An aero-engine gas path abnormality monitoring method based on flight data

Abstract

The application provides an aero-engine gas path abnormality monitoring method based on flight data, comprising the following steps: gas path feature extraction based on kernel slow feature analysis: slow feature analysis principle and kernel slow feature analysis principle; gas path performance evaluation based on deep support vector description algorithm; aero-engine gas path abnormality determination based on box plot; aero-engine abnormality monitoring model based on KSFA-Deep SVDD; data set selection and parameter setting. The aero-engine gas path abnormality monitoring method based on flight data provided by the application can reflect the change process of the gas path to a certain extent, reduces the difficulty of subsequent abnormality monitoring, and the gas path data has a very high dimension in the actual aero-engine operation process. The slow feature analysis adopts a polynomial expansion method for nonlinear processing, and when the dimension is high, a dimension disaster is formed, which greatly affects the feature extraction speed and the accuracy of the actual subsequent modeling.

Description

Technical Field

[0001] This invention relates to the field of aero-engines, and more particularly to a method for monitoring aero-engine air path anomalies based on flight data. Background Technology

[0002] An aircraft engine is a highly complex and precise thermodynamic machine. As the heart of an aircraft, it not only powers the aircraft's flight but also serves as a vital driving force for the development of the aviation industry. Every major transformation in the history of human aviation has been inseparable from the technological advancements in aircraft engines.

[0003] Aircraft engines are an important power source for aircraft. Operating under harsh conditions for extended periods, their safety and reliability face significant challenges. Gas system components are constantly exposed to high temperatures and pressures, making them highly susceptible to early failures. This is one of the main factors leading to dangerous situations such as in-flight engine shutdowns. However, early gas system failures are often masked by various unknown interferences and noises, making it difficult to effectively detect early anomalies.

[0004] Currently, methods for monitoring aero-engine gas path anomalies can be categorized into two types: model-based methods and data-driven methods. Model-based methods require the establishment of an accurate physical and mathematical model of the engine to describe its operating characteristics, and rely on sufficient sensor measurement information to correct the model characteristics to improve model accuracy. The accuracy of the established mechanistic model directly affects the diagnostic effect.

[0005] The performance of the gas path will slowly degrade over time. It is extremely important to capture the original value characteristics of the gas path parameters that reflect the changes in gas path performance. The above methods are not well applied to early faults. The main reason is that the relevant algorithms cannot eliminate interference signals in weak signals and it is difficult to capture changes in early gas path faults. Due to damage accumulation, structural deformation and other reasons, the gas path components of aero engines are affected.

[0006] Therefore, it is necessary to provide a method for monitoring aero-engine air path anomalies based on flight data to solve the above-mentioned technical problems. Summary of the Invention

[0007] This invention provides a method for monitoring aero-engine air path anomalies based on flight data. It addresses the problem that aero-engines are important power sources for aircraft, and their safety and reliability face great challenges when operating under harsh conditions for extended periods. Air path components are constantly under high temperature and pressure, making them prone to early failures, which can lead to dangerous situations such as in-flight engine shutdowns.

[0008] To address the aforementioned technical problems, this invention provides a method for monitoring aero-engine air path anomalies based on flight data, comprising the following steps:

[0009] S1. Gas path feature extraction based on kernel slow feature analysis: slow feature analysis principle and kernel slow feature analysis principle;

[0010] S2. Performance evaluation of gas path based on deep support vector description algorithm: principle of traditional support vector description algorithm, principle of deep support vector description algorithm and Deep SVDD network structure and training;

[0011] S3. Determination of aero-engine gas path anomalies based on box plots;

[0012] S4. An aero-engine anomaly monitoring model based on KSFA-Deep SVDD;

[0013] S5. Dataset Selection and Parameter Setting: Introduction to experimental datasets, parameter selection, dataset preprocessing, DeepSVDD parameter setting, and selection and parameter setting of comparison models;

[0014] S6. Experimental Results and Discussion: Slow Feature Extraction.

[0015] Preferably, the slow characteristic analysis principle in S1 includes the following: During the operation of an aero-engine, the performance of the air path changes slowly over time. Slow characteristic analysis is proposed to find a set of input-output functions that make the output signal change as slowly as possible. The features extracted by the slow characteristic analysis algorithm can effectively reflect the slow change of the air path state, and the magnitude of the change can reflect the change state of the air path.

[0016] Given an n-dimensional gas path input signal Considering a linear mapping function, we can find the mapping relationship between the objective function and the input function using SFA, i.e.

[0017] (1)

[0018] In the formula, Represents the weight of the j-th function;

[0019] Accordingly, the objective function of the SFA mapping function is:

[0020] (2)

[0021] In the formula, Represents the difference value. Represents the time difference of s. It is the average value of s along the time axis;

[0022] Furthermore, the optimization problem of the objective function should satisfy the following conditions:

[0023] (3)

[0024] (4)

[0025] (5)

[0026] Finally, it can be reconstructed into the following generalized eigenvalue decomposition problem:

[0027] (6)

[0028] In the formula, , . , It is a diagonal matrix composed of generalized eigenvalues. After solving equation (6) and obtaining the transformation matrix W, the characteristic S reflecting the slow change of the gas path can be expressed as:

[0029] .

[0030] Preferably, the principle of the slow feature analysis in S1 includes the following: In the actual operation of aero-engines, the air path data has extremely high dimensionality. Slow feature analysis often uses the polynomial expansion method to process it. However, the air path sensor data of aero-engines has a large dimensionality, and polynomial expansion will cause the curse of dimensionality, which will greatly affect the feature extraction speed and the accuracy of subsequent modeling.

[0031] Kernel tricks are a method for processing nonlinear data. Kernel slow feature analysis introduces kernel functions to replace the polynomial extensions in slow feature analysis, thereby enabling the processing of nonlinear data. This involves mapping the original data in the low-dimensional space to a high-dimensional space and transforming the inner product operations in the high-dimensional space into kernel function operations in the input space. Various kernel functions are commonly used; this invention employs the Gaussian kernel function, which has strong generalization and smoothing estimation capabilities. Through multiple experiments, a Gaussian kernel size of 5 was set.

