Fault diagnosis with neural networks. Part 1: Trajectory recognition

The present investigation was focused on formulating a method for designing a fault diagnosis system for chemical plants by using artificial neural networks. Fault diagnosis is aimed at identifying a fault which affects a given process by analysing the signs supplied by process sensors. Neuronal networks are mathematical models which try to imitate the functioning of the human brain. A neural network is defined by its structure and the learning method used. The difficulty with diagnosing faults lies in recognising the tralectories (temporal series of data) followed by process variables when a fault affects the process; when tralectories are recognised, the associated fault is also identified. The theory so developed recommended an optimised structure and training method for the neural networks to use. Both the proposed structure and the training method were tested by carrying out comparative studies between traditional structures and a training method. The results showed the superiority of the neural networks designed and trained with the method proposed in this work. Except for simple processes, fault diagnosis is a more complex problem than simply identifying tralectories, because a fault may cause an infinite set of tralectories (i.e. flow). The fundaments established in this work are thus used in Part II, where the analysis is extended to recognise flows