Performance of Cox models on predictive maintenance of construction machines

Abstract

Manufacturers of construction machines often provide tools to contractors to monitor through telecommunication devices the state of machines. Besides telematics data, repair information are sometimes collected by mechanics with the intention of improving the mainenance strategy. These two sources of data allow the construction of statistical models capable of condition-based maintenance. Nevertheless, either human effort need to be systematized or a cross-disciplinary effort must emerge to train statistical methods for use in production. Concretely, the reconstruction of the failure event from the repair, i.e. to understand the relationship between the modelled repair intervention and the failure event. Additionally, the model yielded poor discrimination and accuracy for production. This study demonstrates the process of building, testing and validating the Cox proportional hazards predictive model in the context of construction machines. It presents the basic framework and validation process which potential production models must abide. Moreover, insights on what can be done to improve data extraction.