Machine Learning Technics for Remaining useful Life Prediction using Diagnosis Data: a Case Study of a Jaw Crusher

Abstract

Predictive maintenance currently involves digital transformation with all the technologies developed to serve the latter. This maintenance strategy is believed to be an efficient solution to end late/early intervention issues. It is for this reason that machine health state monitoring by Remaining Useful Life prognosis is very crucial. However, in the literature, most studies focus on failure diagnosis more than the system's Remaining Useful Life. In addition, to prepare models to serve the prognosis, the use of actual machinery data is critical to assure the later scalability of the application. The literature about predictive maintenance has often evaluated data-driven approaches with machine learning techniques processing simulated Data rather than real ones. To tackle this problem, the authors propose a continuity of previous work treating a jaw crusher default diagnosis in the context of the ore mining industry. The RUL of the crusher components is estimated upon completion of the fault diagnosis data. Smart sensors Data have been preprocessed to serve the evaluation of four regression machine learning models: Bayesian Linear Regression, Poisson Regression, Neural Network Regression, and Random Forest. Poisson regression and Neural Network required data normalization in this case study to improve their performance. Linear regression methods proved their inability to forecast the machine degradation state, while the bagging ensemble method, Random Forest, was able to track the actual values. This paper aims to enhance the Prognosis and Health Management of the machine while contributing to the literature enhancement on failure prognosis using real industrial data. Keywords— Data-driven approach, Industry 4.0, Machine learning, Predictive Maintenance, RUL prognosis, Smart sensors.