Imperial AI Hack 2020

• Dema Ushchapovskyy

• October, 2019

24 hour challenge to predict anomalous behaviors and their cause of an industrial compressor.

Experimented with linear and non linear feature extraction and build linear models such as ARIMA and VAR. Final solution was to build a variational autoencoder and cluster the latent representation to identify running regimes of the compressor. The output distributions of the autoencoder were used to identify causality by ranking features by z-scores.

Final Place: 1st.