[0032] (8)

[0033] In the formula, and For any sample.

[0034] Replacing the polynomial expansion with the Gaussian kernel function, i.e., replacing the second-order polynomial expansion with the expanded data K, the objective function of the mapping function in equation (2) becomes:

[0035] (9)

[0036] The generalized eigenvalue decomposition problem in equation (6) is recast as:

[0037] (10)

[0038] In the formula, for The covariance matrix, Let K be the covariance matrix;

[0039] Solving the generalized eigenvalue decomposition problem in equation (10) yields the transformation matrix W, and the kernel slow eigenvalue can be expressed as:

[0040] (11).

[0041] Preferably, the traditional support vector description algorithm in S2 includes the following principles: In the actual operation of an aero-engine, the ratio of normal samples to abnormal samples is unbalanced, resulting in a sample imbalance problem. High-quality abnormal labeled samples are extremely difficult to obtain. Unsupervised learning algorithms are suitable for situations where both normal and abnormal data exist but no labels are available. When only healthy state data is available, support vector data description can be trained based on normal aero-engine samples and effectively distinguish between normal and abnormal engine states during subsequent operation. The performance evaluation process of the traditional support vector description algorithm is as follows:

[0042] Given an n-dimensional slow feature extracted by KSFA SVDD performs a nonlinear mapping on the samples to find an optimal hypersphere in the high-dimensional feature space that encloses all training samples. The corresponding objective function is:

[0043] (12)

[0044] In the formula, Let O be the radius of the hypersphere, and O be the center. As slack variables, As a penalty factor;

[0045] By introducing the solution formula (12) for the Lagrange function, the original problem can be rewritten as the following dual description:

[0046] (13)

[0047] In the formula, For kernel function operations, the kernel function selected in this invention is the Gaussian kernel function, which has strong generalization and smoothing estimation capabilities;

[0048] Solving equation (13) yields the result that the center of the hypersphere is... and hypersphere radius :

[0049] (14)

[0050] (15)

[0051] Define the test sample for any time. To the center of the supersphere The square of the distance is the health indicator (HI):

[0052] (16).

[0053] Preferably, the deep support vector description algorithm in S2 includes the following principles: Traditional support vector description mainly uses a single-layer nonlinear mapping for a given training sample and uses kernel functions to complete feature extraction. However, single-layer feature extraction is often difficult to uncover the complex relationships between nonlinear data. In response to the shortcomings of traditional support vector description, deep support vector description has been proposed in recent years. In deep SVDD, the goal is the same, but instead of using a linear mapping in a high-dimensional space, a deep neural network is used to learn the nonlinear representation of the input data in a low-dimensional latent space.

[0054] For slow features extracted by KSFA The feature extraction process is represented as follows:

[0055] (17)

[0056] In the formula, Let L be the nonlinear mapping function of the Lth layer in a deep neural network. These are the weights of the Lth layer network;

[0057] Unlike the objective function of traditional SVDD, the optimization objective of deep SVDD models is to jointly learn the network parameters W while minimizing the amount of data enclosing the hypersphere in the output space, so that the final features in the output space are maximized. The objective function is as follows: The object should be enclosed within an optimal hypersphere as much as possible.

[0058] (18)

[0059] In the formula, R is the radius of the hypersphere, o is the a priori center of the hypersphere, and v is the equilibrium parameter. Let W be the penalty factor, and W be the set of network parameters, denoted as... In anomaly detection, the training samples are usually normal samples, and the above objective function can be simplified:

[0060] (19)

[0061] The network weight parameters W are trained using a deep network optimization algorithm to obtain the trained weights. For a given test sample Engine health index (HI) is defined as Features The square of the distance to the center o of the hypersphere is:

[0062] (20).

[0063] Preferably, the Deep SVDD network structure and training in S2 include the following methods: the slow features extracted by KSFA reflect the characteristics of the gas path parameter data changing over time; convolutional neural networks are often used for multi-array signals such as time series signals due to their strong generalization ability; at the same time, the feature attention mechanism can select the information that is more critical to the current task objective from many features. Therefore, for the aero-engine anomaly monitoring task, this invention designs an encoder structure based on 1D-CNN fusion feature attention mechanism for use in Deep SVDD to learn the nonlinear representation of the input data in the low-dimensional latent space.

[0064] The encoder consists of two one-dimensional convolutional layers and one fully connected layer. The specific parameter settings for each layer are as follows:

[0065] The first convolutional layer has 1 input channel and 32 output channels, using a 3x3 convolutional kernel and 1 padding pixel. The first pooling layer has a convolutional kernel size of 2 and a stride of 2, halving the length of the input.

[0066] The second convolutional layer has 32 input channels and 64 output channels. It also uses a 3x3 convolutional kernel and 1 padding pixel. The second pooling layer also has 2 convolutional kernels and a stride of 2.

[0067] The feature extraction layer also includes a Flatten layer to flatten the tensor output by the convolutional layer into a one-dimensional tensor, and a fully connected layer. The feature attention mechanism is implemented using a sigmoid function, with the input being the output of the fully connected layer. Assuming the output of the feature attention mechanism is w, the final feature representation... for , where z is the encoder output and c is the center;

[0068] The training framework for Deep SVDD networks mainly consists of two stages: an autoencoder pre-training stage and a Deep SVDD training stage. The goal is to obtain the optimal Deep SVDD network weights. and the center of the supersphere .

[0069] Preferably, the aero-engine air path anomaly determination based on box plot in S3 includes the following methods: Box plot, as a commonly used outlier screening method, can calculate the normal and abnormal ranges of the obtained aero-engine air path performance health factors.

[0070] The box length of a box plot is the interquartile range (IQR), which is expressed as:

[0071] (twenty one)

[0072] and 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:

[0073] (twenty two)

[0074] In the formula, , These are the lower quartile and upper quartile, respectively, and k is the outlier coefficient. In this invention, k=3 is used to monitor extreme outliers of health factors.

