Comparative study of probabilistic neural network and back propagation network for fault diagnosis of refrigeration systems

Abstract

Fault diagnosis is of crucial importance to mechanical systems. To find an efficient way of conducting fault diagnosis for refrigeration systems, probabilistic neural network, and back-propagation network were employed to diagnose seven types of typical faults for a 90-ton centrifugal chiller. These include system-level faults such as refrigerant leak/undercharge, refrigerant overcharge, and excess oil and also component-level faults such as condenser fouling, reduced condenser water flow, noncondensables in the refrigerant, and reduced evaporator water flow. Eight major features of the chiller system were selected as indicative parameters for diagnosing. The establishment of the fault diagnosis models based on the probabilistic neural network and back-propagation network and the optimization processes of the networks were elaborated. Comparison in terms of diagnostic performance was performed based on the optimized networks. The results showed that the overall diagnostic performance of the probabilistic neural network was better than that of the back-propagation network. The probabilistic neural network has a correct rate that was 3.48% higher than that of back-propagation, and its diagnosis time was lower by more than 400 times. Moreover, the diagnosis of a single probabilistic neural network training was more reliable than that of the back-propagation network. It was also demonstrated that system-level faults were more difficult to be recognized by the model than component-level faults because of their widespread influence on the system operation. The probabilistic neural network model has a better performance on system-level faults with an improvement of correct rates about 4%∼8.7% from those of back-propagation model.