Performance Analysis of Classification Algorithms for Fault Diagnosis in Rotating Machines

Abstract

Classification of any given vibration signal as healthy or faulty can be done by employing classification algorithms available to us. Identification of a fitting classification algorithm is a task that should be done at the time of identification of the problem statement itself, such that required changes can be done in it if the need be. Hilbert Huang Transform (HHT) empowered Adaptive Neuro-Fuzzy Inference System (ANFIS) was used to obtain the most significant features of the vibration signals of both healthy and faulty rotating machines in the time and frequency domain, namely RMS velocity, Kurtosis, and Crest Factor (RKC). They were then fed to classification algorithms to classify the machines as healthy or faulty. Five machine learning techniques such as Probabilistic Neural Network (PNN), decision tree (DT), k- nearest neighbour (KNN), and Radial Basis Network (RBN) are utilized as classification algorithms. Decision Tree algorithm was found to be the optimal classification technique; overfitting was found to be a notable issue. To improve prediction, the decision tree algorithm was parallelly ensembled into Random Forest using the Bootstrap Aggregation method.