In this study

a BP (back-propagation network) neural network

optimized by PSO (particle swarm optimization) was applied to the fault diagnosis of a centrifugal chiller. Seven typical faults

including four component-level and three system-level faults

were investigated. Results showed that the performance of fault diagnosis was significantly improved (for both single- and double-hidden BP layers) compared with the model without PSO. The optimization simplified the structure of the neural network from 18 neurons to 10 neurons for a single-hidden-layer network and from 25 neurons to 12 neurons for a double-hidden-layer network. This increased the correct rate of fault diagnosis from 89.42% to 95.30% and from 97.87% to 98.11% for single-hidden-layer network and double-hidden-layer network

respectively. There are also considerable savings in diagnostic time

especially for the double-hidden–layer network

to only 23% of that before optimization. The cases of "false report" and "leaked report" have been reduced

and the false alarm rate is also lower than before. Moreover

the diagnosis performance of the system-level fault

especially the RefLeak (Refrigerant Leakage)

and the recognition rate of the normal condition are greatly improved. Through PSO

the BP network is able to jump out of the local minimum and greatly improve the fault diagnosis performance for the centrifugal chiller.