Performance Evaluation of Machine Learning Techniques for Fault Diagnosis in Vehicle Fleet Tracking Modules. (14th May 2021)

• Record Type: Journal Article
• Authors: Sepulvene, Luis; Drummond, Isabela; Kuehne, Bruno; Frinhani, Rafael; Leite Filho, Dionisio; Peixoto, Maycon; Reiff-Marganiec, Stephan; Batista, Bruno
• Is Part Of: Computer journal. Volume 65:Number 8(2022)
• Page Start: 2073
• Page End: 2086
• Publication Date: 2021-05-14
• Subjects: machine learning -- fault diagnosis -- feature extraction -- convolutional neural networks; Computers -- Periodicals; 005.1
• DOI: 10.1093/comjnl/bxab047
• Languages: English
• ISSNs: 0010-4620

Abstract

Abstract: With industry 4.0, data-based approaches are in vogue. However, extracting the essential features is not a trivial task and greatly influences the final result. There is also a need for specialized system knowledge to monitor the environment and diagnose faults. In this context, the diagnosis of faults is significant, for example, in a vehicle fleet monitoring system, since it is possible to diagnose faults even before the customer is aware of the fault, minimizing the maintenance costs of the modules. In this paper, several models using machine learning (ML) techniques were applied and analyzed during the fault diagnosis process in vehicle fleet tracking modules. Two approaches were proposed: 'With Knowledge' and 'Without Knowledge', to explore the dataset using ML techniques to generate classifiers that can assist in the fault diagnosis process. The approach 'With Knowledge' performs the feature extraction manually, using the ML techniques: random forest, naive Bayes, support vector machine and Multi Layer Perceptron; on the other hand, the approach 'Without Knowledge' performs an automatic feature extraction, through a convolutional neural network. The results showed that the proposed approaches are promising. The best models with manual feature extraction obtained a precision of 99.76% and 99.68% for detection and detection and isolation of faults, respectively, in the provided dataset. The best models performing an automatic feature extraction obtained, respectively, 88.43% and 54.98% for detection and detection and isolation of failures.