summary
In view of the problems that the existing data-driven aeroengine fault diagnosis algorithm is susceptible to the interference of redundant features and noise in the flight monitoring data, and cannot timely correct the influence of the unbalanced sample distribution in the monitoring data on the generalization performance of the model, a fault diagnosis method for aero engine gas path based on feature optimization and support vector machine is proposed by introducing feature dimensionality and extraction algorithm into the support vector machine model, and the corresponding model is established. The after-flight data of turboprop engine and CFM56-7B engine were input into the model, the actual fault occurrence time was analyzed and predicted, and the prediction results were compared with the real results, and the results were compared with the results obtained by four fault diagnosis methods such as random forest. The results show that the application of feature optimization algorithm can significantly shorten the running time of various fault diagnosis methods by more than 20%. The fault diagnosis method based on feature optimization and support vector machine makes the prediction accuracy reach 99.8%. For the unbalanced measured data, the introduction of feature optimization algorithm and regression prediction idea can significantly improve the performance of the algorithm on the unbalanced dataset, and the fault detection rate is increased to 91.67% compared with the non-regression algorithm.
keyword
fault diagnosis; feature optimization; support vector machines; principal component analysis; deep autoencoder; Aero engines
PART/0
introduction
Aero engine is a complex thermodynamic machine that provides propulsion power for aircraft flight, and in the working process, many components are in a state of high speed, high stress, high temperature and high pressure for a long time, which makes the time of engine failure uncertain. As an important part of the aircraft, the operating status of the engine directly affects its safe service. Therefore, timely and accurate engine fault diagnosis and management can not only achieve condition-based maintenance and reduce maintenance costs, but also improve flight safety. In conclusion, through the research and development of accurate and reliable engine fault diagnosis technology, the safety and economy of aircraft operation can be effectively enhanced [1].
In the 80s of the 20th century, the development of artificial neural networks provided a reliable analytical tool for the diagnosis of air circuit faults of aero engines. By making full use of the knowledge, experience and historical data in the engine field, including algorithms based on neural networks, support vector machines, fuzzy logic, expert knowledge systems, and deep kernel extreme learning machines [2], data-driven engine gas fault diagnosis methods have been greatly developed. Kyriazis et al. [3] combined probabilistic neural networks with fuzzy logic to enhance the ability to identify gas path faults. Sarkar et al. [4] recognized the different characteristics of sudden failure and recessionary failure of engine, proposed a prediction method based on symbolic dynamic filtering model, and analyzed time series data to distinguish the two types of faults. Lin et al. [5] converted the original fault signal into a spectral entropy image and input it into a convolutional neural network, which improved the real-time performance of the fault diagnosis algorithm. Cui et al. [6] pre-trained the deep autoencoder network by using the unlabeled data sample set to assist the Softmax classifier in diagnostic decision-making. Cao et al. [7] used the Gaussian kernel function support vector machine as a weak classifier of the AdaBoost algorithm to realize the fault diagnosis of multi-classification gas path, which effectively solved the problems of nonlinearity and insufficient learning samples. Tian et al. [8] introduced the genetic algorithm and applied it to the tuning of regularization parameters and kernel parameters, which can adaptively select the parameters of the least squares support vector machine to improve the anti-noise ability of the fault diagnosis model. However, the existing problems limit the popularization and application of data-driven engine gas circuit fault diagnosis methods. On the one hand, due to the influence of sensor accuracy and environmental interference, flight monitoring data often contains more redundant features and noise, which increases the burden of model training and reduces the accuracy of fault diagnosis. On the other hand, the number of normal samples in the monitoring data collected during the actual operation of the engine is much larger than the number of fault samples, and the extremely uneven sample distribution affects the learning process of the fault diagnosis method and weakens the generalization ability of the model.
In this paper, we propose a method for diagnosing aeroengine gas circuit faults based on feature optimization and support vector machine, which adopts the method of increasing fault data by fault probability estimation and combining the idea of regression prediction to diagnose and predict the gas path fault of engine post-flight data.
PART/1
An overview of troubleshooting techniques
结合特征优化和支持向量机（Support Vector Machine，SVM）的气路故障诊断模型结构如图1所示。
The direct use of the original dataset as the model input has the following problems: the redundancy and correlation between the original features are not conducive to the training of the machine learning model; The subset of features composed of important features in the dataset plays a leading role in the diagnosis of fault problems. Therefore, the dimensionality enhancement and optimal subset selection of the feature subset in the original sample space can optimize the operation efficiency of the model and improve the diagnostic accuracy.
