基于径向基神经网络的机电系统精确模型辨识方法研究

发布时间：2019-01-25 18:36

【摘要】：本文主要研究伺服系统的精确建模问题,通过分析机理建模的复杂与不精确的问题,指出引入神经网络建模能带来的快速性、精确性以及简便性的提升。而目前针对神经网络辨识的研究虽然有很多改进方案,但是大多都只是在一些特定的仿真模型下效果较好,缺乏实际系统的验证,有些算法甚至并不适用于实际系统辨识,因此本文研究基于神经网络的伺服系统精确模型辨识问题,主要的研究成果可归纳为:首先,对一类以永磁同步电机为执行元件的位置伺服系统,进行了标称模型分析与详细的摄动项环节分析,分析了不同非线性环节以及摄动项会对神经网络辨识造成哪些影响,为优化设计神经网络辨识方法提供了理论依据。其次,对比分析了神经网络辨识的基本结构、神经网络的结构特征与选型依据、神经网络训练的基本方法,通过对比指出选择径向基神经网络辨识的选型依据,通过对比训练方法的优劣为神经网络的参数训练方法提供了改进方向。然后,结合伺服系统的特点,提出了适用于伺服系统的两点差分式串-并联辨识结构,优化了神经网络的结构,改进了神经网络参数的训练算法,提出将正交最小二乘法(OLS)与梯度下降法(GD)相结合,能够有效地改进减少神经网络中心节点数量以及降低对初始位置选取的依赖,然后结合伺服系统工作的频段,给出样本数据、测试数据的选择方法以及给出神经网络模型的评价方法。最后得到一个一步预测的模型结构,该结构使用前几时刻的实际数据作为输入能够准确预测出下一时刻的输出,并且通过仿真实验验证了改进结构和训练算法的有效性。最后,结合提出的针对伺服系统改进的神经网络辨识方案,在实际转台伺服系统当中采集开环的训练样本与测试数据,训练出其神经网络模型,再通过与传统的扫频方案得到的模型进行对比,验证了神经网络用于实际系统建模的可行性。
[Abstract]:In this paper, the exact modeling of servo system is studied. By analyzing the complexity and imprecision of mechanism modeling, the paper points out the improvement of rapidity, accuracy and simplicity brought by the introduction of neural network modeling. However, although there are many improved methods for neural network identification, most of them only work well under some specific simulation models, lacking the verification of the actual system, and some algorithms are not even suitable for the actual system identification. Therefore, in this paper, the exact model identification of servo system based on neural network is studied. The main research results can be summarized as follows: firstly, for a class of position servo system with permanent magnet synchronous motor as the actuator, The nominal model analysis and the detailed analysis of perturbation terms are carried out, and the effects of different nonlinear links and perturbation terms on neural network identification are analyzed, which provides a theoretical basis for the optimization design of neural network identification methods. Secondly, the basic structure of neural network identification, the structural characteristics and selection basis of neural network, the basic training method of neural network, and the selection basis of selecting radial basis function neural network identification are pointed out. By comparing the advantages and disadvantages of the training methods, the improvement direction of the neural network parameter training method is provided. Then, combined with the characteristics of servo system, a two-point differential series-parallel identification structure is proposed for servo system. The structure of neural network is optimized, and the training algorithm of neural network parameters is improved. The combination of orthogonal least square method (OLS) and gradient descent method (GD) can effectively reduce the number of neural network center nodes and reduce the dependence on initial position selection, and then combine with the frequency band of servo system. The sample data, the selection method of test data and the evaluation method of neural network model are given. Finally, a one-step prediction model structure is obtained, which can accurately predict the output of the next moment by using the actual data from the previous time as input, and the effectiveness of the improved structure and the training algorithm is verified by simulation experiments. Finally, combined with the improved neural network identification scheme for servo system, the open-loop training samples and test data are collected in the actual turntable servo system, and its neural network model is trained. Compared with the model obtained by the traditional frequency sweeping scheme, the feasibility of the neural network used in the practical system modeling is verified.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP183;TM921.54

【参考文献】