Convolutional Auto-Encoder Based Degradation Point Forecasting for Bearing Data Set


International Conference on Artificial Intelligence and Applied Mathematics in Engineering (ICAIAME 2019), Antalya, Turkey, 21 April 2019, vol.43, pp.817-829 identifier

  • Publication Type: Conference Paper / Full Text
  • Volume: 43
  • Doi Number: 10.1007/978-3-030-36178-5_71
  • City: Antalya
  • Country: Turkey
  • Page Numbers: pp.817-829
  • Keywords: Bearing dataset, Convolutional neural network, Prognostics and Health Management, Degradation forecasting, Auto-encoder, ROTATING MACHINERY, DEEP, DIAGNOSIS
  • Eskisehir Osmangazi University Affiliated: Yes


In smart manufacturing industry, health analysis and forecasting of degradation starting point has become an increasingly crucial research area. Prognostics-aware systems for health analysis aim to integrate health information and knowledge about the future operating conditions into the process of selecting subsequent actions for the system. Developments in smart manufacturing as well as deep learning-based prognostics provide new opportunities for health analysis and degradation starting point forecasting. Rotating machines have many critical components like spinning, drilling, rotating, etc. and they need to be forecasted for failure or degradation starting times. Moreover, bearings are the most important sub-components of rotating machines. In this study, a convolutional neural network is used for forecasting of degradation starting point of bearings by experimenting with Nasa Bearing Dataset. Although convolutional neural networks (CNNs) are utilized widely for 2D images, 1-dimensional convolutional filters may also be embedded in processing temporal data, such as time-series. In this work, we developed one such autoencoder network which consists of stacked convolutional layers as a contribution to the community. Besides, in evaluation of test results, L-10 bearing life criteria is used for threshold of degradation starting point. Tests are conducted for all bearings and results are shown in different figures. In the test results, proposed method is found to be effective in forecasting bearing degradation starting points.