[0075] The proposed method for monitoring aero-engine anomalies consists of an offline phase and an online phase, with the specific steps as follows:

[0076] Offline phase

[0077] A. Perform kernel slow feature analysis on the normal airway training data to extract slow change features;

[0078] B. Construct a Deep SVDD model based on the slow characteristics of the normal state, and obtain the model parameters;

[0079] C. Calculate health indicators for normal-state data using Deep SVDD parameters to describe the process of aero-engine performance degradation. To improve the sensitivity of the HI indicator to describing aero-engine degradation, EWMA is added for smoothing based on HI.

[0080] (twenty three)

[0081] In the formula: The mean of the first t data points. The smoothing exponent is typically between 0.05 and 0.25. A value of 0.15 can reflect the degradation trend well;

[0082] D. Use the box plot method to calculate the threshold of the training data.

[0083] This invention uses box plots to derive thresholds from health indicators obtained from training samples under normal conditions of different models, and applies these thresholds as the normal state thresholds to subsequent online phase monitoring. The magnitude of the thresholds will vary depending on the magnitude of the health indicators for different engines on different models. However, each engine can establish its own threshold based on its normal state, creating a "personalized" early fault symptom alarm threshold for each engine.

[0084] Preferably, the KSFA-Deep SVDD-based aero-engine anomaly monitoring model in S4 includes the following methods: (twenty four)

[0085] In the formula: This represents the upper quartile of the health factor in a normal state. The interquartile range is denoted as , and k is the outlier coefficient. In this invention, k=3 is used to monitor extreme outliers of health factors.

[0086] Online phase

[0087] A. Perform kernel slow feature analysis on the test data during flight to extract slow change features;

[0088] B. Calculate the distance between the test data and the center of the Deep SVDD hypersphere in a normal state, and use this as a health indicator;

[0089] C. Monitor the health indicators of the aircraft engine based on the threshold. If three consecutive points exceed the threshold, an early warning will be issued.

[0090] Preferably, the parameter selection in S5 includes the following methods: QAR data contains a large number of different performance parameters, but not all parameters are suitable for health status assessment. Since some parameters are not closely related to the health performance of the engine, and the data information of some parameters is incomplete, it is necessary to select parameters that are fully recorded and closely related to the state of the engine when conducting engine health performance assessment. In order to determine the characteristic parameters related to EGT, the Pearson correlation coefficient is used for correlation analysis, as shown in equation (25).

[0091] (25)

[0092] In the formula, The closer the value is to 1, the stronger the correlation between the two sets of data. The correlation between two parameters can be 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 correlation, respectively.

[0093] Preferably, the dataset preprocessing in S5 includes the following: selecting the cruise phase data of each flight as the analysis object; since the data volume is very large, an average value is taken every 600 seconds; at the same time, since the maximum and minimum values ​​cannot be obtained during real-time status monitoring, the present invention uses the z-score normalization method to normalize the real-time status parameters. The formula is as follows:

[0094] (26).

[0095] Compared with related technologies, the air path anomaly monitoring method for aero-engines based on flight data provided by this invention has the following beneficial effects:

[0096] This invention provides a method for monitoring aero-engine air path anomalies based on flight data. The air path features extracted by slow feature analysis can reflect the changes in the air path to a certain extent, reducing the difficulty of subsequent anomaly monitoring. However, in the actual operation of aero-engines, the air path data has extremely high dimensionality. Slow feature analysis often uses polynomial expansion methods to process it nonlinearly. When the dimensionality is high, it will lead to the curse of dimensionality, which will greatly affect the feature extraction speed and the accuracy of subsequent modeling. Considering the characteristics of aero-engine air path parameter data, kernel slow feature analysis with kernel function expansion can enhance its feature extraction capability and achieve effective extraction of feature information. Attached Figure Description

[0097] Figure 1 This is a schematic diagram of a preferred embodiment of an aero-engine air path anomaly monitoring method based on flight data provided by the present invention.

[0098] Figure 2 This is a schematic diagram of the Deep SVDD network training structure;

[0099] Figure 3 This is a diagram illustrating the working principle of a box plot.

[0100] Figure 4 A schematic diagram of an aero-engine anomaly monitoring model based on KSFA-Deep SVDD;

[0101] Figure 5 This is a schematic diagram of the correlation analysis results;

[0102] Figure 6 This is a schematic diagram of the parameter variation curves for the left engine;

[0103] Figure 7 A schematic diagram of the extracted kernel slow features;

[0104] Figure 8 This is a schematic diagram of gas path monitoring based on KSFA-Deep SVDD;

[0105] Figure 9 The diagram illustrates the monitoring effects of different methods on Case 1.

[0106] Figure 10 This is a schematic diagram of gas path monitoring based on KSFA-Deep SVDD;

[0107] Figure 11 This diagram illustrates the monitoring effects of different methods on Case 2. Detailed Implementation

[0108] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0109] Please refer to the following: Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7 , Figure 8 , Figure 9 , Figure 10 and Figure 11 ,in, Figure 1 This is a schematic diagram of a preferred embodiment of an aero-engine air path anomaly monitoring method based on flight data provided by the present invention. Figure 2 This is a schematic diagram of the Deep SVDD network training structure; Figure 3 This is a diagram illustrating the working principle of a box plot. Figure 4 A schematic diagram of an aero-engine anomaly monitoring model based on KSFA-Deep SVDD; Figure 5 This is a schematic diagram of the correlation analysis results; Figure 6 This is a schematic diagram of the parameter variation curves for the left engine; Figure 7 A schematic diagram of the extracted kernel slow features; Figure 8 This is a schematic diagram of gas path monitoring based on KSFA-Deep SVDD; Figure 9 The diagram illustrates the monitoring effects of different methods on Case 1. Figure 10 This is a schematic diagram of gas path monitoring based on KSFA-Deep SVDD; Figure 11 This diagram illustrates the monitoring effects of different methods on Case 2. A method for monitoring aero-engine airflow anomalies based on flight data includes the following steps:

[0110] S1. Gas path feature extraction based on kernel slow feature analysis: slow feature analysis principle and kernel slow feature analysis principle;

[0111] S2. Performance evaluation of gas path based on deep support vector description algorithm: principle of traditional support vector description algorithm, principle of deep support vector description algorithm and Deep SVDD network structure and training;

[0112] S3. Determination of aero-engine gas path anomalies based on box plots;

[0113] S4. An aero-engine anomaly monitoring model based on KSFA-Deep SVDD;

[0114] S5. Dataset Selection and Parameter Setting: Introduction to experimental datasets, parameter selection, dataset preprocessing, DeepSVDD parameter setting, and selection and parameter setting of comparison models;

[0115] S6. Experimental Results and Discussion: Slow Feature Extraction.

