Aero-engine fault diagnosis method based on feature augmentation

Abstract

The application discloses a kind of based on feature augmentation aero-engine fault diagnosis method, belong to aero-engine fault diagnosis technical field.The present application is directed to the saliency of aero-engine fault signal feature gradually reduces, it is difficult to extract useful information in engine actual operation and maintenance data, and then influence fault diagnosis accuracy problem.Original sample is carried out high-dimensional feature augmentation, and feature augmented sample is obtained;Again normalization processing is carried out, and then training sample set is constructed from normalized sample;Normal state sample and fault sample in training sample set are set different labels respectively;Training sample set is used to train fault diagnosis network, when reaching preset iteration number, training after fault diagnosis network is obtained;The operation data of aero-engine is collected, and normalized after processing is obtained after diagnosis data;Normalized after diagnosis data is input into training after fault diagnosis network, and aero-engine fault diagnosis result is obtained.The present application is used for aero-engine fault diagnosis.

Description

Technical Field

[0001] This invention relates to a method for diagnosing aero-engine faults based on feature amplification, and belongs to the field of aero-engine fault diagnosis technology. Background Technology

[0002] Aero engines are the "heart" of an aircraft. As the primary power source and bleed air system, their health is crucial to the safety and economy of flight. With decreasing accident rates, aero engines have achieved higher reliability and lower failure rates. However, due to their complex structure and harsh operating environment, aero engines inevitably experience malfunctions. Furthermore, the diverse models, complex internal mechanisms, and high diagnostic difficulty of aero engines make fault diagnosis a consistently challenging task.

[0003] To accurately diagnose aero-engine faults, airlines and research institutes have conducted in-depth research on methods for diagnosing engine gas path faults. Currently, these methods can be mainly categorized into two types: model-based methods and data-driven methods. Model-based methods can relatively accurately detect sudden faults, requiring no experience or historical data, and the results are interpretable. However, due to the complexity of engine systems, building accurate models is becoming increasingly difficult. Furthermore, building aero-engine diagnostic models requires extensive prior knowledge and accurate understanding of the engine's structural and system design parameters, which are often confidential. Therefore, it is extremely difficult for airlines to build accurate engine fault diagnosis models. With the gradual application of digitalization, networking, and intelligent technologies in the industrial field, more comprehensive and massive amounts of industrial data are becoming easier to collect and store, laying the foundation for data-driven fault diagnosis methods. Currently, more and more data-driven methods are being used in fault diagnosis. Compared to model-based methods, data-driven methods have the advantages of not requiring a mathematical and physical model of the engine, being unaffected by the engine's complexity, requiring no extensive prior knowledge or experience, and possessing excellent nonlinear processing capabilities. Therefore, data-driven fault diagnosis methods based entirely on monitoring data are more practical in the field of aero-engine fault diagnosis. With the rapid development of machine learning methods, especially deep learning methods, data-driven methods have gradually become the mainstream for aero-engine fault diagnosis.

[0004] Data-driven methods require high-quality data samples; ideally, there should be sufficient samples, adequate labels, and good separability. However, actual engine maintenance data often exhibits temporal and nonlinear characteristics that make it difficult to separate. Currently, some deep learning algorithms perform well in this area, including convolutional neural networks, long short-term memory networks, deep belief networks, YOLOv3, and pattern recognition algorithms. These algorithms can effectively extract abstract features from time-series signals, improving the accuracy of fault identification. However, as the reliability of aero-engines continues to increase, the saliency of fault signals gradually decreases. Therefore, it is crucial to fully utilize useful information from actual engine maintenance data to improve the accuracy of fault diagnosis. Summary of the Invention

[0005] To address the problem that the saliency of fault signal features in aero-engines is gradually decreasing, making it difficult to fully extract useful information from actual engine operation and maintenance data, thus affecting the accuracy of fault diagnosis, this invention provides an aero-engine fault diagnosis method based on feature amplification.

[0006] The present invention provides a method for fault diagnosis of aero-engines based on feature amplification, comprising:

[0007] Step 1: Perform high-dimensional feature amplification on the original sample to obtain the feature-amplified sample;

[0008] Step 2: Normalize the amplified samples and then construct a training sample set from the normalized samples; assign different labels to the normal state samples and fault samples in the training sample set.

[0009] Step 3: Train the fault diagnosis network using the training sample set. After reaching the preset number of iterations, the trained fault diagnosis network is obtained.

[0010] Step 4: Collect the operating data of the aero-engine, and perform the high-dimensional feature amplification described in Step 1 and the normalization processing described in Step 2 to obtain the normalized data to be diagnosed; input the normalized data to be diagnosed into the trained fault diagnosis network to obtain the aero-engine fault diagnosis results.

[0011] According to the feature amplification-based aero-engine fault diagnosis method of the present invention, the original sample in step one is extracted from the engine air path performance monitoring data.

