Bayesian Deep Learning for Manufacturing 2.0

Highlights and New Additions

1. Bayesian 3D U-Net model integrating Bayesian layers and attention blocks for uncertainty quantification and superior decoder performance leveraging the where to look capability with multi-task capabilities to estimate bot real-valued(regression) and categorical(classification) based values. The Decoder is used to obtain real-valued segmentation maps
2. Deep Reinforcement Learning using deep deterministic policy gradient (DDPG) and a custom made multi physics manufacturing environment to build agents to correct manufacturing systems
3. Closed Loop Sampling for faster model training and convergence using epistemic uncertainty of the Bayesian CNN models
4. Matlab Python Integration to enable low latency connection between multi-physics manufacturing environments (Matlab) and TensorFlow based DDPG agents
5. Multi-Physics Manufacturing System Simulations to generate custom datasets for various fault scenarios using Variation Response Method (VRM) kernel
6. Uncertainty guided continual learning to enable life long/incremental training for multiple case studies
7. Exploratory notebooks for various case studies
8. 3D Gradient-weighted Class Activation Maps for interpretability of deep learning models
9. Datasets for Industrial multi-station case studies for training and benchmarking deep learning models

Bayesian Deep Learning for Manufacturing

Object Shape Error Response Library