Physics-Informed ML-based Chatter Detection for Thin-Walled Milling

Procedia CIRP, vol. 141, pp. 191–196

Abstract

Regenerative chatter, a self-excited vibration arising from the phase lag between instantaneous chip thickness and the waviness left by preceding tooth passes, continues to limit the precision and productivity of milling thin-walled aero-engine components such as blisks. In these compliant structures, closely spaced modes, tool-spindle interactions, and progressive material removal cause local variations in stiffness and damping. The consequences include compromised surface integrity and dimensional accuracy, accelerated tool wear, elevated noise, and increased scrap. This work seeks to develop and validate a generalized, real-time chatter detection framework that maintains accuracy under the evolving dynamics characteristic of industrial milling processes. The proposed pipeline integrates Empirical Mode Decomposition (EMD) and the Hilbert-Huang Transform (HHT) with a sliding-window anomaly detection scheme based on machine learning. Instantaneous frequency and energy of intrinsic mode functions are used to adapt to time-varying modal content driven by material removal and shifting boundary conditions. Deployment is supported by established Machine Learning Operations (MLOps) practices, including containerized services and MLflow for experiment tracking, versioning, and lifecycle management. System dynamics are characterized a priori through impact testing of the tool-spindle assembly, Finite Element Method (FEM)-based modal analysis of the workpiece, and cutter-workpiece engagement simulations. These analyses provide eigenfrequency maps and stability-lobe diagrams that define chatter frequency bands across cutter locations. Experiments across different milling configurations confirm consistently high detection accuracy and robustness to dynamic shifts, demonstrating strong transferability to new setups. This study provides a practical and scalable approach for real-time chatter monitoring, supporting more reliable and efficient machining of critical milling components.

Authors 6

1. Aakash Singh Aachen Fraunhofer IPT

   Fraunhofer Institute for Production Technology IPT

   Affiliation as printed

   Fraunhofer IPT, Steinbachstr. 17, 52074 Aachen, Germany

2. Michał Słupczyński Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, Computer Science 5 - Databases and Information Systems, Ahornstr. 55, 52074 Aachen, Germany

3. Michal Kovac Aachen Fraunhofer IPT

   Fraunhofer Institute for Production Technology IPT

   Affiliation as printed

   Fraunhofer IPT, Steinbachstr. 17, 52074 Aachen, Germany

4. Viktor Rudel Aachen Fraunhofer IPT

   Fraunhofer Institute for Production Technology IPT

   Affiliation as printed

   Fraunhofer IPT, Steinbachstr. 17, 52074 Aachen, Germany

5. Stefan Decker Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University, Computer Science 5 - Databases and Information Systems, Ahornstr. 55, 52074 Aachen, Germany

6. Thomas Bergs Aachen Fraunhofer IPT

   Fraunhofer Institute for Production Technology IPT

   Affiliation as printed

   Fraunhofer IPT, Steinbachstr. 17, 52074 Aachen, Germany