The method proposed in this paper uses exponential, logarithmic, and combination methods to increase the feature dimensionality of the preprocessed engine post-flight data. In order to select the feature combinations with discrimination from the dimensioned data, the Near Component Analysis (NCA) algorithm and the Auto Encoder (AE) feature analysis and extraction methods were introduced, and the Principal Component Analysis (PCA) method was used to compare the results obtained based on the NCA and AE feature optimization algorithms. Select more separable fault characteristics for fault diagnosis.
1.1 Feature optimization algorithm
In this paper, the feature optimization algorithm is used to evaluate and analyze the engine state characteristics after feature dimensionality, which provides a basis for subsequent feature selection. The basic task of feature optimization is to evaluate the existing features and find out the most effective features by using feature selection and feature extraction algorithms. Feature selection is the process of selecting N (N<M) features from the existing M features to form a subset of features, and achieving the follow-up task indicators by optimizing the feature information. Feature extraction is to generate new features on the basis of the original feature set to help solve the problem.
1.1.1 Nearest Neighbor Component Analysis (NCA) algorithm
The learning of distance metrics in the feature space can greatly improve the performance of the classifier and has high practical application value. As a typical distance measure learning algorithm, the nearest neighbor component analysis method has been applied in the fields of medical detection [9], speech recognition [10], and face recognition [11]. The NCA algorithm randomly selects neighbors in the data samples, and searches for the transformation matrix by optimizing the Leave-One-Out (LOO) method, and uses the matrix to obtain the low-dimensional embedding representation of the original data.
1.1.2深度自编码器（AE）
Deep autoencoder [12] is a type of neural network, which directly uses a single-layer or multi-layer neural network to map the input data, and the converted output vector is used as the feature extracted from the input data. Traditional autoencoders are generally used in tasks such as data dimensionality reduction [13], text extraction [14], and speech recognition [15], which can represent linear and nonlinear transformations at the same time, and have higher adaptability than other dimensionality reduction methods.
The feature compression model of the deep autoencoder is shown in Figure 2. As can be seen from the figure, a typical autoencoder model consists of a simple 3-layer neural network structure, including an input layer, a hidden layer, and an output layer. The input layer and the hidden layer form the coding network h=f(x),
1.2 Support vector machines
Support Vector Machines (SVMs) are commonly used supervised learning models in classification and regression analysis, which have been applied to protein structure prediction [16] and fault diagnosis [17]. The SVM represents instances as points in space and searches the space for a hyperplane that maximizes a certain value, so that instances of a separate class are separated by as wide spacing as possible. If there is a new instance, map it to the same space and predict the category it belongs to based on the location of the interval. The minimum distance between the hyperplane and all instances is called the spacing. The SVM algorithm searches for a separated hyperplane in space, as shown in Figure 3. w·x+b=0 is the separated hyperplane, which is infinite for linearly separable datasets, but the separated hyperplane with the largest geometric interval is unique.
2. Data preprocessing and feature optimization
The parameters of the aero engine gas path performance measurement described in this paper are derived from the simulation data generated by the turboprop engine gas performance model, the CFM56-7B engine post-flight data, and the maintenance records [18-19]. The gas circuit performance simulation model simulates the operation of the turboprop engine in various fault scenarios such as normal state, single component failure, bleed valve failure, and combination failure, and compares the data samples obtained in different failure modes with the health baseline [18] to record the abnormal fluctuation of the engine state. The measurement parameters and related variables of each gas path involved in this paper are shown in Tables 1 and 2. Individual faults (component failure or bleed valve failure) and their corresponding variation in the quantity of individual performance parameters (other performance parameters remain unchanged) are shown in Table 3. Table 4 shows the changes in the combined faults and their corresponding performance parameters. In order to make the simulation data closer to the real flight data, a certain order of magnitude of noise is added to the simulation process of the model.
2.1 Data preprocessing and feature dimensionality
Due to the fact that there are many types of actual engine faults and have different degrees of influence on different measurement parameters, it may not be possible to accurately diagnose faults by directly using the existing measurement parameters as fault characteristics modeling. Therefore, the feature dimensionality method is adopted to increase the fault features, and then the fault features are optimized to mine the hidden information of the fault, so as to improve the accuracy of engine fault diagnosis and the effectiveness of the algorithm.
Feature dimensionality generally increases the number of fault features in the form of linear combination, logarithmic transformation and exponential transformation on the basis of existing features. In this paper, various types of feature enhancement are performed on the existing fault features, including exponential, logarithmic, differential, and combinatorial. The specific characteristics of the two types of aero engines are shown in Table 5 and 6 respectively. In general, the parameters with correlation should be selected for combination of combinatorial dimensionality.
2.2 Feature selection and extraction
The feature augmentation method not only enhances the expression ability of the original features, but also introduces more redundant features to the dataset. In order to eliminate the influence of redundant features, reduce the storage space of data, and improve the performance of the model, the feature selection method and feature dimensionality reduction method were used to optimize the selection of the obtained features.
