Impact of feature engineering and domain adaptation on tool wear prediction accuracy under variable cutting conditions
Journal
Manufacturing Letters
Journal Volume
44
Start Page
1232
End Page
1241
ISSN
22138463
Date Issued
2025-08
Author(s)
Abstract
Traditional models for tool wear monitoring are typically built using fixed cutting conditions, like speed and feed rates, but their accuracy declines under variable conditions, such as changes in material hardness. To address this, the current study adopts domain adaptation techniques using neural network models to enhance tool wear monitoring across different material hardness levels. The study utilizes vibration signals from accelerometers and spindle current signals to predict tool wear. By employing Domain Adversarial Neural Networks (DANN), the domain adaptation model effectively reduces discrepancies between data distributions from different domains, improving prediction accuracy. Prior to model training, frequency-domain signal features were reordered and normalized to enhance predictive accuracy. The results demonstrate that both feature engineering and domain adaptation substantially improve prediction performance. Specifically, the root mean square error (RMSE) of predictions was reduced from 62.913 μm (without preprocessing) to a range of 21.713–26.854 μm. Feature reordering and normalization alone contributed to reducing the RMSE to 31.596 μm. It is observed that domain adaptation and feature engineering have similar impacts on model generalizability. For datasets that have undergone effective preprocessing and feature extraction, domain adaptation offers only slight improvements in tool wear prediction under varying material hardness conditions. The findings also indicate that selecting source domain data similar to the target domain is crucial for optimal model performance.
Publisher
Elsevier Ltd
Type
journal article