Architecture of a fault-tolerant analytical platform for predicting failures in the fleet of road equipment based on hybrid machine learning models

Abstract

The article presents a modular architecture of a fault-tolerant analytical platform for predicting failures in the fleet of road equipment, developed taking into account the specifics of the operation of rollers, pavers and crawler excavators. The architecture integrates classical statistical methods and neural network models into a single hybrid system capable of processing non-stationary, incomplete and noisy data from onboard sensors. Special attention is paid to reducing the number of false alarms due to the mechanism of dynamic calibration of thresholds and estimation of confidence intervals. The experimental verification was carried out on real data collected during the operation of 14 units of equipment for 8 months. The results showed that the proposed platform provides an F1 measure of at least 0.93 with a skip rate of up to 20% and reduces the number of false alarms by 34% compared to traditional solutions based on isolated machine learning models. The developed architecture is compatible with on-board computing modules and can be integrated into existing fleet management systems.