Firstly, the NCA algorithm is used to select the features of the two types of aeroengine datasets after feature dimensionality, and the algorithm uses machine learning to score each feature to avoid the disconnection between the feature selection results and the actual model results. Before using the NCA algorithm to obtain the weight vector, it is necessary to optimize the regularization term λ and select the minimum parameter of the average loss. By selecting the λ value of the objective function at the minimum point, the corresponding weight vector w is obtained as the feature ranking result. The importance scores (score > 0.1) obtained by using the NCA algorithm on turboprop engines and CFM56-7B engines are shown in Fig. 4 and 5, respectively.
According to the feature weight values indicated in the above figure, the combination of features after dimensionality that should be retained is screened out, as shown in Table 7.
In addition, similar to the process in feature selection, the deep autoencoder model is selected as the feature dimensionality reduction tool, and the two types of aeronautical features after feature dimensionality enhancement are also used as inputs. By establishing a typical three-layer neural network, the autoencoder can use the hidden structure of the model to perform feature dimensionality reduction and feature extraction, and map the input features into low-dimensional features.
2.3 Principal component analysis algorithm evaluates feature optimization results
In order to select the most suitable dataset for feature optimization among various feature selection and feature dimensionality reduction methods, the results of various feature optimization methods were evaluated by introducing the PCA method into the model [20]. PCA is the most typical tool for feature extraction for pattern classification. As a classical feature extraction method, PCA transforms the original dataset into "effective" feature components with fewer dimensions without reducing the information contained in the original data, so that it can achieve the purpose of optimal variance under the statistical mean square meaning. At the same time, with the advantage of PCA's easy visualization, the datasets processed by the NCA algorithm and the deep autoencoder are displayed on the two-dimensional plane to evaluate the classification ability and advantages and disadvantages of fault samples.
Two feature optimization methods were used to analyze the results on different engine gas performance simulation datasets. Taking the turboprop engine dataset as an example, the PCA visualization results of the original dataset and the two types of datasets optimized by the two methods are shown in Figure 6~8.
Comparing the sample distribution in the three types of datasets, it is found that compared with the simulation dataset of the original turboprop engine, the feature subset optimized by NCA algorithm and deep autoencoder can achieve better separation between the fault samples and the normal samples. Because the feature subset obtained by the NCA algorithm has better separation of different types of gas path faults, it is more suitable for gas path fault diagnosis.
PART/3
Fault diagnosis and analysis results
Based on the optimized aeronautics dataset with the above characteristics, a support vector machine model is constructed to carry out gas path fault diagnosis, and the effectiveness of the proposed fault diagnosis method is verified by comparison with other common classifiers.
3.1 Fault diagnosis based on feature optimization and support vector classification model
After the importance score obtained by the NCA algorithm > 0.1, the fault diagnosis was carried out in the feature input SVM classifier. For turboprop engines, the input characteristics are △PCNC, △T8M, △WFB-1, △PCNC2, and △PCNC/△T8M, five characteristics (Table 7). The SVM classifier in the model uses the RadialBasisFunc‐tion (RBF), and the dataset is divided into training sets and test sets according to the ratio of 7:3.
In order to verify the effectiveness of the fault diagnosis algorithm based on feature optimization and support vector machine, the proposed algorithm is compared with several common fault diagnosis algorithms on the turboprop engine simulation dataset.
The statistical analysis results of four fault diagnosis algorithms, Random Forest (RF) [21], Perceptron Model (PM) [22], Decision Tree (DT), and Support Vector Machine (SVM), on the turboprop engine dataset are shown in Table 8. According to the feature optimization algorithm in this paper, four variants of the above model were designed for joint experiments. Among them, ORIG means that the algorithm uses original features to train the model; FO indicates that the algorithm uses a subset of features filtered by the feature optimization algorithm to train the model.
From the experimental results in the table, it can be seen that the feature optimization method can significantly reduce the running time of various methods and improve the accuracy of all fault diagnosis algorithms except the perceptron. Among them, from the fault diagnosis performance analysis: the fault diagnosis algorithm based on feature optimization and support vector machine is obtained, and the highest fault detection rate (Falsediscoveryrate, FDR) of 99.80% is obtained. In the experiment, the diagnostic accuracy of each method for non-fault samples is high, and only one normal sample is misclassified in the prediction process, so the false alarm rate (FAR) is 0.658%. In addition, the perceptron model was significantly underfitted in the experiment. The accuracy of the ORIGPC algorithm in the turboprop engine dataset is less than 65%, indicating that the algorithm is in an underfit state and it is difficult to learn the data information. In FOPC, the feature selection algorithm simplifies the feature items of the input data, which makes the underfitting phenomenon more serious, which leads to the further decline of the FDR index.
