Neural Network Based Approaches To Fault Diagnosis Of Control Systems

It is acknowledged that modern industrial processes may involve agreat deal of control loops, whose faults associated with any loopcomponent may lead to deterioration of the overall control performance,even severe fault in production. To accommodate dynamics andnonlinearity of control system faults, this paper proposes a generalizedmulti-model dynamic RBF neural networks based approach to contolloop fault diagnosis.Initially, an in depth analysis of common faults associated withcontrol systems is conducted along with simulation work. A generalizedmulti-model dynamic neural network system which combines bothmetrics of generalized process models and multi-model dynamic neuralnetworks is devised. Therein, two modules are involved. One module isto choose a most compatible generalized process model after identifyingtime constants and gains of the practical process. Another modulecorresponds to the multi-model dynamic RBF neural networks baseddiagnosis system. This neural networks system is trained off-line bymeans of data pertaining to generalized process models before being employed online to diagnosis control system. Additionally, the influencesof time constants and gains of generalized process models on multi-modeldynamic neural networks are investigated. As a result, the system is ableto detect the faults of control system whose characteristics are within aprescribed boundary. Simulation results show that the proposed systemcould be considered as a good alternative to diagnosis control system. Inaddition, fault diagnosis of the stripper level control loop consisting in TEprocess is performed, giving rise to satisfied diagnostic results, whichdemonstrates potential effectiveness of the proposed methods applied toindustrial processes.Moreover, we are particularly concerned with control valve stictionwhich is recognized as a major cause of control loop oscillation. It isworth noting that multi-model dynamic neural networks could beresponsible for nonlinearity of control valve stiction, thereby proposing aneural network based method in detecting viscous stiction of controlvalves. In addition, an S-function of Kano models is developed byutilizing Simulink toolbox, which is validated by means of simulations aswell.