Title:
Digital Apprentice for Chatter Detection: An On-line Learning Approach to Regenerative Chatter Detection in Machining via Human-Machine Interaction

Thumbnail Image
Author(s)
Yan, Xiaoliang
Authors
Advisor(s)
Melkote, Shreyes N.
Advisor(s)
Editor(s)
Associated Organization(s)
Series
Supplementary to
Abstract
Regenerative chatter in machining, which is characterized by self-excited vibration, is a common process anomaly that limits productivity and part quality in machining operations. This thesis proposes an on-line approach for chatter detection via effective human-machine interaction, facilitating knowledge transfer from experienced machinists to the “digital apprentice” through the “learnable skill primitive” (LSP) method that establishes a chatter detection threshold. The research focus is to develop the methodology for chatter-specific knowledge acquisition and a human-machine interface inspired by computing techniques and frameworks such as learning from demonstration, reinforcement learning, and interactive agent shaping. In this work, the milling operation is selected as a case study for the proposed LSP method. Digital audio data is acquired from milling experiments through a studio-style condenser microphone mounted inside a milling machine. The data is pre-processed through various digital filters before Fast Fourier Transform (FFT) is performed to identify the chatter frequency contents. During the training phase, data for the human operator’s natural reaction to chatter is collected via a specially designed human-machine interface. The learned chatter detection thresholds are obtained through the “learnable skill primitive” method by temporally mapping the reaction data to the cutting signal. In addition, a variance mitigation strategy is developed to reduce the negative impact of the high variance in the operator’s reaction time to chatter. During the testing phase, experiments are conducted to evaluate the detection accuracy, detection speed, and robustness of the learned chatter detection thresholds. Experimental data support the claim that the learned thresholds can detect chatter with good detection accuracy and detection speed. Finally, the learned threshold is demonstrated to be robust to milling of different workpiece materials under different cutting conditions such as feeds, speeds, axial and radial immersions (depths of cut), and directions of cutting.
Sponsor
Date Issued
2020-12-07
Extent
Resource Type
Text
Resource Subtype
Thesis
Rights Statement
Rights URI