[0012] According to the feature-amplified aero-engine fault diagnosis method of the present invention, the method for high-dimensional feature amplification of the original sample in step one includes a high-dimensional mapping method based on a polynomial kernel explicit mapping function, specifically:

[0013] Using polynomial kernel function K p For the original sample x iand the original sample x j Perform the inner product operation, x i ≠x j :

[0014]

[0015] In the formula φ p This represents the mapping function corresponding to the polynomial kernel function, where r and d are the undetermined parameters in the polynomial kernel function, respectively.

[0016] Where the i-th original sample x i =[x i,1 ,x i,2 ,…,x i,n ], the j-th original sample x j =[x j,1 ,x j,2 ,…,x j,n ], where n represents the dimension of the sample, and the total number of original samples is N, i = 1, 2, 3, ..., N, j = 1, 2, 3, ..., N;

[0017] φ p (x i The dimension of ) is To reduce φ p (x i The dimension of x is calculated and the original sample x is preserved. i Given the information, choose r=1, d=2, then φ p (x i The expression for ) is:

[0018]

[0019] φ p (x i In the expression, the constant term 1 and the coefficient Discarded, denoted as φ p ′(x i ):

[0020]

[0021] φ p ′(x i () as a sample after feature amplification.

[0022] According to the feature-amplified aero-engine fault diagnosis method of the present invention, the method for high-dimensional feature amplification of the original sample in step one includes an experience-based high-dimensional space modeling method, specifically:

[0023] Based on the failure time of the faulty engine k, extract the m preceding times before the failure of engine k from the OEM data. kThe original sample set A is obtained by collecting the exhaust temperature deviation (DEGT), high-pressure rotor speed deviation (DN2), fuel flow deviation (DFF), and exhaust temperature margin change (EGTM) for each flight cycle. k :

[0024]

[0025] In the formula, the i-th original sample x I for:

[0026] x I =[DEGT I DN2 I ,DFF I EGTM I ], I = 1, 2, 3, ..., m k ;m k Integers greater than 10;

[0027] The original sample x is processed using a Long Short-Term Memory (LSTM) network. I Smoothing and denoising processes are performed to obtain smoothed sample data; K-step differencing is then performed on the smoothed sample data, with K set to 9, to obtain the feature-amplified sample set A. k ′:

[0028]

[0029] In the formula, DEGT′ represents the DEGT value after LSTM smoothing and denoising, DN2′ represents the DN2 value after LSTM smoothing and denoising, DFF′ represents the DFF value after LSTM smoothing and denoising, and EGTM′ represents the EGTM value after LSTM smoothing and denoising.

[0030] According to the feature amplification-based aero-engine fault diagnosis method of the present invention, the normalized samples are also used to construct a test sample set, and the ratio of the training sample set to the test sample set is 4:1.

[0031] According to the feature amplification-based aero-engine fault diagnosis method of the present invention, the sample data input to the fault diagnosis network each time is selected from the training sample set through a sliding window.

[0032] According to the feature amplification-based aero-engine fault diagnosis method of the present invention, the fault diagnosis network is a convolutional neural network (CNN), a long short-term memory network (LSTM), a temporal convolutional network (TCN), or a deep residual shrinking network (DRSN-CW).

[0033] According to the feature amplification-based aero-engine fault diagnosis method of the present invention, the learning rate of the fault diagnosis network is set to 0.001, the Adam optimizer is used to update the weights, and L2 regularization is used to reduce overfitting; the decay coefficient of L2 regularization is 0.0001, and the batch size is 10.

[0034] The beneficial effects of this invention are as follows: The method of this invention improves the classification effect of fault diagnosis by amplifying features from the original data. One approach is based on a polynomial kernel explicit mapping function, which requires no complex operations and provides a clear mapping relationship between each dimension and the original dimensions after high-dimensional mapping. Another approach is based on an empirical high-dimensional space modeling method, which first uses LSTM to smooth and denoise the original data, and then uses the differencing data as the amplified new features. This invention's fault diagnosis method fully utilizes useful information from actual engine operation and maintenance data, thus improving the accuracy of fault diagnosis.

[0035] To demonstrate the effectiveness of the method of this invention, experiments were conducted on engine datasets and public datasets, respectively. Two additional methods using machine learning for high-dimensional mapping were also included for comparison. Experimental results show that Scheme 1 and Scheme 2 of this invention achieve better classification results than the original data in multiple evaluation metrics and outperform the two machine learning methods. On the engine dataset, Scheme 1 and Scheme 2 achieve greater than 90% on all three metrics of the TCN deep learning model. On the public dataset, Scheme 1 and Scheme 2 achieve greater than 97% on all three metrics of the TCN deep learning model. Furthermore, Scheme 1 achieves an accuracy of 99.5% on the TCN deep learning model, and Scheme 2 achieves an accuracy of 98.3%. Moreover, based on the application of the minimum norm method to both datasets, it was found that if the original data has strong orthogonality, the high-dimensional mapping will still retain a certain degree of orthogonality. This indicates that retaining a certain degree of orthogonality after high-dimensional mapping can effectively improve the classification results of fault diagnosis. Attached Figure Description

[0036] Figure 1 This is a schematic diagram of the process of collecting and converting aero-engine air path parameters; N1 in the diagram is equivalent to N1 in the manual, and N2 is equivalent to N2 in the manual.

