fault diagnosis and load detection in electrical machines using vibration analysis and neural nets

rotating machines in particular induction electrical machines are important industry instruments. in manufacturing, electrical motors are exposed to many damages, and this causes stators and rotors not to work correctly. in this paper we addressed modal analysis and an intelligent method to detect motor load condition and also the stator faults such as turn-to-turn and coil-to-coil faults using motor vibration analysis. a three-phase induction motor with a special winding was used to create the faults artificially. the vibration signal of motor in different states such as working without fault, with various faults and with various loads was acquired. some spectral analysis was done using the spectrum and the spectrograph of vibration signals and differences due to different states of motor were observed. suitable features such as linear prediction cepstral coefficients and fourier transform filter bank coefficients were extracted from vibration signals and were then applied to non-supervised (som) and supervised (lvq) neural networks in order to classify motor faults and its load condition. many experiments were conducted to evaluate the effect of neural network type, type and length of feature vector, length of training signal etc. in brief, using som and lvq neural networks, 20 element filter bank feature vectors, and 600ms of the training data, performance of 93.6% and 94.2% were obtained for load and fault detection respectively.