Abstract: Vibration signals from a hydropower unit are non-linear and non-stationary, but they are similar in different scales and typical of fractal features. This paper reports a multi-fractal method of fault diagnosis for hydropower units, analyzing the vibration signals, extracting their generalized dimensional spectral features, and diagnosing the fault with a probabilistic neural network optimized by the artificial fish swarm algorithm. A case study shows that this method, through combination of multi-fractal and probabilistic neural network, can accurately distinguish fault types. Compared with a BP or RBF network, it achieves a higher diagnostic recognition rate and faster speed, thus providing a more reliable tool for unit operation and maintenance personnel.

Key words: hydropower unit, fault diagnosis, multi-fractal, probabilistic neural network (PNN), artificial fish swarm Algorithm (AFSA)