Development of Bayesian Autoencoders to quantify epistemic and aleatoric uncertainties, within the context of real and virtual drifts in industrial sensors.
We use ensembling with anchored priors under the family of randomised MAP (maximum a posteriori) sampling developed by Tim Pearce.
0_download_data.pydownloads and formats the raw data, lastly pickle the data for convenience1_preprocess_resample.pyresamples and splitting into train/test data. Also applying noise and drifts on one sensor at a time.2_bayesian_autoencoder.pytrains the Bayesian Autoencoder, obtain reconstructed signals and evaluates them.
We use a hydraulic condition monitoring dataset recorded at ZEMA Testbed: https://archive.ics.uci.edu/ml/datasets/Condition+monitoring+of+hydraulic+systems
The research presented was supported by European Metrology Programme for Innovation and Research (EMPIR) under the project Metrology for the Factory of the Future (MET4FOF), project number 17IND12 as well as the PITCH-IN (Promoting the Internet of Things via Collaborations between HEIs and Industry) project funded by Research England. We express our gratitude to Tim Pearce for his inputs.
Uncertainties of reconstructed signals in presence of real drifts (degrading cooler condition), and virtual drifts (injected noise and drift in one of pressure sensors, PS1)
Sensitivity of reconstruction loss, epistemic uncertainty and aleatoric uncertainty towards drifts.
- using a Gaussian likelihood with a full covariance matrix, instead of a diagonal only. Although the implementations are there already in this repo, we need to analyse it in a better way.
- explore other choices of probability distribution for likelihood
- leverage the trio: reconstruction loss, epistemic and aleatoric uncertainties for unsupervised classification
- variant Bayesian Autoencoder architectures, possible options include : convolutional, denoising, clustering,...
- investigate latent space
- apply sensitivity analysis to investigate uncertainties due to all sensors instead of only one at a time.