[0116] The slow characteristic analysis principle in S1 includes the following: During the operation of an aero-engine, the performance of the air path changes slowly over time. Slow characteristic analysis is proposed to find a set of input-output functions that make the output signal change as slowly as possible. The features extracted by the slow characteristic analysis algorithm can effectively reflect the slow change of the air path state, and the magnitude of the change can reflect the change state of the air path.

[0117] Given an n-dimensional gas path input signal Considering a linear mapping function, we can find the mapping relationship between the objective function and the input function using SFA, i.e.

[0118] (1)

[0119] In the formula, Represents the weight of the j-th function;

[0120] Accordingly, the objective function of the SFA mapping function is:

[0121] (2)

[0122] In the formula, Represents the difference value. Represents the time difference of s. It is the average value of s along the time axis;

[0123] Furthermore, the optimization problem of the objective function should satisfy the following conditions:

[0124] (3)

[0125] (4)

[0126] (5)

[0127] Finally, it can be reconstructed into the following generalized eigenvalue decomposition problem:

[0128] (6)

[0129] In the formula, , . , It is a diagonal matrix composed of generalized eigenvalues. After solving equation (6) and obtaining the transformation matrix W, the characteristic S reflecting the slow change of the gas path can be expressed as:

[0130] .

[0131] The principle of slow feature analysis in S1 includes the following: In the actual operation of aero-engines, the air path data has extremely high dimensionality. Slow feature analysis often uses the polynomial expansion method to process it. However, the air path sensor data of aero-engines has a large dimensionality, and polynomial expansion will cause the curse of dimensionality, which will greatly affect the feature extraction speed and the accuracy of subsequent modeling.

[0132] Kernel tricks are a method for processing nonlinear data. Kernel slow feature analysis introduces kernel functions to replace the polynomial extensions in slow feature analysis, thereby enabling the processing of nonlinear data. This involves mapping the original data in the low-dimensional space to a high-dimensional space and transforming the inner product operations in the high-dimensional space into kernel function operations in the input space. Various kernel functions are commonly used; this invention employs the Gaussian kernel function, which has strong generalization and smoothing estimation capabilities. Through multiple experiments, a Gaussian kernel size of 5 was set.

[0133] (8)

[0134] In the formula, and For any sample.

[0135] Replacing the polynomial expansion with the Gaussian kernel function, i.e., replacing the second-order polynomial expansion with the expanded data K, the objective function of the mapping function in equation (2) becomes:

[0136] (9)

[0137] The generalized eigenvalue decomposition problem in equation (6) is recast as:

[0138] (10)

[0139] In the formula, for The covariance matrix, Let K be the covariance matrix;

[0140] Solving the generalized eigenvalue decomposition problem in equation (10) yields the transformation matrix W, and the kernel slow eigenvalue can be expressed as:

[0141] (11).

[0142] The traditional support vector description algorithm in S2 follows these principles: In the actual operation of an aero-engine, the ratio of normal to abnormal samples is unbalanced, resulting in a sample imbalance problem. High-quality abnormal labeled samples are extremely difficult to obtain. Unsupervised learning algorithms are suitable for situations where both normal and abnormal data exist but no labels are available. When only healthy state data is available, support vector data description can be trained based on normal aero-engine samples and effectively distinguish between normal and abnormal engine states during subsequent operation. The performance evaluation process of the traditional support vector description algorithm is as follows:

[0143] Given an n-dimensional slow feature extracted by KSFA SVDD performs a nonlinear mapping on the samples to find an optimal hypersphere in the high-dimensional feature space that encloses all training samples. The corresponding objective function is:

[0144] (12)

[0145] In the formula, Let O be the radius of the hypersphere, and O be the center. As slack variables, As a penalty factor;

[0146] By introducing the solution formula (12) for the Lagrange function, the original problem can be rewritten as the following dual description:

[0147] (13)

[0148] In the formula, For kernel function operations, the kernel function selected in this invention is the Gaussian kernel function, which has strong generalization and smoothing estimation capabilities;

[0149] Solving equation (13) yields the result that the center of the hypersphere is... and hypersphere radius :

[0150] (14)

[0151] (15)

[0152] Define the test sample for any time. To the center of the supersphere The square of the distance is the health indicator (HI):

[0153] (16).

[0154] The principle of the deep support vector description algorithm in S2 includes the following: Traditional support vector description mainly uses a single-layer nonlinear mapping for a given training sample and uses kernel functions to complete feature extraction. However, single-layer feature extraction is often difficult to uncover the complex relationships between nonlinear data. In response to the shortcomings of traditional support vector description, deep support vector description has been proposed in recent years. In deep SVDD, the goal is the same, but instead of using a linear mapping in a high-dimensional space, a deep neural network is used to learn the nonlinear representation of the input data in a low-dimensional latent space.

[0155] For slow features extracted by KSFA The feature extraction process is represented as follows:

[0156] (17)

[0157] In the formula, Let L be the nonlinear mapping function of the Lth layer in a deep neural network. These are the weights of the Lth layer network;

[0158] Unlike the objective function of traditional SVDD, the optimization objective of deep SVDD models is to jointly learn the network parameters W while minimizing the amount of data enclosing the hypersphere in the output space, so that the final features in the output space are maximized. The objective function is as follows: The object should be enclosed within an optimal hypersphere as much as possible.

[0159] (18)

[0160] In the formula, R is the radius of the hypersphere, o is the a priori center of the hypersphere, and v is the equilibrium parameter. Let W be the penalty factor, and W be the set of network parameters, denoted as... In anomaly detection, the training samples are usually normal samples, and the above objective function can be simplified:

[0161] (19)

[0162] The network weight parameters W are trained using a deep network optimization algorithm to obtain the trained weights. For a given test sample Engine health index (HI) is defined as Features The square of the distance to the center o of the hypersphere is:

[0163] (20).

