Machine Learning for Weld Acoustics Monitoring
Caterpillar Inc. and University of Illinois Urbana-Champaign · Peoria, IL
A system that classifies weld quality from the sound of the weld itself. Welding produces a characteristic acoustic signature; the question was whether that signature carries enough information to flag a bad weld as it happens, without instrumenting the workpiece.
Pipeline
- MFCC feature extraction from raw acoustic recordings, in Python with Librosa.
- SVM classifiers trained on Caterpillar welding lab datasets.
- Python GUI giving the operator immediate feedback.
- Arduino and LED prototype alert stack for a shop-floor signal.
Result
75–84% test accuracy on the Caterpillar welding lab datasets.