Research On Civil Aviation Engine Fault Diagnosis And Life Predication Based On Intelligent Technologies

Fault diagnosis and life predication are key parts of civil aviation engine heath management system, and are the important means to achieve on-condition maintenance. Civil aviation engine fault diagnosis and life predication based on intelligent technologies are studied in this dissertation, in which engine run data are used in the study of fault detection and life predication and simulated data are used in that of fault classification/isolation. The main jobs are as follows:1) The study on fault detection is close related to civil aviation engineering. Firstly, traditional pattern recognition methods such as probability estimation, k-nearest neighbors (k-NN) and k-means dynamic clustering are introduced into civil aviation engine fault detection. Secondly, self-organizing map (SOM) networks are firstly used to extract the representative samples of engine normal conditions and detect faults. Finally, the hyper-sphere model based on support vector machines (SVM) and the one-class linear programming classifier based on dissimilarity representation are applied to civil aviation engine fault detection, and model selections are done by validation method. The study shows that methods including probability estimation and nearest neighbors are easy to implement while the parameters in the MOG (mixture of Gauss) model and k-NN model must be properly chosen, k-means clustering and SOM use representative samples to describe engine data and SVM based hyper-sphere model and linear programming based one-class classifier build the bound of engine data for fault detection. Among theses methods, linear programming based one-class classifier is the best choice as far as the diagnostic result is concerned.2) SVM is applied to civil aviation engine fault classification research, and the diagnostic model based on the support vector classifier with penalty item (C-SVC) is proposed. This diagnostic model features three points, the first is that it uses one-against-one method for multi-class classification, the second is that cross validation is used for model selection and the third is that three most possible faults can be given. The research shows that the diagnostic accuracy reaches 93.6% while the standard deviation of measurement noise is three times larger than the normal.3) To avoid subjective selection of the architechture of the feed-forward network (FFNN), the FFNN under the Bayesian evidence framework is firstly applied to civil aviation engine fault classification, and three-layer feed-forward network models with different neurons in the hidden layer are evaluated quantitatively and the best among candidate models is selected for diagnosis.4) After the problem of Least squares SVM (LS-SVM) classifier under Bayesian evidence framework in published papers is analysized, LS-SVM classifier under Bayesian evidence framework is derived and its model selection algorithm is given. The study shows the given algorithm is feasible. Then the diagnostic model based on LS-SVM classifiers is proposed for civil aviation engine fault diagnosis. This model uses one-against-all method for multi-classification, three level Bayesian inferences for model selection and logic block for decision.5) The problem of LS-SVM regression under Bayesian evidence framework in published papers is analysized, then the derivation of LS-SVM regression under Bayesian evidence framework is given and the prediction with error bars is derived using the covariance matrix of model parameters. Error bars can indicate the predication accuracy. Their feasibility is demonstrated by toy examples. Finally, the civil aviation engine life prediction model based on LS-SVM regression is built using the test cell data of overhauled engines.