Variable selection for conveyor-belt mean wear rate prediction

Rubber belt conveyors are an integral part of many mining and bulk haulage applications. The belts are designed to wear in-service and thus need to be replaced periodically. Currently, belt life management is done by measuring the thickness reduction and estimating remaining life once the belt has been operating for some time. This approach is not applicable for new conveyor belts. In this paper we present a process for building a model, and results thereof, to predict the life of new conveyor-belts based on a variety of design and operating parameters. The work analyses wear readings from operating conveyor belts and constructs linear regression models for predicting the wear-rate of out-of-set conveyors. Ultrasonic readings from 165 iron-ore rubber, steel-cored conveyor belts, installed in 95 conveyors were modelled against various belt and conveyor features. Nine linear regression models and/or modelling approaches were evaluated with nested cross-validation to measure prediction error. Results show that these models achieved an improvement in performance error of up to 46% compared to relying on the mean wear rate alone and indicate that the mean wear rate of heavy-duty conveyor belts is, at least in part, related to the compound variable of belt-speed squared divided by belt-capacity (length x width). Although the models described herein offer a significant improvement to the current best practice for estimating life of new conveyors, models were only able to account for a maximum of 75% of the observed wear behaviour and prediction error still remains in the same order as the data variance, suggesting additional variables will be required to improve end-of-life prognosis. This work also demonstrates, that for this dataset, in which two explanatory variables dominate, performance error is largely unaffected by variable selection approach. Finally, the work shows how widely used data science methods can be applied to commercially impactful equipment life prediction. The work can be easily replicated by conveyor owners to improve their own belt maintenance planning.