联合粉磨系统磨机负荷辨识方法研究

发布时间：2018-03-07 09:12

本文选题：联合粉磨系统　切入点：磨机负荷　出处：《济南大学》2015年硕士论文　论文类型：学位论文

【摘要】：磨机是联合粉磨系统中的核心设备,然而大部分磨机都处于低效率、高耗能的状态,且粉磨过程具有高耦合性等特点,因此对磨机负荷的准确辨识尤为重要。为了得到最优的磨机负荷模型,分别采用了四种递推最小二乘算法、RBF神经网络以及T-S模糊模型对磨机负荷进行了辨识。本课题完成的主要工作概况如下:通过对联合粉磨系统和磨机负荷辨识发展现状的概述,并结合大量的历史数据,得出了磨机的主电机电流最能表征磨机负荷,以及分析出影响主电机电流的主要变量有:总量给定、选粉机转速、循环风机转速、分料阀开度以及粉煤灰库提升机电流。最终确定了总量给定和选粉机转速对主电机电流影响最大,故作为关键变量,其他三个变量为不确定因素。文中所有的负荷模型都以总量给定和选粉机转速作为输入变量,主电机电流作为输出变量。首先采用了四种递推最小二乘算法对磨机负荷进行辨识,具体的算法有:递推最小二乘法、遗忘因子递推最小二乘法、限定记忆递推最小二乘法和偏差补偿递推最小二乘法。通过仿真结果分析,得出在该工况下,加入遗忘因子和采用偏差补偿策略的模型能很好的跟踪主电机电流的变化情况,其中基于偏差补偿算法的模型最为精确。基于普通递推最小二乘和限定记忆的模型拟合误差相对较大,不适合该工况下的建模。然后采用了神经网络对磨机负荷进行辨识。由于RBF神经网络对非线性系统具有良好的逼近性能等优点,故分别采用了基于高斯核函数,多二次核函数和逆多二次核函数的三种RBF网络模型进行了辨识。径向基函数的中心,基宽度以及连接的权值均采用梯度下降法进行训练。神经元的个数通过反复实验来确定。通过分析平均误差、均方误差等性能指标,最终得到基于高斯核函数的RBF神经网络模型具有较高的精确度,更适合估计该工况下的磨机负荷。最后采用了T-S模糊模型对磨机负荷进行辨识。利用模糊C-均值聚类算法将输入变量划分为四个子空间,并利用加权最小二乘法对模糊后件的参数进行了辨识,得到了较为精确的T-S模糊模型。为了对比文中所用三类辨识方法的有效性和建模精度,在文章的最后分别用最小二乘法中的加权最小二乘法和基于高斯核函数的RBF神经网络对同一段历史数据进行建模。由仿真结果可知,相比于加权最小二乘法辨识出的模型和基于高斯核函数的RBF神经网络模型,T-S模糊模型能很好的反映出该段工况下磨机主电机电流的变化情况。
[Abstract]:The grinding machine is the core equipment in the combined grinding system. However, most of the grinding machines are in the state of low efficiency and high energy consumption, and the grinding process has the characteristics of high coupling, etc. Therefore, it is very important to identify the mill load accurately. In order to get the optimal load model, Four kinds of recursive least square algorithm (RBF) neural network and T-S fuzzy model are used to identify the load of mill. The main work of this paper is as follows: the development of load identification of combined grinding system and mill is summarized. Combined with a large number of historical data, it is concluded that the main motor current of the mill can best characterize the mill load, and the main variables that affect the main motor current are as follows: the total quantity given, the speed of the separator, the speed of the circulating fan. The opening of the valve and the current of the hoist in the fly ash storehouse are determined to be the most important variables for the main motor because the total quantity and the speed of the separator have the greatest influence on the current of the main motor. The other three variables are uncertain factors. All the load models in this paper take the total quantity given and the speed of the separator as input variables. The main motor current is used as the output variable. Firstly, four recursive least square algorithms are used to identify the mill load. The specific algorithms are: recursive least square method, recursive least square method of forgetting factor, The finite memory recursive least square method and deviation compensation recursive least square method are used. Through the analysis of simulation results, it is concluded that the model with forgetting factor and deviation compensation strategy can track the change of main motor current well. The model based on deviation compensation algorithm is the most accurate, and the model fitting error based on ordinary recursive least squares and limited memory is relatively large. It is not suitable for modeling under this condition. Then, neural network is used to identify mill load. Because RBF neural network has good approximation performance to nonlinear system, it adopts Gao Si kernel function respectively. Three kinds of RBF network models of multi-quadratic kernel function and inverse polyquadratic kernel function are identified. The base width and the weight of connection are trained by gradient descent method. The number of neurons is determined by repeated experiments. The average error, mean square error and other performance indexes are analyzed. Finally, the RBF neural network model based on Gao Si kernel function has high accuracy. Finally, T-S fuzzy model is used to identify the mill load. The input variables are divided into four subspaces by using fuzzy C-means clustering algorithm. By using the weighted least square method, the parameters of the fuzzy rear parts are identified, and a more accurate T-S fuzzy model is obtained. In order to compare the effectiveness and modeling accuracy of the three identification methods used in this paper, At the end of the paper, the weighted least square method and the RBF neural network based on Gao Si kernel function are used to model the same historical data. Compared with the model identified by the weighted least square method and the RBF neural network model based on Gao Si kernel function, the fuzzy model of T-S can well reflect the change of the main motor current under this working condition.
【学位授予单位】：济南大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TQ172.63

【参考文献】