[0037] Figure 2 This is a typical CNR report example of a gas circuit fault;

[0038] Figure 3 This is a schematic diagram illustrating the change trend of DEGT when an aircraft engine experiences an EGT indication failure; in the diagram, Flight cycle represents the flight cycle.

[0039] Figure 4 This is a schematic diagram illustrating the changing trend of EGTM when an aircraft engine experiences an EGT indication fault.

[0040] Figure 5 This is a schematic diagram showing the changing trend of DN2 when an aircraft engine experiences an EGT indication fault;

[0041] Figure 6 This is a schematic diagram illustrating the changing trend of DFF when an aircraft engine experiences an EGT indication fault;

[0042] Figure 7 This is a general framework diagram of the aero-engine fault diagnosis method based on feature amplification described in this invention;

[0043] Figure 8 This is a schematic diagram of an LSTM cell structure;

[0044] Figure 9 This is a schematic diagram of the fitting of air circuit parameters in an engine malfunction.

[0045] Figure 10 This is a schematic diagram of a sliding window for extracting samples;

[0046] Figure 11 This is a flowchart of the engine air circuit fault diagnosis method of the present invention;

[0047] Figure 12 The bar chart corresponds to the Accuracy (%) in Table 3;

[0048] Figure 13 The bar chart corresponds to Precision (%) in Table 3;

[0049] Figure 14 The bar chart corresponds to the F1-score (%) in Table 3;

[0050] Figure 15 This is a bar chart corresponding to Computing time(s) in Table 3; in the chart, Computing time represents the computation time.

[0051] Figure 16 The bar chart corresponds to the Accuracy (%) in Table 4;

[0052] Figure 17 This is a bar chart corresponding to Precision (%) in Table 4;

[0053] Figure 18 The bar chart corresponds to the F1-score (%) in Table 4;

[0054] Figure 19This is a bar chart corresponding to Computing time(s) in Table 4;

[0055] Figure 20 The bar chart corresponds to the Accuracy (%) in Table 5;

[0056] Figure 21 This is a bar chart corresponding to Precision (%) in Table 5;

[0057] Figure 22 The bar chart corresponds to the F1-score (%) in Table 5;

[0058] Figure 23 This is a bar chart corresponding to Computing time(s) in Table 5. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0061] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the scope of the invention.

[0062] Specific Implementation Method 1: Combination Figures 1 to 7 As shown, this invention provides a method for diagnosing aero-engine faults based on feature amplification, including:

[0063] Step 1: Perform high-dimensional feature amplification on the original sample to obtain the feature-amplified sample;

[0064] Step 2: Normalize the amplified samples and then construct a training sample set from the normalized samples; assign different labels to the normal state samples and fault samples in the training sample set.

[0065] Step 3: Train the fault diagnosis network using the training sample set. After reaching the preset number of iterations, the trained fault diagnosis network is obtained.

[0066] Step 4: Collect the operating data of the aero-engine, and perform the high-dimensional feature amplification described in Step 1 and the normalization processing described in Step 2 to obtain the normalized data to be diagnosed; input the normalized data to be diagnosed into the trained fault diagnosis network to obtain the aero-engine fault diagnosis results.

[0067] In step one, the original sample was extracted from the engine air circuit performance monitoring data.

[0068] The civil aviation engine studied in this embodiment mainly consists of six major components: fan (FAN), low-pressure compressor (LPC), high-pressure compressor (HPC), combustor (CC), low-pressure turbine (LPT), and high-pressure turbine (HPT). Figure 1 As shown, Figure 1 "Other" indicates other data. Ground monitoring systems primarily monitor the gas path performance of aero engines by monitoring parameters such as exhaust gas temperature (EGT), fuel flow rate (FF), low-pressure rotor speed (N1), and high-pressure rotor speed (N2). In practice, the original equipment manufacturer (OEM) uses thrust setpoints, environmental factors, and operating condition information to convert the originally monitored gas path parameters into baseline deviation values ​​to monitor the engine's performance status; this is known as OEM data.

[0069] Based on Customer Notification Reports (CNRs) provided by OEMs to airlines and relevant literature, OEMs primarily monitor engine airflow performance using four airflow parameters: Exhaust Gas Temperature Deviation (DEGT), Exhaust Gas Temperature Margin (EGTM), Delta Core Speed ​​(DN2), and Delta Fuel Flow (DFF). The CNR records the time of any engine anomalies and their potential causes, and is released to airlines by OEMs based on their case studies and knowledge bases. However, the CNR has a time lag, meaning it cannot be provided to airlines immediately, and service fees are incurred during this process. Therefore, it is crucial for airlines to be able to promptly assess engine health based on these airflow parameters.

[0070] Figure 2 This is a typical CNR report example of a gas circuit fault. Figure 2 This is for illustrative purposes only, therefore a detailed explanation of the specific English text in the diagram is unnecessary. From Figure 2As can be seen in the diagram, Box 1 indicates the engine model, serial number, location, and report information, which facilitates subsequent classification and organization. Box 2 indicates the changes in response parameters when a fault occurs. Box 3 indicates the fault type determined by the OEM based on the diagnostic knowledge base. Box 4 indicates the start and end times of abnormal changes in status parameters, allowing customers to find the specific parameter values ​​for the corresponding dates in the database for a more intuitive judgment. Box 4 is also key data for fault sample collection in this implementation method.

