Large fleets of industrial and construction vehicles require periodic maintenance activities. Scheduling these operations is potentially challenging because the optimal timeline depends on the vehicle characteristics and usage. This paper studies a real industrial case study, where a company providing telematics services supports fleet managers in scheduling maintenance operations of about 2000 construction vehicles of various types. The heterogeneity of the fleet and the availability of historical data fosters the use of data-driven solutions based on machine learning techniques. The paper addresses the learning of per-vehicle predictors aimed at forecasting the next-day utilisation level and the remaining time until the next maintenance. We explore the performance of both linear and non-liner models, showing that machine learning models are able to capture the underlying trends describing non-stationary vehicle usage patterns. We also explicitly consider the lack of data for vehicles that have been recently added to the fleet. Results show that the availability of even a limited portion of past utilisation levels enables the identification of vehicles with similar usage trends and the opportunistic reuse of their historical data.

大量的工业和建筑车辆需要定期维护。安排这些操作可能具有挑战性，因为最佳时间表取决于车辆的特性和使用情况。本文研究了一个实际的工业案例，其中一家提供远程信息处理服务的公司支持车队经理安排约2000辆各种类型的施工车辆的维护作业。机群的异质性和历史数据的可用性促进了基于机器学习技术的数据驱动解决方案的使用。本文介绍了针对每辆车的预测变量的学习，旨在预测第二天的使用水平以及下一次维护之前的剩余时间。我们探索线性模型和非线性模型的性能，表明机器学习模型能够捕获描述非固定车辆使用模式的潜在趋势。我们还明确考虑到缺少最近添加到车队中的车辆的数据。结果表明，即使只有有限的过去利用率水平的可用性，也可以识别具有相似使用趋势的车辆，并有机会重用其历史数据。