Asynchronous Motor Fault Diagnosis Based On Wavelet Neural Network

As the transmission machinery, asynchronous motors are widely used in industrial production and defense field, such as power plants, steel mills, naval vessels, so its safe operation is essential. Asynchronous motor fault detection, especially the initial issuance of the fault detection is one of the key measures which protect the safe operation of asynchronous motors. Therefore, it is important and urgent to prevent the fault from occurring and reduce maintenance expenditure.The paper analyzes faults mechanism for four types of common faults of AC motor. The faults are shown as follows: inter-turn short circuits of stator windings, rotor bar broken, rotor eccentricity and the bearing fault. This paper selects the stator current signal as the signal of motor faults and induces the fault characteristic frequency .Wavelet analysis has the character of time-frequency localization and can decompose the signal into different frequency bands. So the paper uses the wavelet decompose to eliminate signal noise of stator current signal which contains information of the broken rotor bar fault, and uses the wavelet packet to decompose the signal, and then reconstructs the wavelet coefficients. At last the paper uses the FFT to compute the power spectrum. The simulation results show that, wavelet transform has successfully extracted the fault frequency, and it can eliminates influences of the fundamental frequency components covering the fault frequency. But the method that uses the FFT can not detect broken rotor bar fault frequencies. It is proved that the wavelet analysis method is a powerful signal processing and it is a good tool to extract the fault feature vector.The paper uses the wavelet neural network (WNN) to establish the asynchronous motor fault diagnosis model according to the mapping relationship between the common symptoms of fault in the asynchronous motor and fault mode. The model adopts the conjugate gradient descent algorithm, which is optimized by the momentum and adaptive learning rate. The initialization of parameters of the WNN is also analyzed in the paper. It is shown from the simulation results that, compared with conventional wavelet neural network and BP network, this model significantly reduces the training time and is valid for motor fault diagnosis.