Measurement-Driven Low-Label Adaptation for Signal-Based Areal Surface Roughness Prediction After Tool Replacement in Ti–6Al–4V Milling

Abstract

In precision milling, areal surface roughness (Sa\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\textrm{a}}$$\end{document}) is commonly verified through offline surface topography measurement. However, tool replacement can change process signals and reduce the reliability of signal-based Sa\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\textrm{a}}$$\end{document} prediction. This study formulates Sa\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\textrm{a}}$$\end{document} prediction in Ti–6Al–4V milling as a measurement-driven low-label target-tool adaptation problem. Three tools were machined under fixed cutting parameters, and each steady-state pass was paired with one confocal white-light Sa\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\textrm{a}}$$\end{document} measurement. Force and vibration signals were organized using a one-pass–one-sample strategy to avoid overlapping window sample inflation. Seventy-two statistical descriptors were extracted to describe amplitude, energy, fluctuation, distribution, and spectral response characteristics. Descriptor-based regressors and a hierarchical signal statistical fusion model were evaluated using the same leave-one-tool-out (LOTO) protocol. Unlabeled target-tool descriptors were used for per-tool input normalization, and limited target-tool Sa\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\textrm{a}}$$\end{document} labels were used only for affine output calibration. Raw cross-tool prediction remained unreliable for all models, with negative average R2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values. After calibration, random forest achieved the best numerically calibrated performance in the present controlled setting, with average RMSE/R2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$R^2$$\end{document} values of 0.0925 µm/0.4658, 0.0956 µm/0.3331, and 0.0973 µm/0.2908 under pass-uniform calibration, layer-all-pass calibration, and layer-one-pass calibration, respectively. Under the present controlled three-tool LOTO setting, the results indicate that reliable statistical descriptors, unsupervised input alignment, and few-shot affine correction provide a practical route for improving Sa\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\textrm{a}}$$\end{document} prediction after tool replacement. Tool replacement causes severe Sa\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\textrm{a}}$$\end{document} prediction bias despite fixed cutting parameters. One-pass samples link steady-state signals with measured Sa\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\textrm{a}}$$\end{document} values. Few-shot affine calibration improves cross-tool Sa\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\textrm{a}}$$\end{document} prediction with limited labels. Tool replacement causes severe Sa\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\textrm{a}}$$\end{document} prediction bias despite fixed cutting parameters. One-pass samples link steady-state signals with measured Sa\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\textrm{a}}$$\end{document} values. Few-shot affine calibration improves cross-tool Sa\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$S_{\textrm{a}}$$\end{document} prediction with limited labels.