[0164] The Deep SVDD network structure and training in S2 include the following methods: the slow features extracted by KSFA reflect the characteristics of the gas path parameter data changing over time; convolutional neural networks are often used for multi-array signals such as time series signals due to their strong generalization ability; at the same time, the feature attention mechanism can select more critical information for the current task objective from many features. Therefore, for the aero-engine anomaly monitoring task, this invention designs an encoder structure based on 1D-CNN fusion feature attention mechanism for use in Deep SVDD to learn the nonlinear representation of input data in low-dimensional latent space.

[0165] The encoder consists of two one-dimensional convolutional layers and one fully connected layer. The specific parameter settings for each layer are as follows:

[0166] The first convolutional layer has 1 input channel and 32 output channels, using a 3x3 convolutional kernel and 1 padding pixel. The first pooling layer has a convolutional kernel size of 2 and a stride of 2, halving the length of the input.

[0167] The second convolutional layer has 32 input channels and 64 output channels. It also uses a 3x3 convolutional kernel and 1 padding pixel. The second pooling layer also has 2 convolutional kernels and a stride of 2.

[0168] The feature extraction layer also includes a Flatten layer to flatten the tensor output by the convolutional layer into a one-dimensional tensor, and a fully connected layer. The feature attention mechanism is implemented using a sigmoid function, with the input being the output of the fully connected layer. Assuming the output of the feature attention mechanism is w, the final feature representation... for , where z is the encoder output and c is the center;

[0169] The training framework for Deep SVDD networks mainly consists of two stages: an autoencoder pre-training stage and a Deep SVDD training stage. The goal is to obtain the optimal Deep SVDD network weights. and the center of the supersphere .

[0170] The aero-engine air path anomaly determination based on box plot in S3 includes the following methods: Box plot, as a commonly used outlier screening method, can calculate the normal and abnormal ranges of the obtained aero-engine air path performance health factors.

[0171] The box length of a box plot is the interquartile range (IQR), which is expressed as:

[0172] (twenty one)

[0173] and 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:

[0174] (twenty two)

[0175] In the formula, , These are the lower quartile and upper quartile, respectively, and k is the outlier coefficient. In this invention, k=3 is used to monitor extreme outliers of health factors.

[0176] The proposed method for monitoring aero-engine anomalies consists of an offline phase and an online phase, with the specific steps as follows:

[0177] Offline phase

[0178] A. Perform kernel slow feature analysis on the normal airway training data to extract slow change features;

[0179] B. Construct a Deep SVDD model based on the slow characteristics of the normal state, and obtain the model parameters;

[0180] C. Calculate health indicators for normal-state data using Deep SVDD parameters to describe the process of aero-engine performance degradation. To improve the sensitivity of the HI indicator to describing aero-engine degradation, EWMA is added for smoothing based on HI.

[0181] (twenty three)

[0182] In the formula: The mean of the first t data points. The smoothing exponent is typically between 0.05 and 0.25. A value of 0.15 can reflect the degradation trend well;

[0183] D. Use the box plot method to calculate the threshold of the training data.

[0184] This invention uses box plots to derive thresholds from health indicators obtained from training samples under normal conditions of different models, and applies these thresholds as the normal state thresholds to subsequent online phase monitoring. The magnitude of the thresholds will vary depending on the magnitude of the health indicators for different engines on different models. However, each engine can establish its own threshold based on its normal state, creating a "personalized" early fault symptom alarm threshold for each engine.

[0185] The KSFA-Deep SVDD-based aero-engine anomaly monitoring model in S4 includes the following methods: (twenty four)

[0186] In the formula: This represents the upper quartile of the health factor in a normal state. The interquartile range is denoted as , and k is the outlier coefficient. In this invention, k=3 is used to monitor extreme outliers of health factors.

[0187] Online phase

[0188] A. Perform kernel slow feature analysis on the test data during flight to extract slow change features;

[0189] B. Calculate the distance between the test data and the center of the Deep SVDD hypersphere in a normal state, and use this as a health indicator;

[0190] C. Monitor the health indicators of the aircraft engine based on the threshold. If three consecutive points exceed the threshold, an early warning will be issued.

[0191] The parameter selection in S5 includes the following methods: QAR data contains a large number of different performance parameters, but not all parameters are suitable for health status assessment. Since some parameters are not closely related to the health performance of the engine, and the data information of some parameters is incomplete, it is necessary to select parameters that are fully recorded and closely related to the state of the engine when conducting engine health performance assessment. In order to determine the characteristic parameters related to EGT, the Pearson correlation coefficient is used for correlation analysis, as shown in equation (25).

[0192] (25)

[0193] In the formula, The closer the value is to 1, the stronger the correlation between the two sets of data. The correlation between two parameters can be 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 correlation, respectively.

[0194] The dataset preprocessing in S5 includes the following methods: Cruise phase data for each flight is selected as the analysis object. Due to the large amount of data, an average value is taken every 600 seconds. Furthermore, since the maximum and minimum values ​​cannot be obtained during real-time status monitoring, this invention uses z-score normalization to normalize the real-time status parameters. The formula is as follows:

[0195] (26).

[0196] Please refer to the following: Figure 5Based on the correlation analysis results, the characteristic indicators related to EGT were finally determined to be: fuel flow rate (FF), low-pressure rotor speed (N1), high-pressure rotor speed (N2), low-pressure compressor outlet total temperature (T25), and high-pressure compressor outlet total temperature (T3), with a sampling frequency of 1Hz for each parameter.

[0197] Meanwhile, aircraft flight phases typically include several states such as parking, taxiing out, takeoff, climb, cruise, approach, and taxiing in. Considering data stability, data from a cycle phase that approximates steady flight are selected as the research object.

