Heterogeneous vibration data preprocessing method for fault detection

• Published in: Procedia computer science Vol. 253; pp. 2127 - 2136
• Main Authors: Claeyssens, Donatien, Zekri, Dorsaf, Thierry, Delot, El Cadi Abdessamad, Ait
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 2025
• Subjects: deep learning; factory based data set; Fault detection; heterogeneous data; vibration data set; deep learning; Fault detection; factory based data set; vibration data set; heterogeneous data
• ISSN: 1877-0509; 1877-0509
• DOI: 10.1016/j.procs.2025.01.273

The early detection of bearing faults is an important subject for maintenance and many research works have been conducted on laboratory-based homogeneous datasets. Industrial data sets are still uncommon, and so far, only one have been made publicly available so work on such industrial-based data sets are still rare. In this study, we focus on the use of an industrial-based data set and how well a model can classify new and unused data compared to a laboratory-based homogeneous data set. To do so, we propose a novel preprocessing method for heterogeneous vibration data set in order to extract as much information as possible for the model training. This preprocessing method is based on a sliding window method with a variable window length based on several factors of the data set such as the rotational speed of the machine, the duration of the measurement as well as the sampling frequency. Our preprocessing method ensures standardization of the dimensions, minimal loss of information and maximizes the quantity of data available. Because of the highly unbalanced nature of such heterogeneous data sets, we use spectrograms as the data representation and the CNN as a learning model. A validation method and experimental results are proposed to show the superiority of a model trained on heterogeneous data set over homogeneous data set.