Automated product grade transitions, exposing the inherent and latent dangers of neural networks in manufacturing process control: an industrial case study

Abstract

The proliferation of neural network based solutions in manufacturing process control is causing concern for many engineers in the industry [American Automatic Control Conference (paper FP18-2), Denver, 2003; AIChE Spring Meeting (paper 144f), New Orleans, 2003]. The intrinsic and latent dangers of neural nets (within a control framework) are often overlooked when assessing a suitable technology for real-time control application. This case study examines the intrinsic properties of a neural net that make it wholly unsuitable as a technology directly or indirectly involved in the manipulation of process valves. Of particular concern is the optimization of product grade transitions. Apart from start up and shutdown, product grade transitions are the most safety critical procedures performed on a manufacturing plant. Any control system involved in the automation of such procedures should be of the highest integrity and adhere to sound, intrinsically safe control engineering principles—something neural networks categorically fail to provide. The objectives of this case study are twofold. Firstly, the aim is to raise awareness in the manufacturing industry of the inherent and latent dangers of neural networks in control. The second objective is to provide a platform for best practice guidance on what is and is not acceptable as a modeling paradigm in a manufacturing process control scheme.