A proposed method for representing the real-time behaviour of technical processes in neural nets

The application of neural networks in supervision and control of technical processes requires not only their abilities to classify process states and identify possible faulty or dangerous ones but also the possibility to monitor changes of process variables over time in order to predict eventually developing dangerous states. However, the traditional methods of teaching neural nets have shown that nets did nor provide for this later capability, when they were trained by a data set in which all evolutionary states from which faulty or dangerous situations can arise are involved. The following paper presents a method of teaching neural nets by means of sequences of sets of process values which converges towards process states that are known to be faulty or dangerous. The method had originally been developed with the aim of improving the quality and speed of the detection of static process states, but can also be applied to the early detection of changes in the process that may lead to dangerous states. Until now, measurements with a simulated coal-fired power plant have shown very promising properties of the proposed mechanism. So a neural net ruled supporting and warning system has been conditioned by data sets representing the plant when all parts are working at their operation points and by some sets representing clearly presenting faulty states to create an undressed basis structure/concept of the neural net classificator. This basis structure was successively sensitized by teaching evolutionary states of the faulty states which were younger and younger in its development history. The test results showed that now even slightly from each other differing sensor patterns and/or evolutionary stales of arising faults can be detected. Especially the net can separate even faulty states when 2 of 157 data of the sensor representation of the plants working condition changed about 2% only