[0071] Based on engine maintenance reports and CNR reports, the currently collected aero-engine gas path faults include three categories: EGT indication fault, TAT indication fault, and high-pressure turbine blade ablation fault. Analysis of the fault data corresponding to these three types of faults reveals that when an engine experiences a gas path fault, four gas path parameters—DEGT, EGTM, DFF, and DN2—will undergo certain changes. To further illustrate this, the following analysis uses one type of fault as an example to examine the changes in these four gas path parameters when a fault occurs. Figures 3 to 6 This diagram illustrates the changes in four gas path parameters when an aero-engine experiences an EGT indication fault. T1 represents the actual moment the engine malfunctions; T2 represents the moment the engine is diagnosed with a fault; point A shows the values ​​of each gas path parameter when the engine begins to malfunction; point B shows the values ​​of each gas path parameter when the engine is diagnosed with a fault; and Δt represents the number of flight cycle intervals between T1 and T2. As shown in the diagram, from T1 to T2, when the engine malfunctions, the four gas path parameters undergo certain changes, such as: DEGT showing a decreasing trend, EGTM showing an increasing trend, and DN2 and DFF showing a trend of first increasing and then decreasing.

[0072] Through the Figures 3 to 6 Analysis shows that when an aero-engine experiences a gas path failure, the failure type is correlated with changes in four gas path parameters: DEGT, EGTM, DFF, and DN2. Therefore, changes in these four parameters can be used to classify multiple gas path failures.

[0073] Figure 7This is a schematic diagram of the method of the present invention, which mainly includes three parts: (1) using two feature amplification techniques to perform high-dimensional mapping on the original fault samples; (2) analyzing the orthogonality of the feature-amplified fault samples at the theoretical level; and (3) using the feature-amplified fault sample set to train a deep learning model to achieve fault diagnosis. A detailed description is as follows: First, considering the highly nonlinear and difficult-to-separate problem of the original monitoring data, the two feature amplification techniques proposed in this embodiment (polynomial kernel explicit mapping method and empirical method) are used to map the original engine samples into a high-dimensional space to improve the separability of the original samples. Then, the orthogonality of the feature-amplified sample data is determined by the minimum relative norm method to see if the orthogonality of the data after high-dimensional mapping weakens. Theoretically, the two amplification methods proposed in this embodiment can ensure that the orthogonality of each dimension of the feature-amplified data is strong, and each dimension has strong independent information features. Finally, the feature-amplified sample data is normalized and used as input to a deep learning model for fault diagnosis, determining whether the input data is normal data or a certain type of fault data. Finally, in the experimental section, the feature-amplified sample data and the original data were input into the same deep learning model, and multiple evaluation metrics were used to demonstrate the effectiveness of the feature augmentation method proposed in this invention.

[0074] Furthermore, the methods for high-dimensional feature amplification of the original sample in step one include two approaches:

[0075] The first method is a high-dimensional mapping method based on a polynomial kernel explicit mapping function, where the polynomial kernel function refers to a kernel function expressed in polynomial form. Its core idea is to perform a high-dimensional mapping on the sample data, thereby making originally linearly inseparable data linearly separable. Specifically:

[0076] Using polynomial kernel function K p For the original sample x i and the original sample x j Perform the inner product operation, x i ≠x j :

[0077]

[0078] In the formula φ p This represents the mapping function corresponding to the polynomial kernel function, where r and d are the undetermined parameters in the polynomial kernel function, respectively.

[0079] Where the i-th original sample x i =[x i,1 ,x i,2 ,…,x i,n ], the j-th original sample x j =[x j,1 ,x j,2,…,x j,n ], where n represents the dimension of the sample, and the total number of original samples is N, i = 1, 2, 3, ..., N, j = 1, 2, 3, ..., N;

[0080] φ p (x i The dimension of ) is To reduce φ p (x i The dimension of x is calculated and the original sample x is preserved. i Given the information, choose r=1, d=2, then φ p (x i The expression for ) is:

[0081]

[0082] Based on φ p (x i ) observation, φ p (x i One term in the equation is a constant of 1. Since this constant term does not contain useful information, it can be discarded. When normalizing the sample, the coefficients... The presence or absence of a coefficient does not affect the processed data, therefore the coefficient is discarded. Get φ′ p (x i ), φ′ p (x i Compared to φ p (x i The latter is even simpler:

[0083]

[0084] φ′ p (x i () as a sample after feature amplification.

[0085] For φ p (x i φ′ obtained through modification p (x i ) Basically, φ is preserved p (x i The mapping form of ) is still φ′ p (x i φ can be considered as a form of explicit polynomial kernel mapping function. In this embodiment, the explicit polynomial kernel mapping function mentioned later specifically refers to φ. p ′(x iAnalysis of the Gaussian kernel explicit mapping function and the polynomial kernel explicit mapping function reveals that the polynomial kernel explicit mapping function consists of the original data, the squares of the original data, and the dot product of a subset of the original data. The polynomial kernel explicit mapping space expands upon the original space, differing in composition from the Gaussian kernel explicit mapping space.