From the analysis of model running efficiency, compared with the diagnostic model using the original features, the running time of various methods based on feature optimization is shortened. In the feature-optimized model, the operation efficiency of the SVM model is only lower than that of the decision tree algorithm. Since the decision tree algorithm is a tree-branch-based method, it is not necessary to calculate the similarity between features, etc. The support vector machine algorithm selected in this paper is the RBF function kernel to map the input low-dimensional sample features to the high-dimensional space, and make classification decisions by measuring the similarity between the samples. Based on the accuracy and operation efficiency of the model, the algorithm proposed in this paper has good fault diagnosis performance.
In order to evaluate the reliability of the evaluation results of the PCA algorithm, this paper explores the influence of the selection of feature optimization algorithm on the evaluation results. The classification results of the comparison model after training with the optimized feature subsets based on NCA and AE algorithms are shown in Table 9. Among them, AE represents the use of features optimized by the auto-encoding algorithm to train the model; NCA means that the algorithm uses the 5-dimensional features screened by the nearest neighbor component analysis algorithm to train the model. Compared with the NCA algorithm, the optimized fault diagnosis model based on the autoencoder AE algorithm requires a longer running time, and the accuracy of 3 of the 4 models is very low. Therefore, the optimization quality of each feature optimization algorithm evaluated by the PCA algorithm is consistent with the performance in the actual fault diagnosis.
3.2 Non-equilibrium fault diagnosis algorithm based on feature optimization and support vector regression model
Due to the huge difference between the number of samples in the two categories of faults and normal in the dataset obtained in reality, the number of normal samples is much larger than the number of fault samples, which is a typical unbalanced dataset. Taking the data of CFM56-7B turbofan engine as an example, among the 1055 valid samples, only 13 samples of a single gas circuit failure were included, that is, the engine flow path was washed due to the decrease of its overall performance index (exhaust temperature margin) due to the accumulation of dirt. The setting of the washing fault label, that is, the relationship between the probability of engine failure and the maintenance record, is shown in Figure 9, and it can be seen from the figure that the engine has produced fault signs for a period of time before the engine is washed, and it is set to the fault state. The state before and after washing is considered to be healthy. The N flight cycles before washing are set as the fault state, which is used as the sample required for the subsequent fault diagnosis algorithm. Generally, the number of cycles is determined by repeated tests, and the number of cycles is used as a variable to test the fault diagnosis accuracy under each cycle, and finally the number of cycles with higher accuracy is selected.
In this paper, the characteristics of the imbalance dataset are considered, and engine gas circuit failures can generally be regarded as the result of engine gas circuit performance degrading to a certain extent, that is, the engine failure process is regarded as different degradation states [24]. In the above-mentioned fault diagnosis algorithm, the fault sample data is extended and generated according to the engine fault maintenance record, and for the engine gas circuit fault, in addition to the injury of foreign objects, in the actual flight process, the closer to the fault maintenance point, the greater the possibility of the fault occurring. Therefore, after optimizing the characteristics of the original data, a certain confidence interval is set for the fault point and part of the data before the fault point, and the normal distribution is used to fit it, so as to expand the fault sample data. Then, the 4-dimensional feature data extracted by the NCA algorithm includes four features: Diff(△EGT), Diff(△N2), Diff(△EGT) and Diff(△N2), and the SupportVectorRegression (SVR) is used to diagnose the fault, and the appropriate classification threshold is set according to the results of the fault diagnosis model in the training set.
The fault diagnosis results of the diagnostic algorithm based on feature optimization and SVM and the diagnostic algorithm combined with regression prediction on the CFM56 engine dataset are shown in Table 10. From the experimental data, it can be seen that the fault diagnosis algorithm combined with regression prediction can significantly improve the fault detection accuracy on the unbalanced dataset.
In order to further demonstrate the accuracy of the fault diagnosis algorithm combined with regression prediction on the analysis of the sample failure rate, one fault point in the test data and 80 points before the fault point were selected for observation, and the failure probability change curve was plotted, as shown in Figure 10. As can be seen from the figure, the probability of failure predicted by the SVR model increases as the time is closer to the fault point. The visual failure probability curve shows that the diagnostic algorithm combined with regression prediction has a good fitting effect on the prediction of fault probability.
PART/4
conclusion
(1) Compared with other fault diagnosis algorithms, the fault diagnosis model based on feature optimization and SVM established in this paper has the best accuracy and short running time.
(2) For the two common types of unbalanced datasets, this paper combines the maintenance records, expands the fault data sample, and combines it with the idea of regression prediction to improve the accuracy of the diagnostic algorithm in such cases.