[0198] Deep SVDD parameter settings:

[0199] The designed Deep SVDD model was trained using a preprocessed dataset. During training, the efficient Adam optimizer was employed with a batch size of 100 and an initial learning rate of 0.001. The ReLU activation function was used in both convolutional and fully connected layers, which, combined with the Adam optimizer, accelerated convergence. The loss function was defined as a combination of the average Euclidean distance between the embedding and its center and a penalty term based on the embedding radius; the goal was to minimize this loss. Optimal parameters were obtained through continuous testing with empirical data, as shown in Table 1.

[0200] Table 1. Deep SVDD Structure Parameter Settings

[0201] 1 Input layer The input layer （6,1） 2 Conv1D Kernel size=3padding=1stride=1 （1,32） 3 MaxPool1d Kernel size=2stride=2 （32,32） 4 Conv1D Kernel size=3padding=1stride=1 （32,64） 5 MaxPool1d Kernel size=2stride=2 （64,64） 6 Linear The linear layer （64,16） 7 Linear The linear layer of attention （16,4） 8 Output layer The output layer （4,16）

[0202] Comparison of model selection and parameter settings:

[0203] To verify the effectiveness of the proposed aero-engine air path anomaly monitoring method, a monitoring model based on traditional SFA and traditional SVDD was constructed. Isolation forests, due to their excellent accuracy, are widely used for anomaly detection in structured data. This invention uses their anomaly score as a health indicator and selects appropriate model parameters based on the data characteristics. Furthermore, to highlight the effectiveness of the Deep SVDD feature extraction network constructed in this invention, a deep extraction model based on DBN was constructed, with the following model parameter settings:

[0204] (1) KSFA-Isolation Forest

[0205] The forest has 100 trees, and the random number generator uses a seed of 42.

[0206] (2) KSFA-SVDD

[0207] The SVDD kernel function is a Gaussian kernel function, and the penalty factor is set to 0.9;

[0208] (3) KSFA-DBN-Deep SVDD

[0209] Taking into account both network performance and complexity, the network structure is a three-layer fully connected neural network. The first layer is a fully connected layer with an input dimension of 6 and an output dimension of 64. The second layer is a fully connected layer with an input dimension of 64 and an output dimension of 32. The third layer is a fully connected layer with an input dimension of 32 and an output dimension of 16. Each layer is followed by a batch normalization layer and a ReLU activation function. The last layer uses the hyperbolic tangent function (tanh) as the activation function.

[0210] Please refer to the following: Figure 6 It was found that the overall trend of the monitoring parameters of the left engine was not obvious, and it was impossible to determine whether the engine was abnormal based solely on the change curve of the left engine parameters. Therefore, the KSFA algorithm was used to extract slow features from the parameters of the left engine and to observe the extracted kernel slow features.

[0211] Please refer to the following: Figure 7 It was determined that the three slowest-changing kernel features were extracted. These features were stable in the initial part. The left engine performance features extracted using KSFA reduced interference from changes in external environment or flight conditions. Simultaneously, the slow-changing features exhibited abrupt changes of varying degrees at the tail end, indicating an anomaly in the left engine at the point of abrupt change. This demonstrates that compared to the performance parameters of a single engine, KSFA-extracted features improve the sensitivity of identifying engine anomalies. The left engine performance features extracted using KSFA showed significant fluctuations in the latter half, leading to false alarms and other abnormal changes during fault monitoring.

[0212] Please refer to the following: Figure 8 , Figure 9 , Figure 10 and Figure 11 Based on the above analysis, the top 30% of normal flight data from Case 1 were selected as training data input into the model. During model training, this invention used four comparative methods—"SFA-Deep SVDD", "KSFA-Isolation Forest", "KSFA-SVDD", and "KSFA-DBN-Deep SVDD"—to monitor the gas path system. The results are as follows: Figure 9 As shown. Observing the test results of each model, we can see that:

[0213] Compared to the anomaly monitoring model proposed in this invention, after completion, all test data are input into the model as test groups to calculate the corresponding HI. Figure 8It can be observed that the constructed left-side engine HI remained stable for the first 919 cycles, but experienced a sudden change in cycle 920, corresponding to a real-world date of March 29, 2020. Following a short period of abnormal fluctuation, the airline received an aircraft alarm in cycle 958, leading to engine maintenance, corresponding to a real-world date of June 2, 2020. Therefore, this invention achieves early warning of engine failure symptoms, shifting the fault threshold forward and providing a reference for maintenance planning.

[0214] The HI (Hypermetric Indicators) constructed by SFA-Deep SVDD and KSFA-SVDD exhibit significant fluctuations in the latter half, leading to numerous false alarms during normal condition monitoring. Furthermore, the HI constructed based on KSFA-Isolation Forest and KSFA-DBN-Deep SVDD show overall drastic fluctuations, easily resulting in intermittent false alarms, making them unreliable for airlines in developing maintenance plans.

[0215] Similar to the parameter settings in Case 1, the first 30% of the regular flight data from Case 2 were selected as the training data input into the model. After the model training was completed, all test data were input into the model as the test group, and the corresponding HI was calculated. Figure 10 Observations revealed a significant degradation trend in the constructed left-engine HI (High-Intensity Dynamics) pattern. Early degradation occurred in the aero-engine during cycle 263, corresponding to June 15, 2020. The degradation worsened rapidly after cycle 263, triggering continuous warnings. The airline received an aircraft alarm on August 20, 2020, during cycle 319, and initiated engine maintenance. This demonstrates that the method of this invention can identify engine anomalies in different cases in advance, allowing sufficient time to develop a reliable maintenance plan.

[0216] In Case 2, this invention also used four comparative methods to monitor the gas path system: “SFA-Deep SVDD”, “KSFA-IsolationForest”, “KSFA-SVDD”, and “KSFA-DBN-Deep SVDD”. The results are as follows: Figure 11 As shown. Observing the test results of each model, we can see that:

[0217] Compared to the anomaly monitoring model proposed in this invention, while the anomaly monitoring methods based on KSFA-Isolation Forest, KSFA-SVDD, and SFA-Deep SVDD show a degradation trend, they all exhibit varying degrees of false alarms, making them unreliable for airlines in developing maintenance plans. The anomaly monitoring method based on KSFA-DBN-Deep SVDD is 35 cycles behind the method proposed in this invention. These two anomaly cases demonstrate that the method proposed in this invention has broader applicability and is better suited for monitoring various faults in aero-engines, possessing significant engineering application value.