[0086] The minimum relative norm method is used below to prove whether the orthogonality of features in the high-dimensional mapping space weakens after feature amplification:

[0087] Let the matrix form of the original sample be A: A = (DEGT, DN2, DFF, EGTM).

[0088] DEGT, DN2, DFF, and EGTM are all column vectors. A is orthogonalized column-wise using the Gram-Schmit orthogonalization method. The orthogonalization formula is expressed as:

[0089] DEGT″=DEGT

[0090]

[0091]

[0092]

[0093] Replacing the corresponding variables with x and y, in the formula<x,y> Let x represent the inner product of x and y; DEGT″, DN2″, DFF″, and EGTM″ are orthogonalized column vectors that intersect each other pairwise. This leads to the matrix B = [DEGT″, DN2″, DFF″, EGTM″].

[0094] Find the 1-norm of the column vectors of matrices A and B respectively, forming vectors C = [||DEGT||1,||DN2||1,||DFF||1,||EGTM||1] and D = [||DEGT″||1,||DN2″||1,||DFF″||1,||EGTM″||1]. ||.||1 represents the 1-norm of the vector. Calculate K. f Value: K f The evaluation index for the strength of orthogonality between different dimensions of the data:

[0095]

[0096] Define a lower threshold K1 and an upper threshold K2, where K1 < K2. If K f If K1 ≤ K1, then there is severe multicollinearity among the column vectors of matrix A, meaning the orthogonality between the column vectors is weak; if K1 ≤ K f If K ≤ K2, further analysis is needed in conjunction with other criteria; if K fIf K ≥ K2, then there is a certain degree of orthogonality between the column vectors.

[0097] K1, K2, K f The values ​​are shown in Table 1.

[0098] Table 1. Numerical Table of the Least Relative Norm Method

[0099]

[0100] As can be seen from Table 1, K f >K2 indicates that there is a certain degree of orthogonality between the column vectors of the original sample data, that is, each dimension of the original data has good independent information.

[0101] Then, the minimum relative norm method is used to calculate the K of the feature-amplified sample after mapping by the Gaussian kernel explicit mapping function. f , obtain K f The value is 1.3 × 10 -4 The value is less than the lower threshold K1. This indicates that the orthogonality of features weakens in the Gaussian kernel explicit mapping space, which weakens the strength of individual features to some extent. This suggests that the Gaussian kernel explicit mapping function is not suitable for high-dimensional mapping of engine data.

[0102] After calculating the high-dimensional mapping using the polynomial kernel explicit mapping function, the K value of the feature-amplified sample is obtained. f The calculated K f The value is 0.12, which is greater than the upper threshold K2. This indicates that the use of the polynomial kernel explicit mapping function can still maintain a certain degree of orthogonality among the features after high-dimensional mapping. Therefore, the data can utilize useful information from the original feature space in the subsequent classification process after high-dimensional mapping.

[0103] The second approach is an experience-based high-dimensional space modeling method. After analyzing the explicit mapping functions of polynomial and Gaussian kernels, the dimensions of the original samples can be analyzed first. If the original sample dimensions possess a certain degree of orthogonality, the original samples can be combined with amplified new features to form a high-dimensional space, thereby improving the accuracy of fault diagnosis. Therefore, this implementation proposes a new feature amplification method, termed an experience-based high-dimensional space modeling method. Through analysis of… Figures 3 to 6Analysis shows that when an aero-engine experiences a gas path failure, the failure type is correlated with the trend changes of four gas path parameters: DEGT, EGTM, DFF, and DN2. Therefore, the trend changes of these four gas path parameters can be considered as new amplified features. Considering the drastic changes in the original data, this implementation selects the difference term of the original data as a possible amplified feature. Directly differencing the original data may introduce errors. Since LSTM has excellent smoothing and denoising capabilities, the original data is first smoothed and denoised. The difference term is then constructed from the smoothed and denoised data and combined with the original dimensions to form a high-dimensional space.

[0104] Specifically:

[0105] The failure time of faulty engine k is obtained through CNR and maintenance reports. Based on the failure time of faulty engine k, the preceding m steps before the failure of faulty engine k are extracted from OEM data. k The original sample set A is obtained by collecting the exhaust temperature deviation (DEGT), high-pressure rotor speed deviation (DN2), fuel flow deviation (DFF), and exhaust temperature margin change (EGTM) for each flight cycle. k :

[0106]

[0107] In the formula, the i-th original sample x I for:

[0108] x I =[DEGT I DN2 I ,DFF I EGTM I ], I = 1, 2, 3, ..., m k ;m k Integers greater than 10;

[0109] The original sample x is processed using a Long Short-Term Memory (LSTM) network. I Smoothing and denoising processes are performed to obtain smoothed sample data. LSTM deep learning networks can extract effective temporal information from long-span time series and are better suited for processing time series information. An LSTM cell contains three gates: a forget gate, an input gate, and an output gate. The forget gate and the input gate control the neuron state C. t The output gate is composed of the accumulated input h t-1 Current input x t Neuron state C t The results are determined jointly by all parties. A schematic diagram of the LSTM unit structure is shown below. Figure 8 As shown. Figure 8 This is an example diagram showing the smoothing of air circuit parameters in a faulty engine. Figure 8Chinese f t The output of the forget gate; σ represents the sigmoid function; i t is the input gate output result; tanh represents the tanh function; O t This outputs the result of the output gate.

