Feature Estimation for Punching Tool Wear at the Edge
summary
Pulses of 20 kHz accelerometer data produce one output number per punch and models run on Raspberry Pi–class edge hardware. Feature sets from MiniRocket produce classification accuracy from about 84% to over 96.5%, while TSFEL features yield 35–56%; logistic regression on MiniRocket features works in real time. Full MiniRocket gives ~10,000 features but ten selected features nearly match performance. Dataset is very limited; a finite‑element simulator exists for augmentation.
AI-written summary · about
chapters (11)
- 0:00Punching tool wear diagnostics
- 1:28Condition-based maintenance
- 3:15Punching process and tool
- 5:06Vibration data challenges
- 6:34Feature extraction libraries
- 8:01Overfitting risks and selection
- 9:05Edge and cloud analytics
- 11:13Q&A: Edge maintenance
- 12:58Q&A: Feature examples
- 15:09Q&A: Business models
- 16:53Closing remarks
related talks
45mUsing Edge Machine Learning to save power and development time
22mRunning, walking, sitting or biking? - Motion prediction with acceleration and rotation
03mPowered by TensorFlow: Creating a custom, machine learning-powered drumming arm
05mBuilding Machine Learning models with Edge Impulse
24mAccelerating AI on Edge
55mEdgeAI - AI in the Palm of Your Hand
24mAshfaq Munshi, ML7 Fellow, Pepperdata
39mLeveraging the power of C++ for efficient machine learning on embedded devices