[0218] Compared with related technologies, the air path anomaly monitoring method for aero-engines based on flight data provided by this invention has the following beneficial effects:

[0219] This invention provides a method for monitoring aero-engine air path anomalies based on flight data. The air path features extracted by slow feature analysis can reflect the changes in the air path to a certain extent, reducing the difficulty of subsequent anomaly monitoring. However, in the actual operation of aero-engines, the air path data has extremely high dimensionality. Slow feature analysis often uses polynomial expansion methods to process it nonlinearly. When the dimensionality is high, it will lead to the curse of dimensionality, which will greatly affect the feature extraction speed and the accuracy of subsequent modeling. Considering the characteristics of aero-engine air path parameter data, kernel slow feature analysis with kernel function expansion can enhance its feature extraction capability and achieve effective extraction of feature information.

[0220] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for monitoring aero-engine air path anomalies based on flight data, characterized in that, Includes the following steps: S1. Gas path feature extraction based on kernel slow feature analysis: slow feature analysis principle and kernel slow feature analysis principle; S2. Performance evaluation of gas path based on deep support vector description algorithm: principle of traditional support vector description algorithm, principle of deep support vector description algorithm and Deep SVDD network structure and training; S3. Determination of aero-engine gas path anomalies based on box plots; S4. An aero-engine anomaly monitoring model based on KSFA-Deep SVDD; S5. Dataset Selection and Parameter Setting: Introduction to experimental datasets, parameter selection, dataset preprocessing, Deep SVDD parameter setting, and selection and parameter setting of comparison models; S6. Experimental Results and Discussion: Slow Feature Extraction; The principle of the deep support vector description algorithm in S2 includes the following: Traditional support vector description mainly uses a single-layer nonlinear mapping for a given training sample and uses kernel functions to complete feature extraction. However, single-layer feature extraction is often difficult to uncover the complex relationships between nonlinear data. In response to the shortcomings of traditional support vector description, deep support vector description has been proposed in recent years. In deep SVDD, the goal is the same, but instead of using a linear mapping in a high-dimensional space, a deep neural network is used to learn the nonlinear representation of the input data in a low-dimensional latent space. For slow features extracted by KSFA The extraction process is shown below: （17） In the formula, Let L be the nonlinear mapping function of the Lth layer in a deep neural network. These are the weights of the Lth layer network; Unlike the objective function of traditional SVDD, the optimization objective of deep SVDD models is to jointly learn the network parameters W while minimizing the amount of data enclosing the hypersphere in the output space, so that the final features in the output space are maximized. The objective function is as follows: The object should be enclosed within an optimal hypersphere as much as possible. （18） In the formula, R is the radius of the hypersphere, o is the a priori center of the hypersphere, and v is the equilibrium parameter. Let W be the penalty factor, and W be the set of network parameters, denoted as... In anomaly detection, the training samples are usually normal samples, and the above objective function can be simplified: （19） The network weight parameters W are trained using a deep network optimization algorithm to obtain the trained weights. For a given test sample Engine health index (HI) is defined as Features The square of the distance to the center o of the hypersphere is: （20）； The Deep SVDD network structure and training in S2 include the following methods: the slow features extracted by KSFA reflect the characteristics of the gas path parameter data changing over time; convolutional neural networks are often used for time series signals and multi-array signals due to their strong generalization ability; at the same time, the feature attention mechanism can select the information that is more critical to the current task objective from many features. Therefore, for the aero-engine anomaly monitoring task, an encoder structure based on 1D-CNN fusion feature attention mechanism is designed for Deep SVDD to learn the nonlinear representation of the input data in the low-dimensional latent space. The encoder consists of two one-dimensional convolutional layers and one fully connected layer. The specific parameter settings for each layer are as follows: The first convolutional layer has 1 input channel and 32 output channels, using a 3x3 convolutional kernel and 1 padding pixel. The first pooling layer has a convolutional kernel size of 2 and a stride of 2, halving the length of the input. The second convolutional layer has 32 input channels and 64 output channels. It also uses a 3x3 convolutional kernel and 1 padding pixel. The second pooling layer also has 2 convolutional kernels and a stride of 2. The feature extraction layer also includes a Flatten layer to flatten the tensor output by the convolutional layer into a one-dimensional tensor, and a fully connected layer. The feature attention mechanism is implemented using a sigmoid function, with the input being the output of the fully connected layer. Assuming the output of the feature attention mechanism is w, the final feature representation... for , where z is the encoder output and c is the center; The training framework for Deep SVDD networks mainly consists of two stages: an autoencoder pre-training stage and a Deep SVDD training stage. The goal is to obtain the optimal Deep SVDD network weights. and the center of the supersphere ; The aero-engine air path anomaly determination based on box plot in S3 includes the following methods: Box plot, as a commonly used outlier screening method, can calculate the normal and abnormal ranges of the obtained aero-engine air path performance health factors. The box length of a box plot is the interquartile range (IQR), which is expressed as: （21） and 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: （22） In the formula, , These are the lower quartile and upper quartile, respectively, and k is the outlier coefficient, which is set to k=3 to monitor extreme outliers of health factors. The proposed method for monitoring aero-engine anomalies consists of an offline phase and an online phase, with the specific steps as follows: Offline phase A. Perform kernel slow feature analysis on the normal airway training data to extract slow change features; B. Construct a Deep SVDD model based on the slow characteristics of the normal state, and obtain the model parameters; C. Calculate health indicators for normal-state data using Deep SVDD parameters to describe the process of aero-engine performance degradation. To improve the sensitivity of the HI indicator to describing aero-engine degradation, EWMA is added for smoothing based on HI. （23） In the formula: The mean of the first t data points. To select the smoothing exponent, choose... A value of 0.15 can reflect the degradation trend well; D. Use the box plot method to calculate the threshold of the training data; The threshold is obtained by using the box plot method to obtain the health indicators from the training samples of different models under normal conditions. This threshold is then used as the threshold for normal conditions and applied to subsequent online stage monitoring. The threshold magnitude will vary depending on the magnitude of the health indicators of different engines on different models. However, each engine can establish its own threshold based on its normal condition, thus establishing a "personalized" early fault symptom alarm threshold for each engine. The KSFA-Deep SVDD-based aero-engine anomaly monitoring model in S4 includes the following methods: (twenty four) In the formula: This represents the upper quartile of the health factor in a normal state. The interquartile range is denoted as , and k is the outlier coefficient. We take k=3 to monitor extreme outliers of health factors. Online phase A. Perform kernel slow feature analysis on the test data during flight to extract slow change features; B. Calculate the distance between the test data and the center of the Deep SVDD hypersphere in a normal state, and use this as a health indicator; C. Monitor the health indicators of the aircraft engine based on the threshold. If three consecutive points exceed the threshold, an early warning will be issued.

