A real-time recurrent learning network structure for dealing with missing sensor data

A lot of effort has been put into the modeling of highly nonlinear dynamical systems. In every part of the process industry, soft sensors have became a hot item. These soft sensors (neural models) have the ability to map a variety of input-output patterns quite easily. Although they have many advantages over other, more classical modeling techniques or hard sensors, neural networks also have a number of drawbacks. They are insensitive to the availability of sensor data. The practical use of soft sensors, however, shows that this is hardly the case. Network performance reduces rapidly due to lack of sensor data. To cope with this kind of problems a network structure for real-time recurrent learning networks was developed. Two recurrent networks, a model network and an identity network, are merged into one large clustered recurrent net with combined advantages of robustness and high modeling performance. This technique has been tested on a real-life modeling problem from the chemical process industry. Compared to the original model network the clustered recurrent net enhances the robustness to handle missing data and its accuracy of prediction has been greatly improved.