A Weighted Sparse-Input NN Applied to Identify Important Features for Vortex-Induced Vibration

on YouTube ↗28m

" summary

Proposes a WSPINN weighted sparse-input neural network that applies group sparsity on the input first layer to pick physics features from experiments on a 38 m flexible cylinder with Re=10^4–10^5. Starts from 18 candidate features, forces three trusted priors (Reynolds, shear, damping) with smaller penalty and recovers two additional important features for a 5-feature model that matches brute-force search and keeps prediction error near 10 percent. Optimization works with standard optimizers (FGIO, Adam).

AI-written summary · about

" chapters (13)

1. 0:00Intro
2. 0:27Vortex induced vibration (VIV) for long flexible cylinders
3. 2:44Can we simplify the model?
4. 3:44Brute forces compare prediction for each combination of feature sets
5. 7:19Weighted sparse input Neural network (WSPINN)
6. 8:50Flexible cylinder VIV modeling
7. 10:51Spatiotemporal analysis of typical VIV signals
8. 16:16Dataset description
9. 19:26Feature selection results
10. 22:00Physical insight interpretations
11. 24:28Conclusion
12. 25:562 Neural network setup
13. 27:572. Sparse input Neural Network (SPINN)

" related talks

• 34mData-Driven Solutions in EngineeringVladislav Rosov · Data Science fwdays'19
• 40mSparse models are fast models: Improving DNN inference performance by over 10XLawrence Spracklen · Data Con LA 2021
• 01h 01mAdaptive Wavelet Distillation from DNNs and Dictionary LearningBin Yu · Deep Learning Theory Symposium 2021
• 36mU-Net-style neural networks for feature identification in 1D time-seriesMichael Byington · PyData Seattle 2023
• 35mUncertainty in deep learningMadhu Gopinathan · Anthill Inside 2018
• 57mThe Variational Method of MomentsNathan Kallus · Adversarial Approaches in Machine Learning 2022
• 17mFeature Estimation for Punching Tool Wear at the EdgeOlli Saarela · eSAAM 2023
• 36mBayesian Deep learning with 10% of the weightsRob Romijnders · PyData Amsterdam 2018