Application of machine learning methods for automatic prediction of failures of construction and road machines

Abstract

The extension of the actual service life of the fleet of construction and road machines shifts the burden from equipment renewal to maintaining its operability, under which overhaul intervals assigned by running time cease to correspond to the condition of assemblies. The national regulatory framework – GOST 27.002-2015, GOST R 53564-2009, GOST R 27.606-2013 and GOST R ISO 13381-1-2016 – establishes the terminology, the forecasting stages and the requirements for monitoring systems, leaving the computational procedure for recognising a pre-failure condition outside the scope of regulation. The purpose of the work is to compare quantitatively families of machine learning algorithms on a single array of operational observations and to convert classification quality indicators into operational and cost values. An author's model of a fleet of 120 machines of four types has been constructed with daily registration of sixteen parameters for the period from January 2022 to May 2026; the array contains 191,760 records, 991,275 engine hours of running time and 1,389 failures. Logistic regression, random forest, gradient boosting and a multilayer perceptron were fitted on the training sample up to the end of 2024, and validation was carried out over the period from 2025 to May 2026. The area under the ROC curve ranged from 0,8890 to 0,9041, and the area under the precision-recall curve from 0.5202 to 0,5755 at a positive class share of 0,1360; the median warning time lies within the range of 41,8 to 44,2 engine hours. The largest contribution to recognition is made by the iron content in the oil (47,0%) and the root-mean-square value of vibration velocity (21,9%), whereas the total running time since the beginning of operation proved to be almost uninformative. The transition to a predictive strategy reduces the number of sudden failures by 86,9%, raises the technical availability factor from 0,9437 to 0,9744 and lowers unit costs from 590,4 to 273,3 thousand roubles per 1,000 engine hours. The difference between the linear model and the tree ensembles turned out to be smaller than the difference between the variants of feature description, which shifts the centre of gravity of the problem from the choice of algorithm to the composition and quality of the recorded diagnostic parameters.