Fault detection for ironmaking process based on stacked denoising autoencoders

It is quite challenging to monitor an ironmaking process due to some of its special characteristics such as lack of direct measurements and strong disturbances. Hence extracting robust features of the normal process from complex historical data is vitally important. Denoising autoencoder (dA), a recently developed deep learning technique, has become a popular tool to extract and compose robust features. However, its application to fault detection in process control fields are still limited. In this paper, a denoising autoencoder based monitoring approach is proposed for a practical ironmaking process, in which peak-like disturbances due to the switchings between two arbitrary distinct host-blast stoves are involved. To validate the proposed monitoring method, the data corresponding to a cold furnace fault of the process is used and comparative fault detection performances with the existing methods are presented.