| Abstract [eng] |
Predictive maintenance of ball bearings is considered critical for the long-term and efficient operation of machinery and equipment. Due to their high performance and reliability requirements, these components are widely used in a wide range of industries including automotive, aviation and shipbuilding. Bearing diagnostics and condition monitoring in Lithuania is still a new and under-researched area compared to the global scale, and this research contributes to an important scientific and industrial development. Mathematical models for predictive maintenance, which enable optimisation of maintenance processes, reduction of unexpected failures and prolongation of equipment lifetime, have been investigated in this work. The main objective of this study was to determine the effectiveness of the application of machine learning methods for bearing defect detection, defect location and defect size using real data. Random forest, LightGBM and artificial neural network algorithms were used in the study. The results showed that these models were ideal for fault identification with 100% accuracy. The LightGBM model gave the best results in identifying defect locations, with an accuracy of 96.39% and a macro-average area under the curve of 0.9951. As for the defect sizes, the LightGBM model recorded the lowest errors and a coefficient of determination of 0.9153. The cost-benefit analysis also showed that the proposed predictive maintenance solution is financially viable per train car. The predictive maintenance system is estimated to pay for itself in 4.95 years. |
