Engine components fault diagnosis using an improved method of deep belief networks

In the aeronautical industry, there are multiple diagnosis methods available to improve the accuracy in diagnosing malfunctions due to degradation of rotating components in an engine. This study introduces an improved deep learning algorithm - ad_DBN, it is based on the Deep belief network (DBN) method, mimics the multilayer structure of a human brain. In addition to this, it adjusts connection weights adaptively during unsupervised and supervised learning phase. A comparison study of its performance to other methods, such as back propagation (BP) method, Radical Basis Function (RBF) method and Support Vector Machine (SVM) method is carried out. This is used to demonstrate the efficacy of the proposed algorithm to detect faulty component(s) in an aeronautical turboshaft engine. Three aspects are compared in this study: diagnosis accuracy, noise filtering capability and training time. It was found that the ad_DBN algorithm has a supreme diagnosis accuracy of 96.84%, better noise filtering ability but longer training time than other methods.