U-Net-style neural networks for feature identification in 1D time-series
summary
U‑Net–style 1D networks identify discrete events (pipe joints) by expanding the convolutional receptive field >10,000 points so each label sees many repeats; architecture is ~12.6M weights trained on 111 surveys with 91,000 labeled joints augmented into ~40 million distinct examples and produces clear joint spikes that peak‑pick reliably. A piper device collects magnetics with no GPS; transfer learning works after freezing layers and fine‑tuning (400 EKG labels → successful ECG peaks), while earthquakes fail without more data.
AI-written summary · about
chapters (7)
- 0:00Pipeline inspection problem
- 7:22Neural networks and U-Net basics
- 13:02Adapting U-Net to 1D time-series
- 16:51Data augmentation for small datasets
- 19:19Transfer learning across domains
- 26:20Key takeaways and applications
- 26:41Q&A
related talks
27mU-Net is all you need: applying the same AI approach for many different business domains
13mEnhanced EEG Data Classification with 3D SNN and On-Line Learning ESN
59mDeep Neural Networks for Video Applications
24mProject GaitNet: Ushering in the ImageNet moment for human Gait kinematics
21mLand Boundary Delineation using U-Nets
38mThe Joy of Designing Deep Neural Networks
01h 30mDeep Learning Techniques for Monocular Depth Estimation
34mWhy AI in the Medical Field Both Sucks and Rocks...