Novel aeroengine fault diagnosis method based on feature amplification

Abstract

Data-driven fault diagnosis methods have high requirements for data samples. The ideal state is that the input samples have good separability. However, because of the harsh working environment and complex operating conditions, the collected data are highly nonlinear inseparability. The existing methods often cannot achieve a good classification effect. To solve this problem, two feature amplification methods that make full use of the useful information from the aeroengine service data are proposed. One is a high-dimensional mapping method based on explicit mapping of kernel function to amplify features. The other is an experiential method to amplify features. Both methods map the input samples to high-dimensional space to make them more separable in high-dimensional space and retain the useful information of raw data. Each feature after feature amplification still covers strong independent information that can describe fault feature information from different dimensions. Then, four deep learning algorithms that have a good effect on processing time-series data are selected as the classifier. The aeroengine service dataset and the bearing vibration dataset from Case Western University are used to verify the effectiveness of the two feature amplification methods. The experimental results show that the fault diagnosis accuracy can be improved if sample features after high-dimensional mapping possess good orthogonality, otherwise, the accuracy will be reduced.