基于增量改进BP神经网络微波深度干燥模型及应用研究

【作者】 李英伟；

【导师】 彭金辉；

【作者基本信息】 昆明理工大学 ， 冶金工程控制， 2011， 博士

【摘要】 BP神经网络作为人工智能研究领域的重要分支,具有良好的非线性映射能力和高度的并行信息处理能力,在多学科交叉技术领域得到广泛的应用。微波干燥不同于传统干燥方式,其热传导方向与水分扩散方向相同。与传统干燥方式相比,具有干燥速率大、节能、生产效率高、干燥均匀、清洁生产、易实现自动化控制和提高产品质量等优点,但是在干燥过程中,影响微波干燥的因素包括微波输入功率、微波作用时间、物料初始含水量、物料质量、物料转速等,其在干燥过程中影响程度不同,致使微波预处理富硒渣的干燥过程试验周期长,试验量大且参数不易优化。选择具有非线性映射能力的BP神经网络,对微波干燥过程建立仿真模型,预测分析试验过程。标准的BP算法是基于梯度下降法,通过计算目标函数对网络权值和阈值的梯度修正网络权值,在训练过程中存在收敛速度慢和局部最小的问题；且对于复杂的问题,在训练过程中会陷入局部最小点,以致无法收敛。因此,需对BP神经网络算法进行改进。本论文以微波深度干燥富硒渣的工业化试验过程为研究对象,建立基于增量改进BP神经网络的预测模型和Smith补偿PID控制模型,及研究预测模型在微波煅烧领域的应用,主要研究内容是：1)采用Levenberg-Marquardt (L-M)算法对BP神经网络加以改进,提高了神经网络的收敛速度,同时针对在训练神经网络的过程中,无法一次性提供所需的训练数据且当样本规模较大时,系统内存的限制使得对所有样本的训练不可行等问题,提出基于增量学习的BP神经网络。结合在单个神经网络中设置权有效区域和添加隐含层单元数的方法实现增量学习,在新样本知识与原有样本知识接近时,可以在一定范围内修改权值和阈值,同时采用灵活的方式确定隐含层节点数目。在利用训练样本集对神经网络进行训练的过程中,不断增加隐含层节点数目,并计算输出误差,若误差达到要求就停止网络训练,此时网络隐含层节点数即为最优的隐含层节点数。此方法既可以使网络学习新样本知识,又可以使网络保持原有样本的知识。基于增量学习和L-M优化算法的改进BP神经网络,弥补了传统BP神经网络的不足,它具有更快的收敛性,较好的预测精度和更好的拟合结果,并能够避免误差总和不再更新,网络不再训练以致使网络瘫痪等问题,在调整网络参数时不会陷入局部最小,使网络迅速收敛,能够有效地解决训练数据不能一次性提供,以较少的先验知识学习过程的特点等问题,选择有代表性的样本在占用较少内存的前提下训练神经网络,即能够保持原有知识,又能够学习新的知识。2)在微波深度干燥富硒渣的工业化试验中,建立了增量改进BP神经网络的非线性系统预测模型,以微波输入功率、微波作用时间、物料初始含水量、物料质量和物料盘转速为输入条件,用以预测微波深度干燥富硒渣的工业化试验结果。3)建立了增量改进BP神经网络的能耗预测模型,以微波干燥工业化试验中微波功率、微波作用时间、物料质量、物料初始含水率和所需物料最终含水率为输入条件,预测工业化试验过程的能耗。4)在控制领域中,传统的PID控制器结构简单,对模型误差具有鲁棒性及易于操作等优点,被广泛应用于冶金、化工、电力、轻工和机械等工业过程控制领域中。随着工业的发展,被控对象的复杂程度不断加深,尤其对于大滞后、时变的、非线性的复杂系统,传统PID控制已经无法满足目标控制精确化的要求。建立增量改进BP神经网络的Smith补偿PID控制模型,根据工业化试验过程测定的被控对象参数作为神经网络的输入,对控制系统进行离线系统辨识,利用简化了的微波深度干燥富硒渣工业化生产过程的控制模型对增量改进BP神经网络的Smith补偿PID控制器进行仿真研究,实现在线动态整定PID控制参数。5)建立了增量改进BP神经网络反预测模型,以物料的最终质量、物料的最终温度和物料的相对脱水率为输入条件,预测微波干燥工业化试验所需的工作时间、物料初始含水率和试验能耗。6)将工业化微波深度干燥富硒渣的神经网络预测模型应用于微波煅烧重铀酸铵(ammonium diuranate, ADU)和三碳酸铀酰铵(ammonium uranyl carbonate, AUC)试验中,用以预测微波煅烧ADU和AUC的试验过程结果。更多还原

