TY - CONF

T1 - Condition Monitoring using Domain-Adversarial Networks with Convolutional Kernel Features

AU - Caceres-Castellanos, Cesar

AU - Kortmann, Karl-Philipp

AU - Fehsenfeld, Moritz Johannes

N1 - Funding Information: The authors of the Institute of Mechatronic Systems would like to thank the company Lenze SE for their expertise and the provision of part of the test benches. The research for this paper was partly funded by the German Federal Ministry for Digital and Transport (BMDV) under the grant number 19F2132F within the mFund initiative.

PY - 2023

Y1 - 2023

N2 - The data-based condition monitoring and diagnosis of a mechatronic system can be a challenge due to the amount of labeled data traditional methods require. Moreover, transferring a trained classification model from its source domain to another mechatronic system is a difficult task due to even minor differences between sensors, dimensions, or environmental conditions. Additionally, labeled data may not be available or difficult to obtain in this new target domain. In this paper, a novel approach to time series based domain adaptation is proposed by modifying a Domain-Adversarial Neural Network. Therefore, a MiniRocket transform is combined with an artificial neural network as a composed feature extractor. This model aims to extract domain invariant features from multivariate time series data that can be used for cross-domain condition monitoring of mechatronic systems. The model is tested for belt tension monitoring using data from two belt drives considering two types of excitation. Experimental results for wideband excitation show that the proposed model estimates the tension of the belt with high accuracy in the target domain (unsupervised). For the jerk-limited excitation, accuracy is improved for the target domain in a semi-supervised setting.

AB - The data-based condition monitoring and diagnosis of a mechatronic system can be a challenge due to the amount of labeled data traditional methods require. Moreover, transferring a trained classification model from its source domain to another mechatronic system is a difficult task due to even minor differences between sensors, dimensions, or environmental conditions. Additionally, labeled data may not be available or difficult to obtain in this new target domain. In this paper, a novel approach to time series based domain adaptation is proposed by modifying a Domain-Adversarial Neural Network. Therefore, a MiniRocket transform is combined with an artificial neural network as a composed feature extractor. This model aims to extract domain invariant features from multivariate time series data that can be used for cross-domain condition monitoring of mechatronic systems. The model is tested for belt tension monitoring using data from two belt drives considering two types of excitation. Experimental results for wideband excitation show that the proposed model estimates the tension of the belt with high accuracy in the target domain (unsupervised). For the jerk-limited excitation, accuracy is improved for the target domain in a semi-supervised setting.

KW - Fault detection and diagnosis

KW - Time series modeling

KW - Machine learning

KW - Domain adaptation

KW - Unsupervised learning

KW - Semi-supervised learning

UR - http://www.scopus.com/inward/record.url?scp=85184961317&partnerID=8YFLogxK

U2 - 10.1016/j.ifacol.2023.10.1180

DO - 10.1016/j.ifacol.2023.10.1180

M3 - Paper

SP - 7746

EP - 7752

T2 - IFAC World Congress 2023

Y2 - 9 July 2023 through 14 July 2023

ER -