2. The method for monitoring aero-engine air path anomalies based on flight data according to claim 1, characterized in that, The slow characteristic analysis principle in S1 includes the following: During the operation of an aero-engine, the performance of the air path changes slowly over time. Slow characteristic analysis is proposed to find a set of input-output functions that make the output signal change as slowly as possible. The features extracted by the slow characteristic analysis algorithm can effectively reflect the slow change of the air path state, and the magnitude of the change can reflect the change state of the air path. Given an n-dimensional gas path input signal Considering a linear mapping function, we can find the mapping relationship between the objective function and the input function using SFA, i.e. （1） In the formula, Represents the weight of the j-th function; Accordingly, the objective function of the SFA mapping function is: （2） In the formula, Represents the difference value. Represents the time difference of s. It is the average value of s along the time axis; Furthermore, the optimization problem of the objective function should satisfy the following conditions: （3） （4） （5） Finally, it can be reconstructed into the following generalized eigenvalue decomposition problem: （6） In the formula, , . , It is a diagonal matrix composed of generalized eigenvalues. After solving equation (6) and obtaining the transformation matrix W, the characteristic S reflecting the slow change of the gas path can be expressed as: 。 3. The method for monitoring aero-engine air path anomalies based on flight data according to claim 1, characterized in that, The principle of slow feature analysis in S1 includes the following: In the actual operation of aero-engines, the air path data has extremely high dimensionality. Slow feature analysis often uses the polynomial expansion method to process it. However, the air path sensor data of aero-engines has a large dimensionality, and polynomial expansion will cause the curse of dimensionality, which will greatly affect the feature extraction speed and the accuracy of subsequent modeling. Kernel tricks are a method for processing nonlinear data. Kernel slow feature analysis introduces a kernel function to replace the polynomial extension in slow feature analysis, thereby enabling the processing of nonlinear data. This involves mapping the original data in the low-dimensional space to a high-dimensional space, transforming the inner product operation in the high-dimensional space into a kernel function operation in the input space. Various kernel functions are commonly used; this method employs the Gaussian kernel function, which has strong generalization and smoothing estimation capabilities. After multiple experiments, a Gaussian kernel size of 5 was selected. （8） In the formula, and For any sample; Replacing the polynomial expansion with the Gaussian kernel function, i.e., replacing the second-order polynomial expansion with the expanded data K, the objective function of the mapping function in equation (2) becomes: （9） The generalized eigenvalue decomposition problem in equation (6) is recast as: （10） In the formula, for The covariance matrix, Let K be the covariance matrix; Solving the generalized eigenvalue decomposition problem in equation (10) yields the transformation matrix W, and the kernel slow eigenvalue can be expressed as: （11）。 4. The method for monitoring aero-engine air path anomalies based on flight data according to claim 1, characterized in that, The traditional support vector description algorithm in S2 follows these principles: In the actual operation of an aero-engine, the ratio of normal to abnormal samples is unbalanced, resulting in a sample imbalance problem. High-quality abnormal labeled samples are extremely difficult to obtain. Unsupervised learning algorithms are suitable for situations where both normal and abnormal data exist but no labels are available. When only healthy state data is available, support vector data description can be trained based on normal aero-engine samples and effectively distinguish between normal and abnormal engine states during subsequent operation. The performance evaluation process of the traditional support vector description algorithm is as follows: Given an n-dimensional slow feature extracted by KSFA SVDD performs a nonlinear mapping on the samples to find an optimal hypersphere in the high-dimensional feature space that encloses all training samples. The corresponding objective function is: （12） In the formula, Let O be the radius of the hypersphere, and O be the center. As slack variables, As a penalty factor; By introducing the solution formula (12) for the Lagrange function, the original problem can be rewritten as the following dual description: （13） In the formula, For kernel function operations, this formula selects the Gaussian kernel function, which has strong generalization and smoothing estimation capabilities. Solving equation (13) yields the result that the center of the hypersphere is... and hypersphere radius : （14） （15） Define the test sample for any time. To the center of the supersphere The square of the distance is the health indicator (HI): （16）。 5. The method for monitoring aero-engine air path anomalies based on flight data according to claim 1, characterized in that, The parameter selection in S5 includes the following methods: QAR data contains a large number of different performance parameters, but not all parameters are suitable for health status assessment. Since some parameters are not closely related to the health performance of the engine, and the data information of some parameters is incomplete, it is necessary to select parameters that are fully recorded and closely related to the state of the engine when conducting engine health performance assessment. In order to determine the characteristic parameters related to EGT, the Pearson correlation coefficient is used for correlation analysis, as shown in equation (25). （25） In the formula, The closer the value is to 1, the stronger the correlation between the two sets of data. The correlation between two parameters can be 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 correlation, respectively.

6. The method for monitoring aero-engine air path anomalies based on flight data according to claim 1, characterized in that, The dataset preprocessing in S5 includes the following methods: Cruise phase data for each flight is selected as the analysis object. Due to the large amount of data, an average value is taken every 600 seconds. Furthermore, since the maximum and minimum values ​​cannot be obtained during real-time status monitoring, the z-score normalization method is used to normalize the real-time status parameters. The formula is as follows: （26）。

Images