Intelligent monitoring of ball bearing conditions

Abstract Ball bearings are widely used in various kinds of robots, manufacturing machines, and equipment. In order to enhance productivity and improve product quality, an on-line monitoring system is essential to check the status of ball bearings. In this work, peak amplitude in the frequency domain, peak RMS, and the power spectrum were used as indirect indices to develop a system for monitoring and classifying ball bearing defects. These indices were then processed by artificial neural networks. Six different cases of ball bearing states were observed. The data from these observations were then input into neural networks with different architectures to train these neural networks in a learning process. All the trained neural networks are capable of distinguishing the normal bearings from defective bearings with a 100 percent success rate. They can also classify the bearing conditions into six different states with a success rate of up to 97 per cent. The effects of training set sizes and neural network structures on the monitoring performance have also been investigated.