[0110] Due to individual differences in civil aviation engines during actual operation and maintenance, to mitigate the impact of these individual differences on the smoothing and noise reduction of air path parameters, smoothing and noise reduction processing is performed on the independent air path parameters of each individual engine. This is achieved through observation... Figure 9 It can be observed that the fluctuation range of the smoothed value is smaller compared to the original value.

[0111] K-step differencing is performed on the smoothed sample data, which helps to smooth the time series. However, some information loss is inevitable during data fitting and differencing. Therefore, this implementation retains the original data and uses the differencing data as the amplified new features. Based on the analysis of CNR reports from OEM manufacturers, 10 cycles as intervals for fault indicator data can meet most fault diagnosis needs. Therefore, the value of K is chosen as 9, resulting in the amplified sample set A. k ′:

[0112]

[0113] In the formula, DEGT′ represents the DEGT value after LSTM smoothing and denoising, DN2′ represents the DN2 value after LSTM smoothing and denoising, DFF′ represents the DFF value after LSTM smoothing and denoising, and EGTM′ represents the EGTM value after LSTM smoothing and denoising.

[0114] Feature-amplified sample set A k While retaining the original data, the data is expanded from 4 dimensions to 8 dimensions, mapping the low-dimensional data to a high-dimensional space. To demonstrate the orthogonality between features after the high-dimensional mapping, the minimum relative norm method is still used for verification. The calculated K... f The value is 0.42, which is greater than the upper threshold K2, indicating that the experience-based high-dimensional mapping space modeling method can make each feature independent of information.

[0115] Furthermore, the normalized samples are also used to construct the test sample set, with a ratio of 4:1 between the training sample set and the test sample set.

[0116] Furthermore, the sample data input to the fault diagnosis network for each iteration is selected from the training sample set using a sliding window. Analysis of the engine dataset and the public dataset reveals that both datasets are typical multidimensional time-series data. To better capture their sequential characteristics, this implementation uses a sliding window method to extract the samples input to the fault diagnosis network for each iteration, such as... Figure 10 As shown. 4 represents the dimension of the engine dataset, and n represents the dimension after high-dimensional mapping. When using a polynomial kernel explicit mapping function, n = 14; when using an empirical method, n = 8; assuming the sliding window length is l and the sliding window step size is S. When the starting point of the sliding window reaches the q-th column, the starting point is pi, and the ending point is pi+l-1. After obtaining the data segment of the current window, the sliding window will move forward S steps. At this time, the starting point and ending point of the sliding window will become pi+s and pi+l+s-1, respectively. As the window slides forward continuously in the time series data, it will continuously generate data segments. In this embodiment, l is 10 and S is 5.

[0117] As an example, the fault diagnosis network is a Convolutional Neural Network (CNN), a Long Short-Term Memory (LSTM) network, a Temporal Convolutional Network (TCN), or a Deep Residual Shrinkage Network (DRSN-CW). The structural parameters of the four fault diagnosis networks are shown in Table 2. The abbreviations in Table 2 are as follows: Batch Normalization (BN), Rectifier Linear Unit (ReLU), Global Average Pooling (GAP), and Fully Connected Layer (FC). In Table 2, 2 / 4 indicates a value of 2 for binary classification and a value of 4 for multi-class classification. The learning rate for all four network models is set to 0.001. The Adam optimizer is used to update the weights during training, and L2 regularization is used to reduce overfitting. The decay coefficient of L2 regularization is 0.0001, and the batch size is 10.

[0118] Table 2 Parameters of Deep Learning Models

[0119]

[0120] The fault diagnosis process and implementation steps of this invention are as follows:

[0121] Combination Figure 11 As shown, the method steps of the present invention are as follows:

[0122] First, a high-dimensional mapping method based on a polynomial kernel explicit mapping function or an empirical high-dimensional space modeling method is used to map the engine air path state data to a high-dimensional space.

[0123] Step 2: Normalize the input of the sample to reduce the interference caused by the difference in magnitude between parameters;

[0124] Step 3: Construct training and testing sample sets, setting the label for engine normal operating status data to 0. For binary classification, set the label for all fault sample data to 1; for multi-class classification, assign different labels to different fault types. The ratio of training to testing sets is 4:1.

[0125] Step 4: Train the LSTM, TCN, CNN or DRSN-CW network using the training sample set. After the network training is completed, keep the network parameters unchanged.

[0126] Step 5: Input the test set into the trained network model, compare the classification results with the labels in the test set, and output the network model's accuracy, precision, and F1 score in the test set.

