Real Time Bearing Defect Classification Using Time Domain Analysis and Deep Learning Algorithms

Structural Health Monitoring (SHM), especially in the field of rotary machinery diagnosis, plays a crucial role in determining the defect category as well as its intensity in a machine element. This paper proposes a new framework for real-time classification of structural defects in a roller bearing test rig using time domain-based classification algorithms. Along with the bearing defects, the effect of eccentric shaft loading has also been analyzed. The entire system comprises of three modules: sensor module – using accelerometers for data collection, data processing module – using time-domain based signal processing algorithms for feature extraction, and classification module – comprising of deep learning algorithms for classifying between different structural defects occurring within the inner and outer race of the bearing. Statistical feature vectors comprising of Kurtosis, Skewness, RMS, Crest Factor, Mean, Peak-peak factor etc. have been extracted from the 1-D time series data for different defect cases. These features are then fed as input vectors to algorithms comprising of Support Vector Machines (SVM’s) and Multi-layered Perceptron (MLP) for defect classification. A dedicated hardware setup has been built to test the efficiency of the developed algorithms in real-time. These algorithms have been evaluated based on two criteria – examining the simultaneous defect classification accuracy for two sets of bearings and individually monitoring the class labels for a particular defect. It was observed that the developed framework was able to classify between different bearing defects with a classification accuracy of 97.8%.