开关磁阻电动机的神经网络建模及算法研究

The Neural Networks Modeling of Switched Reluctance Motor and its Arithmetic

针对标准BP算法存在收敛速度慢,容易陷入局部最小值和前向网络拓扑结构中,隐节点选取困难的问题,采用一种由Levenberg-Marquardt算法与改进自构形算法相结合而成的快速自构形算法训练BP神经网络,建立了训练收敛快,泛化能力强,网络规模小,便于实时控制的开关磁阻电机非线性BP神经网络模型.经与样机实测数据对比,验证了该模型的准确性.该模型有助于进一步优化能量转换,减小转矩脉动.

In order to avoid the problems of slow rate of convergence, easy falling into local minimum and difficulty in determining the number of hidden nodes of the feed-forward neural network in BP algorithm, a fast self-configuring arithmetic was used to train BP neural networks. This arithmetic was the integration of Levenberg-Marquardt arithmetic and the improved self-configuring arithmetic. Furthermore, the nonlinear model of Switched Reluctance Motor（SRM） was developed, which made use of the measured accurate flux-linkage data and the nonlinear mapping ability of BP neural network based on fast self-configuring arithmetic. This neural network model was of fast training convergence, generalization and small network scale handy for real-time control. The accuracy of this model was proved with experiment results. This model will help to optimize energy conversion and will reduce torque ripples.