[0127] The following experiments validate the engine dataset: To verify that the high-dimensional mapping method based on the polynomial kernel explicit mapping function and the empirical high-dimensional space modeling method can improve the separability of the data, actual engine operation and maintenance datasets and bearing vibration datasets are used for validation. For ease of description, the high-dimensional mapping method based on the polynomial kernel explicit mapping function is referred to as Scheme 1; the empirical high-dimensional mapping space modeling method is referred to as Scheme 2. Convolutional autoencoders (CAE) and autoencoders (AE) are selected as high-dimensional mapping comparison methods. Therefore, in the experiment, Scheme 1, Scheme 2, CAE, and AE are used as feature augmentation methods, and Raw data represents the original data without feature augmentation.

[0128] Since the high-dimensional mapped data obtained by CAE and AE are difficult to calculate using the minimum norm method, their orthogonality is not analyzed. The experimental results of fault diagnosis using actual engine operation and maintenance datasets are shown in Tables 3 and 4, and the corresponding bar charts are as follows: Figures 12 to 15 and Figures 16 to 19 .

[0129] Table 3. Correctness of binary classification of engine maintenance data

[0130]

[0131] Table 4. Accuracy of Multi-Classification of Engine Maintenance Data

[0132]

[0133] For the binary fault diagnosis task, analyze Table 3 and... Figures 12 to 15 As can be seen, the feature augmentation methods (Scheme 1 and Scheme 2) proposed in this embodiment achieve better classification results than other methods on all three evaluation metrics (Accuracy, Precision, F1-score) for the four classifiers (CNN, LSTM, TCN, DRSN-CW). When using CNN as the classifier, Scheme 2 achieves the best classification result, Scheme 1 achieves the second best classification result, and CAE achieves the worst classification result; when using LSTM as the classifier, Scheme 1 achieves the best classification result, Scheme 2 achieves the second best classification result, and Raw data achieves the worst classification result; when using TCN as the classifier, Scheme 2 achieves the best classification result, Scheme 1 achieves the second best classification result, and CAE achieves the worst classification result; when using DRSN-CW as the classifier, Scheme 1 achieves the best classification result, Scheme 2 achieves the second best classification result, and CAE achieves the worst classification result.

[0134] For multi-class fault diagnosis tasks, analyze Table 4 and... Figures 16 to 19 It can be seen that Scheme 1 and Scheme 2 still outperform other methods in classification results. When using CNN as the classifier, Scheme 2 achieved the best classification result, Scheme 1 achieved the second best, and AE achieved the worst. When using LSTM as the classifier, Scheme 1 achieved the best, Scheme 2 achieved the second best, and AE achieved the worst. When using TCN as the classifier, Scheme 1 achieved the best, Scheme 2 achieved the second best, and CAE achieved the worst. When using DRSN-CW as the classifier, Scheme 1 achieved the best, Scheme 2 achieved the second best, and CAE achieved the worst. Through the above analysis, it can be concluded that, regardless of whether it is binary classification or multi-class classification, under the same deep learning model, the classification results of Scheme 1 and Scheme 2 are superior to those of Raw data, CAE, and AE. Even Scheme 1 and Scheme 2 both achieved scores greater than 90% on all three metrics of the TCN deep learning model.

[0135] The above analysis shows that, in both binary and multi-class classification tasks, regardless of which of the four models is used as the final fault diagnosis classifier, using the two feature augmentation methods proposed in this paper to augment the features of the original samples before training the classifier for fault diagnosis yields significantly better diagnostic results than other methods. Experimental results demonstrate that using the proposed feature augmentation methods to map the original samples to a suitable high-dimensional space effectively improves the separability of the samples while avoiding the curse of dimensionality.

[0136] In terms of computation time, the computation time required after high-dimensional mapping of the data is greater than that of the original data. This is because, with the same number of input samples, the larger the dimension of the samples, the greater the computation time required. Additionally, since Scheme 2 requires data smoothing and denoising, this step further increases the computation time, but the maximum required time is only 166.3 seconds. This time includes the model's training and testing time. After the model training is complete, the actual testing time is very short, not exceeding 5 seconds.

[0137] Experimental verification using public datasets:

[0138] To further demonstrate that the high-dimensional mapping method for Scheme 1 and Scheme 2 data can effectively improve the classification results of fault diagnosis, a bearing vibration dataset publicly available from the Bearing Data Center at Case Western Reserve University was used for validation. Before conducting experimental validation, the orthogonality of the high-dimensional mapping of Scheme 1, Scheme 2, and raw data was analyzed. The K0 of the raw data... f The value is 0.99, K after the high-dimensional mapping of Scheme 1. f The value is 0.89, K after the high-dimensional mapping of Scheme 2. f The value is 0.62, which is greater than the upper threshold of 0.1. This indicates that the features after high-dimensional mapping of Scheme 1, Scheme 2, and Raw data all possess strong orthogonality. The experimental results of the bearing vibration dataset are shown in Table 5, and the corresponding bar chart is shown below. Figures 20 to 23 As shown.

[0139] Table 5. Multi-class classification accuracy of the bearing vibration dataset.

