Faster, better, safer – new machine learning methods revolutionize complex systems

A Machine Learning Approach for automative fault nowcasting

By John Pavlopoulos, assistant professor at the Department of Informatics of Athens University of Economics and Business (elected) and affiliated with Stockholm University.

Fault diagnosis is the task of detecting the fault that caused a problem or unexpected behaviour to a subject. If the subject is a human being and the nature of the problem is medical, a reasonable diagnostics process comprises the physician who reads or listens to the patient’s symptoms, looking at any radiographs or echocardiograms, studying the medical records, and then concluding regarding the root cause.

Human-in-the-loop-architecture of automotive fault nowcasting

Our study, conducted by Tony Lindgren, John Pavlopoulos and Korbinian Randl from Stockholm University,  presented a large-scale analysis, demonstrating the effectiveness of Natural Language Processing (NLP) in addressing multilingual fault diagnosis. Different from medical fault diagnosis, which is most often based on image input, we investigated a case from the automotive industry, using data from Scania. Specifically, we classified textual descriptions of problems, as these were registered through work orders in workshops, regarding the actual root cause for the vehicle malfunction.

The large-scale multi-class and -linguality nature of automated fault report management in the automotive domain, combined with the terminology, constitute a specific task and a challenging problem from an NLP perspective. We addressed this challenge by fine-tuning a pre-trained multilingual Transformer, to produce helpful rankings of probable causes of these faults. The experimental results were analysed also from a macroscopic perspective considering both dimensions, languages and classes, showing the feasibility of a novel human-in-the-loop workflow, which surpassed current industry standards for fault detection in the automotive industry.

Relation with the EFRA project

Although the focus of our study was within the automotive domain, the applicability of the proposed method goes beyond, concerning any troubleshooting point where fault claims are received in textual form. A direction for future work is the application to other areas, where results are of high value for all actors with a need for the diagnosis of complex systems, especially when the actor is global. Such an area is the agricultural domain, where AI-enabled food risk prevention is on its way.

Send us a message

Get our latest news

to our newsletter.