Feature Optimization for Machine Learning Based Bearing Fault Classification

The most critical and essential parts of rotating machinery are bearings. The main problem of the bearing fault classification is to select the fault features effectively because all extracted features are not useful, and the high-dimensional features give poor performances and slow down the training process. Due to the effective feature selection problem, the bearing fault diagnosis method does not achieve a satisfactory result. The main goal of this paper is to extract the effective fault features with an optimization technique to classify the bearing faults using machine learning algorithms. Since wavelet entropy can determine complexity and degree of order of a vibration signal, this research uses it in features optimization. The proposed wavelet entropy-based optimization technique reduces the dimensionality of input, elapsed time and raises the learning process. Four Machine learning algorithms (naïve Bayes, support vector machine, artificial neural network and KNN) are applied to classify the bearing faults using the optimized features. To evaluate the proposed method, Case Western Reserve University’s (CWRU’s) bearing dataset is used which consists of three types of bearing faults. The accuracy and robustness of the bearing fault classification are tested by adding noise to the vibration raw signals at various levels of Signal-to-Noise Ratio (SNR). Experimental results show that the proposed method is very highly reliable in detecting bearing faults compared to the conventional methods.