← Work
Machine Learning & Perception
Real-Time Machining-Chatter Detection
~97% accuracy on noisy real-world signals, trained only on clean data, where standard models collapse to ~46%.
Chatter is unstable vibration that ruins surface finish, breaks tools, and drives up scrap. The hard part is detecting it on a real machine whose signals look nothing like clean lab data.
I built a real-time detector that survives that clean-to-noisy gap: a multi-view representation of each 50 ms window — waveform, log-spectrum, and amplitude/energy — fused by a CNN-BiLSTM plus a gradient-boosted-tree ensemble. It runs signal-only and CPU-only under a 100 ms budget for live edge deployment.
