J. Sens. Sens. Syst., 9, 143–155, 2020 https://doi.org/10.5194/jsss-9-143-2020 © Author(s) 2020. This work is distributed under the Creative Commons Attribution 4.0 License. Data-driven vibration-based bearing fault diagnosis using non-steady-state training data Kurt Pichler 1 , Ted Ooijevaar 2 , Clemens Hesch 1 , Christian Kastl 1 , and Florian Hammer 1 1 Linz Center of Mechatronics GmbH, Altenberger Straße 69, 4040 Linz, Austria 2 DecisionS, Flanders Make, Gaston Geenslaan 8, 3001 Leuven, Belgium Correspondence: Kurt Pichler (kurt.pichler@lcm.at) Received: 20 September 2019 – Revised: 31 March 2020 – Accepted: 9 April 2020 – Published: 12 May 2020 Abstract. This paper presents the extension of an empirical study in which a universally applicable fault di- agnosis method is used to analyse vibration data of bearings measured with accelerometers. The motivation for extending the previously published results was to provide a profound analysis of the proposed approach with regard to a more feasible training scenario for real applications. For a detailed assessment of the method, data were acquired on two different test beds: a gearbox test bed equipped with various bearings at different health states and an accelerated lifetime (ALT) test bed to degrade a bearing and introduce an operational fault. Features were extracted from the raw data of two different accelerometers and used to monitor the actual health state of the bearings. For that purpose, feature selection and classifier training are performed in a supervised-learning approach. The accuracy is estimated using an independent test dataset. The results of the gearbox test bed data show that the training of the method can be performed with non-steady-state data and that the same feature set can be used for different revolution speeds if a small decrease in accuracy is acceptable. The results of the ALT test bed show that the same features that were identified in the gearbox test start to change significantly when the bearing starts to degrade. Thus, it is possible to observe the identified features for applying predictive maintenance. 1 Introduction Manufacturing companies continuously try to increase their productivity, by avoiding machine downtime among other things. The former involves considerable costs because of the resulting loss of turnover. Monitoring the condition of, for instance, bearings and gears plays a vital role in the main- tenance programme of rotating machines. Early fault detec- tion could allow for moving from a time-based preventive- maintenance programme to a condition-based predictive- maintenance strategy and reducing unexpected machine downtime and cost. Vibration-based condition monitoring is an established ap- proach that has been employed by industries for many years in their maintenance programmes (Randall, 2011). However, up to this day, machine operators often still base their main- tenance decisions on data from the periodical and manual inspection of single machines, which does not always result in correct conclusions. The common practice is that vibra- tion measurements are periodically recorded using portable vibration sensors, and measurement signals are analysed by an expert to interpret the machine’s health condition. This ap- proach can, however, lead to serious misinterpretation, where rapidly growing impairments could be missed. A continuous condition-monitoring approach enables early detection of machine faults. In this way, the machine condition is continuously tracked, and total failures can be anticipated in advance, hence allowing appropriate mainte- nance actions. Despite their advantages, continuous monitor- ing programmes are still not well adopted by industry. Firstly, this is because it often involves a high investment cost. Al- though recent advancements in sensor, acquisition and pro- cessing hardware have demonstrated cost-effective solutions (Albarbar et al., 2008; Ompusunggu et al., 2018), the eco- nomic benefit of the investment is still not clear and hard to quantify. Secondly, this is because many of those systems still require an expert to interpret the analysis results. Finally, this is also because it is not straightforward to select the most Published by Copernicus Publications on behalf of the AMA Association for Sensor Technology.