Research On Motor Flux Observer Base On Neural Network

In fact, the amplitude and phase of stator flux can not be obtaining by sampling directly, it usually be calculated indirectly by sampling the factors easily be obtained such as stator voltage, current and speed. Torque is the key quantity of DTC, and it is relate to the stator flux. So, accurate stator flux obtaining is the key factor to achieve good dynamic performance of torque response in DTC system.An observing method of motor stator flux based on neural network theory is studied in this paper. The main work is as follows:Three flux observers of conventional DTC is described in this paper at first, and then analyses the merits and demerits of these methods, especially detailed analyze the reason of the DC offset and initial phase issue bring by the pure integrator of the voltage-based method. Then introduce some modified methods of the pure integrator proposed by the other scholars, including the first-order low-pass filter, amplitude limited method, amplitude and phase compensation method, adaptive integrator method, and analyze the advantages and disadvantages of them, beside, introduce the selection of the cut-off frequency. Among those methods, stator flux with amplitude and phase compensation is easy to implement, it can make a good DC offset suppression if select a appropriate cut-off frequency. But, its adaptive ability is bad; it's difficult to observe an accurate flux when the stator frequency and load change because of the fix cut-off frequency.To resolve the shortage of the stator flux observer with amplitude and phase compensation, a RBF neural network flux observer is proposed in this paper. At last, an experiment to test the method proposed in this paper, the results show that, the trained RBF neural network can achieve the function well of variable cut-off flux observer with amplitude and phase compensation. Not only that, the point is RBF neural network with great generalization and adaptive ability improves the accuracy of flux observation in the region of two discrete cut-off frequecy. Consistent with the theoretical analysis, the test results demonstrate the validity of the method proposed in this paper.