【Abstract】 As an important branch of artificial intelligence research area, Back-Propagation (BP) neural network has favourable nonlinear mapping and high parallel information processing capabilities, and it is widely used in technology areas of multi-disciplinary subjects.Microwave drying technology, is different from the traditional drying methods, has characteristics of instantaneity, integrity, efficiency, safety, non-pollution, easy to realize the automatic control and improving the products quality, the direction of heat conduction is the same as the direction of moisture diffusion. But in the industrial process of microwave deep drying of selenium-rich slag, the influencing factors include the microwave input power, the acting time, the initial moisture content, the average material mass, the average material surface area and the rotational frequency. The affecting degrees of these factors are different, so that it can be guaranteed the testing cycle longer, the testing quantity larger and the parameters difficult to be optimized. The BP neural network, which has the ability of non-linear mapping, is chosen to build up the simulation model to predict the experimental process results.However, the traditional BP algorithm based on gradient descent method revises the network weights through computing the network weights’and thresholds’values of the objective functions, needs more convergent time and sometimes the convergent results can not be obtained in local minimum areas. Therefore, the BP neural network algorithm should be improved.In present thesis, the prediction model and the Smith compensation Proportion> Integration Differentiation (PID) control model based on incremental improved BP neural network are built up taking the industrial process of microwave deep drying of selenium-rich slag as the research target, and the universality of prediction model in the field of microwave calcination is researched. The main contributions are summarized as follows:1) The improved BP neural network based on Levenberg-Marquardt (L-M) algorithm overcomes these limitations. In the process of training the network, many training data probably are offered by the way of increment batch and the limitation of the system memory can make the training data infeasible when the sample scale is large, the incremental learning is implemented by adjusting the weights of the BP neural network as demonstrated.The incremental learning is realized by setting up the knowledge effective extent in single neural network and increasing the number of the nodes in hidden layers. When the new sample knowledge closes to the prior sample knowledge, the weights’ and thresholds’value can be changed in the effective extent, and the number of the nodes in hidden layers is fixed flexibly. In the process of training neural network using the training sample, the number of the nodes in hidden layers is increased and the output error is calculated. If the indication value approaches to the target value, the training is ending and the number of the nodes in hidden layers is the optimization. By using the method, the network can not only learn the new sample knowledge but sustain the original knowledge.The incremental improved BP neural network has the faster convergence, better prediction accuracy, better fitting results and can avoid the error sum squares no longer be updated, the phenomenon of network paralysis and the network not be trained, be out of the local minimum when adjusting the network parameters, make the network be converged rapidly, can effectively solve the problem such as the training data can not be provided for one-time, choose the representative samples to train the network in the case of occupying less memory source.2) The incremental improved BP neural network non-linear prediction model is built up during the industrial process of microwave deep drying of selenium-rich slag. Taking the microwave input power, the acting time, the initial moisture content, the average material mass, the average material surface area and the rotational frequency as input variables, the model is used to predict the experimental results.3) The incremental improved BP neural network prediction model for industrial electricity energy consumption is built up. Taking the microwave input power, the acting time, the initial moisture content, the final moisture content and the average material mass as input variables, the model is used to predict the experimental electricity energy consumption.4) The conventional PID control model is widely applied in the fields of industry process control because of its simple structure, robustness and easily operation, such as metallurgy, chemical engineering and electric power and mechanics. With the development of industry, the complexity of the controlled plant is serious, especially for the time-delaying, time-varying and non-liner complex systems, the conventional PID control model can not be met the control accuracy of the target.The industrial Smith compensation PID control model using incremental improved BP neural network is put forward. The parameters of controlled plant are evaluated in the experimental process and identified the neural network Smith compensation PID control off-line. Using the simplified control model of microwave deep drying selenium-rich slag to research the Smith compensation PID controller, the PID control parameters are tuned on-line.5) The industrial anti-prediction model using incremental improved BP neural network is proposed in microwave deep drying of selenium-rich slag processing. Taking the final material mass, the final temperature and the final relative dehydration rate as input variables, anti-prediction model is used to predict the acting time, the initial dehydration rate and the electricity energy consumption.6) The industrial prediction model of microwave deep drying of selenium-rich slag using incremental improved BP neural network is used for the microwave calcination of Ammonium Diuranate (ADU) and Ammonium Uranyl Carbonate (AUC) in order to investigate the universality of the prediction model. The model is used to predict the experimental processing results of microwave calcination.更多还原