Manufacturing systems consist of machinery composed of various types of mechanical components, the reliability of which is crucial to achieving the overall production goals. Failure of critical components leads to shutting down a machine, which might lead to shutting down the entire manufacturing system, resulting in reduced availability and financial losses. This paper presents a machine learning-based probabilistic degradation prediction method. An artificial neural network (ANN) has been constructed to model the multi-cycle degradation path of a mechanical component when a preventive replacement strategy is applied. The ANN is trained by the neuron-by-neuron (NBN) training algorithm and by the maximum likelihood estimation (MLE) method that estimates the distribution parameters of the degradation distribution exploiting the Broyden-Fletcher-Goldfarb-Shanno (BFGS) second-order optimization algorithm. The expected degradation values are calculated from the predicted probability parameters, and then, in a longer time horizon, the expected failure time can also be predicted. In this way, the remaining useful life (RUL) and the reliability of the component can also be estimated. This new alternative method provides a machine learning-based solution that models the complete degradation path and its deviation over against classical statistics-based solutions that model the time of the failure occurrence and its deviation.

Title

Procedia CIRP

Publisher

Elsevier

Publisher Country

Netherlands

Indexing

Scopus

Impact Factor

None

Publication Type

Online only

Volume

130

Year

2024

Pages

331–335