Go back to publications

Detecting Faulty Air Pressure System Sensor Data in Scania Trucks

4 July 2019 L. Dupont, M. Golombeck, S. Tübben Technical Report 10.13140/RG.2.2.15888.56320 University of Applied Sciences and Arts Dortmund German Supervised by Prof. Dr. Christoph Friedrich
#Machine Learning#Anomaly Detection#Predictive Maintenance#KNIME
PDFResearchGate

The air pressure system of a truck regulates safety-critical components including the brakes and the transmission. It is also built from numerous fault-inducing parts, which makes diagnosis complicated: for self-diagnosis and automatic intervention to work at all, an APS fault has to be detected reliably and distinguished from a fault elsewhere in the vehicle.

This project tackles that classification problem on the “APS Failure at Scania Trucks” dataset published by Scania CV AB for the Kaggle competition of the same name. It was completed as a DiffPro-ML mini-project within the machine learning module of the Wirtschaftsinformatik master’s programme, and the work later earned a BMBF research grant (01S17075).

Poster

The project was also presented as a poster, which walks the whole CRISP-DM cycle in one view: business and data understanding on the left, the six modelling workflows and the comparison against the Kaggle leaderboard on the right.

Poster titled Erkennung fehlerhafter Air Pressure System Sensordaten von Scania Lastkraftwagen, by Lars Dupont, Marius Golombeck and Stefan Tuebben. Panels cover the introduction, the CRISP-DM cycle, business understanding, data understanding and data preparation, then modelling, an evaluation carrying a ROC curve and a table comparing the six models against the competition scores, and a closing conclusion.
Conference poster, in German. Open the full version (PDF, 390 KB).

Modelling

Modelling was carried out in the data science tool KNIME, with optimised workflows built for six models:

  • Decision Tree
  • Random Forest
  • Isolation Forest
  • Probabilistic Neural Network
  • Multilayer Perceptron
  • Naive Bayes

Evaluation

Because the dataset comes from a public competition, the results can be compared directly against the Kaggle leaderboard. In that context the Isolation Forest implementation gave the most promising result, achieving the second-highest score in the associated competition.

The trade-off is worth stating precisely: the Isolation Forest solution produces considerably fewer type 1 errors than the alternatives, but performs worse on type 2 errors. For a system whose purpose is deciding whether to intervene in a truck’s braking behaviour, which of those errors matters more is a domain question rather than a statistical one.

Cite this work

@techreport{dupont2019scania,
  author      = {Dupont, Lars and Golombeck, Marius and Tübben, Stefan},
  title       = {{Erkennung fehlerhafter Air Pressure System Sensordaten von
                 Scania Lastkraftwagen}},
  institution = {University of Applied Sciences and Arts Dortmund},
  address     = {Dortmund, Germany},
  year        = {2019},
  month       = jul,
  language    = {german},
  note        = {BMBF research grant 01S17075},
  doi         = {10.13140/RG.2.2.15888.56320},
  url         = {https://www.researchgate.net/publication/334223963}
}