Neural network-based incipient fault detection of induction motors

An incipient fault detection scheme of induction motors through the recognition of frequency spectra of the stator current has been developed in this thesis. It is based on the adaptive resonance theory of neural networks. This fault diagnosis scheme is not only capable of detecting a fault but also can report if it cannot identify a particular fault so that necessary preventive steps can be taken to update the underlying neural network to adapt to this undetected fault. Moreover, it can update itself to cope with this dynamic situation retaining already acquired knowledge without the need of retraining with the old patterns.;A laboratory experimental set-up using a digital signal processing (DSP) technique has been employed to collect the frequency spectra of the stator current at different fault conditions. A wound-rotor induction motor has been used as the test motor to create different types of faults making unbalance in the stator and rotor circuits. A 24-bit high speed DSP board has been used with a personal computer to develop a real-time interactive software to collect the spectra. A driver for the HP-plotter has also been developed to directly plot the frequency spectra of the stator current.;Adaptive resonance theory (ART) based network is a recent addition to the neural network family. A new software has been successfully developed and implemented in the laboratory experiment using ART neural network. Its performances in training, recalling and dynamic updating have been studied with a set of example patterns. The incipient faults of a 3-phase wound rotor induction motor have been successfully diagonized by this neural network.