dummy033485/manufacturing-digital-twin

Manufacturing Digital Twin — Tool-Wear Estimation & Maintenance Policy Study

A digital-twin study on the NASA PCoE Milling Data Set (UC Berkeley BEST Lab: 16 cases, 167 runs, 6 sensor channels per run, measured flank wear VB). Built around the research areas of the MOSAIC Lab (Dr. Michail Katsigiannis, Department of Mechanical Engineering, University of South Carolina): shop-floor sensing, simulation for real-time operations, digital twins, and decision support.

Verified results

Module 1 — wear estimation from single-cut sensor features (per-signal mean/std/RMS/max/kurtosis/skew + DOC/feed/material/cut number; 146 measured cuts):

• Leave-one-case-out: RF RMSE 0.192 mm · MLP 0.218 · no-sensor wear-curve baseline 0.218
• Leave-one-condition-out: RF RMSE 0.197 mm · MLP 0.292 · baseline 0.243

Wear model: empirical per-cut VB increments per condition: 0.039–0.155 mm/cut; nominal life to VB = 0.45 mm is only ~4–16 cuts (accelerated-wear laboratory data).

Modules 2+3 — SimPy digital twin + policy study (milling → inspection → assembly; 20 shifts × 30 replications; the condition-based controller sees only the ML wear estimate, with noise = the measured cross-condition RMSE 0.197 mm):

Condition	Policy	Good parts	Scrap	Unplanned stops
DOC 1.5 / feed 0.5 / cast iron	run-to-failure	1253	13.9%	148.3
	fixed interval	2231	15.6%	11.2
	condition-based	1862	3.9%	0.0
	condition-based, sensor sigma/2	1922	3.4%	0.0
DOC 0.75 / feed 0.25 / steel	run-to-failure	831	16.4%	168.8
	fixed interval	1636	13.7%	4.2
	condition-based	1524	8.5%	0.2
	condition-based, sensor sigma/2	1551	7.9%	0.0

Condition-based maintenance eliminates unplanned stops and cuts scrap by half or more, but pays with extra planned tool changes — estimator noise forces conservative replacements, and a well-tuned fixed interval can beat it on raw throughput. Halving the estimator error recovers only part of the gap (the EMA-filtered controller is already noise-tolerant).

Files

• wear_rf.pkl — final RF wear model + feature list
• twin_report.json — full measured results, twin setup, and policy definitions
• wear_model.json — per-condition wear-increment statistics and nominal tool lives
• milling_runs.csv — per-run features extracted from the raw signals

Live demo

https://huggingface.co/spaces/dummy033485/manufacturing-twin-demo

Limitations

Accelerated-wear laboratory data (absolute throughputs reflect that regime; policy rankings are the transferable result). Scrap-vs-VB is a modelled step function, stated in the code. Cycle times are stochastic assumptions; wear is the measured component.

By Zain — Mechanical Engineering + AI, Fall 2027 PhD applicant.