number of output channels; S5-5. Combining formula (8), the time parameter features extracted by multi-layer gated TCN can be expressed as formula (10): （10）； In the formula, For the first Output characteristics of layer-gated TCN For the first The convolution kernel of the layer, Indicates the first Tensor causal convolution with layer dilation rate d This is the sigmoid activation function, used to control the propagation ratio between layers.

4. The method for detecting anomalous performance of aero-engine air path based on the fusion of spatial and temporal characteristics according to claim 3, characterized in that, The implementation steps of step S7 are as follows: select a Gaussian distribution to map the accuracy of the prediction error, define the error X as following a normal distribution with expected value u and variance σ, denoted as N(u,σ2), and the probability density function p(x) of the Gaussian distribution is given by formula (16): （16）； The threshold is calculated based on a 99% confidence level of the error using the validation dataset, which means that the false alarm rate of the predicted exhaust temperature under normal operating conditions is 1%.

Images