Remaining Useful Life Estimation for Roller Bearings through a Takagi-Sugeno Fuzzy Logic Model with Convolutional Neural Networks Consequents

Abstract

Remaining useful life (RUL), which can be used for optimization, is the prediction of the anticipated time that a machine, system, or object will fail, breakdown, or degrade. Furthermore, by giving managers knowledge about machine health management, it can enhance the dependability of maintenance systems. While many works had tackled RUL estimation through data driven and physical model-based approaches, there has been a lack of RUL models on expert systems like fuzzy logic. While an expert system-based model allows for reconfiguration to suite a specific setup with little recorded dataset, the main challenge is defining the appropriate rules brought by the non-linearity of the RUL estimation problem. In this study, a Fuzzy Logic model with Deep Learning consequents for estimating the RUL of roller bearings with run-to-failure conditions was developed to allow manufacturing sites to develop optimized schedules for preventive maintenance activities. Input membership functions were set according to datasets provided by existing literatures, while the consequents were developed through model training on the convolutional neural networks (CNNs). Based on the PRONOSTIA test set results, the fuzzy logic RUL model's average RMSEs was 20.24%. Additionally, it was discovered that test results generally improved with bearing longevity under the tested conditions. Overall, the created RUL model can assist manufacturing plants in making wise choices for the best scheduling of its preventative maintenance.