In recent years, there have been many attentions in developing technologies with the aim of monitoring and predicting emerging issues such as break downs, component failures, and quality degradations e.g., R, Prytz et al. (2015), as a means to provide predictive maintenance solution in modern vehicle industries. These existing technologies exploit several fault predictions and diagnostic pipelines ranging from statistics methods to machine learning algorithms e.g., M, You et al. (2010), Y, Lei et al. (2016). However, these solutions have not particularly concentrated on the ability to predict the component failures and the cause of the failures taking into consideration vehicle usage patterns and history of failures over time in different seasons.

This is not a trivial task since modern vehicles with their huge functionalities and dependency among their components bring out a challenge to the manufacturer to plan their maintenance strategy in this complex domain. This is truly a complex challenge since failures can be sourced and affected by multiple features, which are highly related to each other and change over time in different contexts (e.g., location, time, season).  

Under such conditions, an advanced early prediction capability is desired, because manufacturers can exceedingly serve from early prediction of potential vehicle component failures, and more specifically the chain of the features and their dependencies which may lead to a failure over time in different seasons.  This is considered important due to the fact that different seasons may have a potential effect on certain component failures, so predicting these dependencies and the actual failure enables a higher level of maintenance for optimally planning and managing total cost and more importantly safety. 

In this study, we build a probabilistic prediction model in a time series, on top of vehicle usage pattern, which is represented by the Live Vehicle Data (LVD). LVD logged and captured using multiple sensors located in Volvo vehicles that includes usage and specification of the vehicles aggregated in a cumulative fashion. We exploit and apply a type of supervised machine learning algorithm called Bayesian Network N, Friedman. (1997), on the engineered LVD (we applied a type of data engineering process to extract hidden patterns from LVD), which is logged through different seasons. These result a very complex network of dependency in each time stamp that indicates how a failure sourced by different features and their quantitative influences. In addition, integrating all these networks reveal how the usage can influence failure over time. Moreover, the quantitative influences allow us to extract the main chain of effect on a failure. This is strongly beneficial for the manufacturers and maintenance strategy to find out the main reason of failures, which can be extracted by vehicle usage pattern during their operation. © ESREL2020-PSAM15 Organizers