Learning Deep Dynamical Models of a Waste Incineration Plant from In-furnace Images and Process Data

This paper presents an approach for predicting in-furnace images and sensor signal readings for a waste incineration plant, utilizing a deep dynamical model based on Kalman Variational Auto-Encoders that considers a range of process signals, control inputs, and time-series sequences of infurnace image data. This is motivated by the need for automatic control systems to be able to anticipate abnormalities in incoming waste to prevent potential instabilities during and after combustion. Experimental results with real plant data show that the proposed strategy provides an improved prediction accuracy for both process signals and in-furnace images compared to a Long Short-Term Memory neural network.