Application of data-driven models to predictive maintenance: Bearing wear prediction at TATA steel

Highlights

• We show the potential of using incomplete sensor data to improving predictive maintenance.

• We present an industry application comparing data-driven methods for predictive maintenance.

• We compare the performance of Neural Networks, Partial least squared regression and Random Forest.

• We find that Partial least squared regression predicts the bush wear with a good accuracy.

• We demonstrate techniques for data cleaning in real world large and complex data sets.

Abstract

Industries that are in transition to Industry 4.0 often face challenges in applying data-driven methods to improve performance. While ample methods are available in literature, knowledge on how to select and apply them is scarce. This study aims to address this gap reported on the design and implementation of data-driven models for predictive maintenance at TATA Steel, Shotton. The objective of the project is to predict the wearing behaviour of the components in the steel production line for maintenance activity decision support. To achieve the predictive maintenance goal, the approach applied can be summarized as follows: 1. business understanding and data collection, 2. literature review, 3. data preparation and exploration, 4. modelling and result analysis and 5. conclusion and recommendation. The data-driven methods that were analysed and compared are: Partial Least Squares Regression (PLSR), Artificial Neu- ral Network (ANN) and Random Forest(RF). After cleaning and analysing the production line data, predictive maintenance with the current available data in TATA Steel, Shotton is best feasible with PLSR. The study further concludes that, predictive maintenance is likely to be feasible in similar industries that are in transition to industry 4.0 and have growing volumes of production data with varying quality and detail. However, as illustrated in this case study, careful understanding of the industrial process, thorough modeling and cleaning of the data as well as careful method selection and tuning are required. Moreover, the resulting model needs to be packaged in a user friendly way to find its way to the job floor.