Improvements in failure detection of DAMADICS control valve using neural networks

This paper shows the results on the detection and isolation of failures in the DAMADICS control valve. The mathematical model extracted for the valve was identified as a first order ARX. The blocking, sedimentation and erosion failures were selected in order to perform the experimentation. With the goal to identify and isolate faults on the valve, a fault Detector was designed by parametric estimation of the model, which allowed to determine the values of the parameters “a” and “b”, and their behavior associated with the failures. As result of this procedure, we found that those parameters affect the gain from the valve. Furthermore, we noticed that this conventional Detector presents problems in the threshold regions of each failure. For solving that problem, a radialbased artificial neural network was added as a complement to the conventional Detector by parameter identification, which allowed us the correction of the error in the thresholds. The proposed Detector that includes the neural network showed better performance detecting and isolating failures.