A Deep Learning Approach for Data-Driven Predictive Maintenance of Rolling Bearings

It is crucial for industrial companies that their systems are available and healthy as most as possible. However, it is inevitable that machines will degrade over time, leading to a fault, or even a complete breakdown if the fault is not identified and addressed in time. In this context, predictive maintenance, i.e., maintenance scheduled and implemented accordingly to the machine’s estimated condition and degradation, is considered a promising approach, as it can extend machines’ availability, productivity, overall product quality, and reduce the waste of material and human resources related to maintenance, among other benefits. It is, for this reason, a key object of Circular Manufacturing, which is an emerging discipline that aims at creating more clean and sustainable manufacturing environments. Deep Learning in this area has been increasingly researched, showing promising results and the ability to extract hidden and abstract information that can improve the performance of health status prediction. In this work, a predictive maintenance approach using Deep Learning is developed for the PRONOSTIA-FEMTO benchmark, regarding the prediction of the current health status of rolling bearing components. The dataset contains vibration data from several run-to-failure experiments. The preprocessing stage is carried out using local mean decomposition, enabling better feature extraction. The approach is then compared to another non-Deep Learning approach for performance assessment.