M2-2.4 - Enable Transferability of interpretable ML Models for Bearing Diagnostic in Changing Operating Conditions

Event

23. ITG/GMA-Fachtagung Sensoren und Messsysteme 2026
2026-06-09 - 2026-06-10
Nürnberg

Band

Vorträge

Chapter

Modellbildung und Signalverarbeitung

Author(s)

J. Schauer, H. El Moutaouakil - Centre for Mechatronics and Automations Technology, Saarbrücken, A. Schütze - Saarland University, Saarbrücken

Pages

420 - 425

DOI

10.5162/sensoren2026/M2-2.4

ISBN

978-3-910600-11-9

Price

free

Abstract

In industrial machine learning applications, changing environmental and operating conditions significantly reduce the robustness and predictive quality of machine learning models, leading to performance drops under domain shifts. Interpretable machine learning algorithms based on physically meaningful features allow us to build models that are more robust to data distribution shifts. However, for significant domain shifts, the model should be able to adapt to the novel domain to avoid a performance decrease. Especially in industrial settings, the influences of external sources are hard to control. This study utilizes a novel approach to represent interpretable machine learning algorithms as deep neural networks, enabling transfer learning to adapt to novel domains. In this study, an interpretable machine learning algorithm comprising feature extraction, feature selection, and classification was developed to predict the condition of a roller bearing using accelerometer data. The novel approach demonstrates the ability to adapt the model via transfer learning to a new, unseen rotational speed with only 32 new transfer samples. While keeping information from the source domain out, the novel approach increases the model's accuracy from 77.6% to 10⁰% in the unseen operating condition.