The machine made the part, and the recorded duration is real — which makes it look like the ideal ground truth for a cycle-time model.
But real is not the same as representative. A longer observed program duration may indicate an inaccurate estimate, or it may reflect maintenance, operator intervention, shift effects, tool problems or other operational disruption. Recent research into CNC cycle-time anomaly analysis explicitly filters exactly these kinds of disturbances — server issues, maintenance, shift changes — before computing actual program-duration distributions for analysis.1
That does not mean every outlier should be deleted. A recurring tool failure may itself reveal that a process choice is not robust, and a consistently difficult setup may point to a genuine structural problem rather than noise to be discarded. The useful distinction is therefore closer to representative variation versus explained exception versus unexplained anomaly.
For manufacturing AI, production actuals need context and provenance: a measured result is evidence of what happened once, and learning what is likely to happen next requires understanding why.
Notes & references
Vlaminck, N., Nicolas, M., Benamara, T. & Raddoux, H. (2025), “Analyzing Machining Cycle Time Anomalies via CNC and Operational Data,” Procedia CIRP, 133, 471–476 (20th CIRP Conference on Modeling of Machining Operations). DOI: 10.1016/j.procir.2025.02.081