Intelligent Prediction of Multiple Defects in Rolling Element Bearing using ANN Algorithm

Rolling Element Bearing (REB) failure in heavy rotating machines or assembly lines might cause the machine to shut down, affecting the total cost and quality of the product. Bearing condition monitoring prevents failure and reduces the time and expense of preventative maintenance. With the advent of machine learning in the area of condition monitoring of mechanical systems, Artificial Neural Network (ANN) is becoming more widely used, redefining the state-of-the-art in defect diagnosis and classification. This work provides a real-time online fault detection approach for rolling bearings based on the ANN algorithm in order to realize the predictive maintenance of rolling element bearings in the industry. The training of the fault diagnostic model and the real-time fault diagnosis are the two primary steps of the procedure. The vibration signal is first preprocessed, which includes data categorization, data segmentation, and feature extraction before the defect diagnosis model is trained and optimized. The extraction of the characteristic features of the vibration signal is utilized to implement the online fault diagnosis through the fault diagnostic model. The fault data for analysis is accessed from the open-source Paderborn university bearing data. The fault categories taken are of multiple faults on the inner and the outer raceways. Optimization is carried out in the ANN model with regard to its number of hidden layers, size of the layers and the activation function to seek for a set of these parameters, which result in the best performance in diagnosing the bearing defects for the fault categories under consideration. The ANN approach is proved to be a useful technique for detecting faults and improving classification accuracy, making it more suitable for rolling bearing fault diagnosis. A maximum fault classification accuracy of 98.5% was achieved with a single hidden layer of size 25 and the Tanh activation function. The corresponding training time was 3.9831 seconds, and the prediction speed was found to be 24000 observations per second.