A data label expansion strategy for tool wear prediction through data-driven approaches

In-process tool wear prediction is of great importance in intelligent machining, which can fully utilize tool life as well as reduce costs. Data-driven approaches are now playing a significant role in tool wear prediction. When data-driven approaches were utilized, wear labels should be embedded towards the purpose of training, the method of frequently measuring tool wear to obtain labels is very cumbersome. A novel strategy for wear label expansion is proposed to improve the performance of data-driven tool wear prediction models and PHM2010 datasets are utilized for evaluation. During the stable machining, tool wear progression exhibits continuity and monotonicity. Therefore, numerical interpolation methods are used to generate computed labels to fill the gaps between sparse measurements of wear. The cubic spline interpolation (CSI) and the piecewise cubic Hermite interpolation (PCHI) are compared and the mean squared errors (MSE) are evaluated. The measured-label samples, generated-label samples, and sparse-label samples were applied for tool wear prediction models training base on physics-informed neural networks (PINNs) methods, respectively. The performance of them is were compared and evaluated. The results show that the proposed strategy can reduce the difficulties in terms of the acquisition of tool wear label data via regular measurement. Therefore, this strategy can help the application of data-driven approaches in industries, as well as provides new insights for expanding the dataset of tool wear prediction models.