Hybrid digital twin with Bayesian synchronisation for uncertainty-aware tool-wear prognostics and predictive tool replacement in milling
Year: 2026
Type: preprint
Abstract: Tool wear in milling is usually monitored with regressors that map sensor features to flank wear. Such models rarely follow the individual tool, seldom report calibrated uncertainty and are almost never judged by the replacement decisions they support. This study presents a hybrid digital twin of a face-milling tool. Its virtual entity couples a three-stage flank-wear law, a population prior that conditions the wear parameters on feed, depth of cut and work material, and a probabilistic virtual sensor whose error is split into a persistent per-tool bias and heavy-tailed noise. A particle filter with kernel smoothing updates the wear state, tool-specific parameters and sensor bias after every cut and accepts occasional wear inspections. The predicted remaining-useful-life distribution drives cost-based replacement rules and a what-if advisory on feed and depth of cut. On the NASA Ames milling data (15 tools), under nested leave-one-tool-out validation, the twin estimated flank wear with a root-mean-square error of 63 µm, against 74 µm for the virtual sensor alone and at least 104 µm for four machine-learning regressors, with 90% of the measurements inside its nominal 95% intervals. Inspecting every third measured run cut the error on the other runs to 46 µm. The median relative error of the predicted remaining life was 24%. Replacing tools when the expected remaining life fell below a tuned multiple of the next run length lowered the cost rate by 31% relative to a condition-based schedule, for a failure costing five replacements, and avoided all failures.
Source: Research Square
Author: Rabah BOUDIAF
Institution: Badji Mokhtar-Annaba University
Language: English
Subfield: Mechanical Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235266323
