Motor Bearing Fault Analysis Using Machine Learning Techniques

Abstract: Motors are the driving force of our industrial world, as they power approximately 85% of all rotating machines. This revolutionary invention has been through radical changes before entering into the commercial industries, and their present form are very reliable. However, despite being so robust, induction motors are not entirely fault proof and are more vulnerable to internal faults than external ones. Certain types of bearing faults are more frequent among the internal faults, and their effects range from various performance-related issues to frequent motor breakdowns. Fortunately, the recent advancement in Digital Signal Processing and Machine Learning relatively allow us to detect these bearing faults and figure out their origins, which in turn enables us to preserve their health and take measures against breakdowns. Through vibration analysis, this paper proposes a powerful method to detect these faults and differentiate among them based on the location of their occurrence within the bearing. We applied it to the rolling bearing fault data provided by the Case Western Reserve University (CWRU) Lab to demonstrate the method's cogency. We employed several supervised learning algorithms to classify the bearing fault signals, namely Decision Tree (DT), Random Forest (RF), and Gradient Boosting (GB). Prior to classification, we proposed time-domain signals using various frequency domain-transformation techniques, such as Discrete Cosine Transform (DCT), Fourier Transform, and Cepstrum. The obtained results show that the methods can identify bearing faults with 99.40%, 98.90% and 95.98% accuracy respectively. The obtained F1-scores are 99.40%, 100%, and 99.8% respectively.

Result Published:  2021

• Supervision