[0140]

[0141] Analysis of Table 5 shows that, compared to the engine dataset, the bearing dataset generally shows improved classification performance across all four classifiers. This is because the bearing dataset is based on data obtained under laboratory conditions, while the engine dataset is based on data obtained under actual engine operating conditions. In terms of data quality, the bearing dataset is superior to the engine dataset. However, based on Table 5 and... Figure 10Analysis shows that Scheme 1 and Scheme 2 outperformed Raw data, CAE, and AE in classification results across four classifiers across different evaluation metrics. This aligns with conclusions drawn from actual engine maintenance datasets, further demonstrating that the high-dimensional mapping methods of Scheme 1 and Scheme 2 can effectively improve fault diagnosis classification results. Further analysis of Scheme 1 and Scheme 2 reveals that both methods augment features based on the original data, preserving a certain degree of orthogonality after high-dimensional mapping. This indicates that in the process of mapping low-dimensional data to a high-dimensional space, if the original data exhibits strong orthogonality and retains a certain degree of orthogonality after high-dimensional mapping, then high-dimensional mapping can effectively improve fault diagnosis classification results.

[0142] While the invention has been described herein with reference to specific embodiments, it should be understood that these embodiments are merely examples of the principles and applications of the invention. Therefore, it should be understood that many modifications can be made to the exemplary embodiments, and other arrangements can be designed without departing from the spirit and scope of the invention as defined by the appended claims. It should be understood that different dependent claims and features described herein can be combined in ways different from those described in the original claims. It is also understood that features described in conjunction with individual embodiments can be used in other described embodiments.

Claims

1. A method for fault diagnosis of aero-engines based on feature amplification, characterized in that... include, Step 1: Perform high-dimensional feature amplification on the original sample to obtain the feature-amplified sample; Step 2: Normalize the amplified samples and then construct a training sample set from the normalized samples; assign different labels to the normal state samples and fault samples in the training sample set. Step 3: Train the fault diagnosis network using the training sample set. After reaching the preset number of iterations, the trained fault diagnosis network is obtained. Step 4: Collect the operating data of the aero-engine, and perform the high-dimensional feature amplification described in Step 1 and the normalization processing described in Step 2 to obtain the normalized data to be diagnosed; input the normalized data to be diagnosed into the trained fault diagnosis network to obtain the aero-engine fault diagnosis results. The method for high-dimensional feature augmentation of the original sample in step one includes a high-dimensional mapping method based on a polynomial kernel explicit mapping function, specifically: Using polynomial kernel functions For the original sample and original sample Perform inner product operations. : ， In the formula This represents the mapping function corresponding to the polynomial kernel function, where r and d are the undetermined parameters in the polynomial kernel function, respectively. The i-th original sample The j-th original sample , where n represents the dimension of the sample, and the total number of original samples is N, i=1,2,3,……N, j=1,2,3,……N; The dimension is In order to reduce Dimensionality and preserve the original sample Given the information, choose r=1, d=2, then The expression is: ， Will The constant term 1 and coefficient in the expression Discard, and record as : ， Will As a sample after feature amplification; The method for high-dimensional feature augmentation of the original sample in step one also includes an empirical high-dimensional space modeling method, specifically: Based on the failure time of faulty engine k, extract the time preceding the failure of faulty engine k from the OEM data. The original sample set was obtained by collecting the exhaust temperature deviation (DEGT), high-pressure rotor speed deviation (DN2), fuel flow deviation (DFF), and exhaust temperature margin change (EGTM) for each flight cycle. : ， In the formula, the I-th original sample for: , ; Integers greater than 10; The original samples were processed using a Long Short-Term Memory (LSTM) network. Smoothing and denoising processes are performed to obtain smoothed sample data; K-step differencing is then performed on the smoothed sample data, with K set to 9, to obtain the feature-amplified sample set. : ， In the formula Indicates the noise reduction after using LSTM smoothing. value, Indicates the noise reduction after using LSTM smoothing. value, Indicates the noise reduction after using LSTM smoothing. value, Indicates the noise reduction after using LSTM smoothing. value.

2. The aero-engine fault diagnosis method based on feature amplification according to claim 1, characterized in that, In step one, the original sample was extracted from the engine air circuit performance monitoring data.

3. The aero-engine fault diagnosis method based on feature amplification according to claim 1, characterized in that, The normalized samples are also used to construct the test sample set, with a ratio of 4:1 between the training sample set and the test sample set.

4. The aero-engine fault diagnosis method based on feature amplification according to claim 3, characterized in that, Each input sample data of the fault diagnosis network is selected from the training sample set through a sliding window.

5. The aero-engine fault diagnosis method based on feature amplification according to claim 4, characterized in that, The fault diagnosis network is a convolutional neural network (CNN), a long short-term memory network (LSTM), a temporal convolutional network (TCN), or a deep residual shrinking network (DRSN-CW).

6. The aero-engine fault diagnosis method based on feature amplification according to claim 5, characterized in that, The learning rate of the fault diagnosis network is set to 0.001, the Adam optimizer is used to update the weights, and L2 regularization is used to reduce overfitting; the decay coefficient of L2 regularization is 0.0001, and the batch size is 10.

Images