Combustion anomalies detection for a thermal furnace based on Recurrent Neural Networks

This paper describes the application of Recurrent Neural Networks (RNN) for effectively detecting anomalies in time series data obtained from experimental study of the combustion and gasification of mechanically activated coal fuel in a thermal furnace. We train Recurrent Neural Networks (RNN) with Long Short-Term Memory (LSTM) units to learn the normal time series patterns and predict anomaly values. The resulting prediction errors between real and expected values are analyzed to give anomaly scores. To investigate the most suitable configuration of RNN and evaluate the effectiveness of the anomaly detection model, we used three datasets of real-world data that contain several types of anomalies. The developed RNN algorithm detected 9 out the 9 collective anomalies in the hold-out sample with one false positive anomaly event.