About this website

Every milling operation depends on a judgment that nobody can observe directly: whether the cutter currently in the spindle is still fit for the next part. Replacing it too late lets the flank wear land grow far enough to push the machined surface out of tolerance, and a worn edge can chip or fracture, taking the workpiece and sometimes the spindle with it. Replacing it too early discards cutting life that has already been paid for. Shops resolve that asymmetry with a fixed change interval taken from published tool-life figures and experience, changing early because the early error is the cheaper one. Condition monitoring replaces that interval with a measurement, and the flank-wear land width V_B is the quantity such a system estimates.

Sensing is no longer the obstacle. Drive currents, position deviations and spindle torque are values the CNC control already computes to close its own loops, and they can be read without installing anything on the machine. The limit is elsewhere: fit a monitor to one machine, one controller generation or one batch of tools, and it rarely carries over to the next combination without being recalibrated first. Machine rigidity, firmware version, tool supplier, workpiece material and the parameter envelope of the part program all leave traces, so a model fitted at one site learns the site as much as it learns the wear.

Widening the training data is the obvious remedy, and no single operator can supply that width. The traces that would supply it sit with organizations that have good reason not to hand them over. A high-frequency controller trace is not a neutral table of numbers. It exposes tool paths, cutting parameters, material choices and cycle times, which firms treat as production know-how rather than as research data. The statistical argument for pooling and the commercial argument against it bear on the same data and point in opposite directions.

Federated learning is the established response. Clients train on their own data and exchange only model updates, so the raw traces never leave the site. Sovereign data ecosystems such as Gaia-X and Pontus-X add what federation alone leaves open: who received what, under which terms, and with what recourse.

This website belongs to the Master's thesis “Development and demonstration of a sovereign federated learning concept in tool condition monitoring” by Robin Lucas Uhl, written at the Institute for Production Management, Technology and Machine Tools (PTW) of the Technical University of Darmstadt. It makes that work inspectable: three machine models trained on a public milling dataset, the rules that combine them, and the artifacts a sovereign exchange would have to describe.

The site is non-commercial and serves educational and research purposes only. It offers no goods or services for sale. Nothing is retrained in the browser; every number shown comes from artifacts prepared by the thesis pipeline.