Title: Performance improvement techniques for neural networks in tool condition monitoring

Authors: Arup Dey; Nita Yodo

Addresses: Department of Industrial and Manufacturing Engineering, North Dakota State University, Fargo, ND 58102, USA ' Department of Industrial and Manufacturing Engineering, North Dakota State University, Fargo, ND 58102, USA

Abstract: The performance of data-driven algorithms in tool condition monitoring often depends on the combinations of different factors such as data quality, input dimensions, and model architecture. Several performance improvement techniques such as data denoising, feature selection, and regularisation techniques are known to enhance prediction accuracy. Moreover, selecting model architecture and tuning hyperparameters also significantly impact the prediction performance. Although the prediction accuracy of a data-driven method can be improved using these techniques, their importance is rarely discussed for tool condition monitoring. In this paper, the importance of various performance improvement techniques is extensively analysed by applying them to a CNC milling machine dataset for tool wear prediction. The investigation results and performance measurement metrics showed data denoising techniques, feature reduction techniques, and regularisation methods improved prediction accuracy up to around 55%. The selection of techniques for improving the accuracy depends on the nature of a dataset and applied algorithms.

Keywords: tool wear; condition monitoring; neural networks; model performance; denoising; regularisation; feature extraction; feature selection.

DOI: 10.1504/IJMMS.2022.124921

International Journal of Mechatronics and Manufacturing Systems, 2022 Vol.15 No.2/3, pp.107 - 126

Received: 14 Oct 2021
Accepted: 14 Feb 2022
Published online: 16 Aug 2022 *