Prognostics applications in the automotive industry are growing rapidly and customers have begun to expect this capability. Remaining useful life (RUL) models are an important aspect of a prognostic as they affect both how far in advance and with what confidence failures can be predicted. Model selection and design must include technical considerations such as mathematical complexity and training data availability, as well as business considerations such as implementation plans, constraints, and risks of inaccurate predictions.
This paper compares different RUL models that have been developed for turbo actuators on diesel engines, with the business objective of advising bus fleet customers on preventive maintenance intervals. The design, development, validation, and resulting prediction accuracy of each RUL model is detailed. A selection process is then applied to choose the model best suited to the intended purpose. In doing so, the paper sheds light on strengths and weaknesses of deep learning RUL models over statistical RUL models. The paper also focuses on the state-of-the-art deep learning network “Tabnet” and its results for useful life predictions. Among the different methods, Accelerated Weibull Failure Time model provides better predictions with a concordance of 0.94 and ~15% less error than any other model.
Automotive industry is focused on monitoring health and performance of vehicles to help customers to improve uptime and reduce the downtime with planned maintenance. To achieve this, technologies like telematics are used for continuous data flow. The data depicts the details of functioning of various components in the engine and subsystems; and the fault codes - the diagnostic troubleshooting codes, associated with them. Traditional method by service representatives to address these fault codes- is to follow the recommended troubleshooting trees. With the electronic engine and subsystems performance interdependencies, it gets challenging to address the same in the traditional manner. Also, the grouping of the fault codes is not known if the fault codes that occurred are related to a specific cause. We implemented unsupervised machine learning and data mining techniques to address such issues. First, the co-occurrence theory that helps in understanding the fault codes that occur together and exhibit dependencies and relations amongst them. We implemented market basket analysis to study the fault codes co-occurrences. Second, we implemented clustering algorithms to know groups/categories of fault codes based on their functional states. These studies provide insights on the failures, component states, and help in troubleshooting the problems experienced by the engine and subsystems. Also, these methods help to address the issues in the early stage, in turn helping technicians to identify the issue and improving the uptime (early repairs and diagnostics). Further, this paper presents the results of the experiments aligning to the domain needs.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.