【作者】 吕洪波；

【导师】 贾培发；

【作者基本信息】 清华大学 ， 计算机科学与技术， 2011， 博士

【摘要】 机器人修磨的特点是面向复杂曲面，且需精确控制磨削量。由于砂带和接触轮具有弹性，且串联工业机器人刚性不好，所以是一种非刚性加工，磨削量受多种因素的影响，其中动态因素不能测控是系统实现难题。现有系统一般基于经验公式对磨削量进行控制，根据实验数据对动态因素进行补偿，精度低，适应能力差。人工智能方法是解决这一难题的有效手段。本文围绕着应用智能方法解决机器人修磨系统磨削量的精确控制问题而展开，包括方法框架、适应建模方法和应用实现三个层面。主要工作成果和创新点如下：1.面向动态因素的适应建模问题提出机器人智能修磨方法框架。该框架的关键点包括知识库、在位测量、适应建模与磨削参数选择。在位测量把新样本及新模型反馈至知识库，使基于知识库的建模方法具有适应动态变化的基础。2.提出基于机器学习的稳态修磨过程建模方法。该方法通过对砂带磨损的稳态近似获得满足独立同分布条件的样本，并用于训练基于支持向量回归和回声状态网络的学习机模型，精度满足系统要求。同时，该方法解决了知识库中数据和模型的组织问题，使之可用于适应建模。3.提出基于迁移学习的适应建模方法。该方法利用新样本去量化新旧模型的相似度，在新旧模型单调性一致的前提下，通过对相似度最高的旧模型的迁移来实现新目标函数的近似估计。局部迁移方法中，面向磨削量的归一化欧氏距离解决了相邻点对待求点迁移量的加权问题。当新样本在输入变量作用域内分布均匀时该方法具有较高精度。4.提出融合先验知识的适应学习建模方法。该方法通过虚拟样本的形式把半经验公式中蕴含的信息融合到学习机中，解决了新样本点分布不均时，新旧模型间不能准确量化迁移量而导致的无法适应建模的问题。虚拟样本保证了学习机的稳定性，机器学习方法提高了建模精度。5.基于粒子群算法实现修磨参数的优化。参数选择是实现智能修磨的关键步骤，多变量在线可控使参数可优化，但变量在线调节速度影响磨削量控制的准确度。提出针对不同变量进行加权的相邻点差值平方和的优化目标，并基于粒子群算法实现参数优化，有效提高了系统响应速度。更多还原

【Abstract】 The robotic profile belt grinding (RPBG) system is often employed for machiningcomplex surface whose removal rates need to be controlled accurately. Because theabrasive belt and the contact wheel are flexible, and the industrial robot is not rigidenough, RPBG is non-rigid and its removal rate is affected by a large amount of factors.So far, empirical formulas are applied to calculate the removal rate in the RPBGsystems mostly. However, they are low accuracy and lack of adaptability. Artificialintelligence is effective for solving this problem. The central topic of this thesis is tocontrol the accuracy of the RPBG system using artificial intelligence technology. Itincludes the method framework, the methods of adaptive modeling and how to put theminto use. The contributions of this dissertation consist of the following parts:1. To solve the problem of adapting to dynamic factors, an intelligent methodframework is presented. The keys of the framework include the knowledge database, in-situ measurement, adaptive modeling and parameters’ selection. The in-situmeasurement gets some new samples for the knowledge database in each grinding. So itis possible for the model to adapt the dynamic factors.2. In order to solve the adaptive problem, methods based on support vectorregression and echo state network are firstly applied to model the removal rate on quasi-steady state which means meeting the i.i.d. condition approximatively in this thesis.And the accuracies of these two methods meet the need. Basing on the quasi-steadylearners, samples in the database can be organized and the adaptive learning problemcan be formulized.3. An adaptive modeling method based on transfer learning is proposed for theRPBG system. This method uses some new samples to measure the difference betweenthe old learner and the new target function. With same monotonicity, the most similarold learner is transferred to predict the target points. The normalized Euclidean distancefor removal rate is important for calculating the weights. When the new samplesdistribute evenly in the input space, this method is effective.4. Another adaptive method is put forward by incorporating prior knowledge inmachine learning by creating virtual examples based on semi-empirical formula. The method can be applied for adaptive modeling when the new samples in input space areunevenly. In principle, the virtual samples make the model stable and the machinelearning algorithm improve the accurate.5. It is a key step to choose the grinding parameters in the intelligent framework,which is also the purpose of the removal model. And there are more than one variablecan be adjusted online, so the grinding parameters can be optimized. The objectfunction is presented in order to reduce impact of the system response speed on thegrinding accuracy. The particle swarm optimization algorithm is applied to gain theoptimal parameters.更多还原