【作者】 徐向阳；

【导师】 邵文柱；

【作者基本信息】 哈尔滨工业大学 ， 材料与化工（专业学位）， 2022， 硕士

【摘要】 在非结构化的自动化作业场景中,智能机器人需要与环境或目标对象交互以实时获取外界信息并对下一步行动做出决策。触觉是动物获取环境信息不可替代的方式之一,也能为智能机器人赋予材质探测和在深海、沙尘、深空等光线不足的场景中进行环境交互的能力。机器人触觉这一横跨了材料、算法和电子通信的领域蕴含了大量科学问题和技术难点,国内外专家正在不断地展开理论研究和技术开发。本课题围绕机器人触觉领域的材质识别问题进行研究,创新地使用有限元高通量模拟的方法分析材质识别任务中采集触觉信号的接触器最优外形,并批量计算和获取了触觉响应信号。分析了触觉数据中反映材质信息的特征,运用卷积神经网络对特征进行识别,并开发触觉识别模型的验证系统对卷积神经网络进行了验证。具体研究内容和相关工作如下:针对材质识别机器学习模型的样本数据来源问题,本文建立了用于采集触觉响应信号的有限元模型,并分析接触器的几何形状与弹性模量对触觉信号的影响,对有限元模型进行了优化迭代。同时将有限元模型的部分参数进行了随机化处理,编写成高通量模拟程序,批量计算并获取接触器触摸不同材料和表面形状物体时的触觉响应信号。此方法能够不受现实条件限制并且不受人为因素影响地获取准确的触觉响应信号。为了建立高性能的材质识别机器学习模型,本文建立了包括支持向量机、决策树、感知机、多层感知机和卷积神经网络的多种机器学习模型,经过对比分析,本文选择了卷积神经网络,然后对网络结构和样本特征作了针对性优化,并用样本扩增和人工数据集的方法扩充了样本集。经过优化后的卷积神经网络能够以97.5%的准确率识别材料为金属或橡胶、表面形状为点状或线条纹理的物体,在材质识别问题中表现出较好的性能。为了验证材质识别机器学习模型在现实接触场景中的识别性能,本文设计并搭建了触觉识别验证系统。系统的核心为树莓派开发板,信号采集单元为以6×6阵列排布薄膜式压力传感器。触觉响应信号于传感器中产生,经过模数转换芯片、数据选择芯片传入树莓派,树莓派中运行了该系统的驱动程序和机器学习模型,能够可视化地输出触觉响应信号和识别结果。结果显示,搭载了材质识别卷积神经网络的触觉识别验证系统能以93.75%的准确率识别物体材质。更多还原

【Abstract】 In unstructured automated job scenarios,intelligent robots need to get real-time external information by interacting with the environment or target objects to make decisions about next actions.For animals,tactile perception is an irreplaceable way to get environmental information,and it also gives intelligent robots the ability to detect materials and interact with the environment in low-light scenarios such as deep ocean,sand,and deep sky.Robot tactile perception,which spans the fields of materials,algorithms and signal communication,contains a large number of scientific problems and technical difficulties.Experts at home and abroad are continually developing theoretical research and technical development.This topic focuses on the problem of material and texture recognition in the field of robot tactile perception.The optimal shape of the contactor that collects tactile signals in the task of recognition is analyzed innovatively by using the finite element highthroughput simulation method,and the tactile response signal is calculated and obtained in batch.The features reflecting material information in tactile data are analyzed and the convolution neural network is used to recognize the features.And a verification system for the tactile recognition model is developed to verify the accuracy of convolution neural network.The specific research content are as follows:To solve the problem of sample data source for material and texture recognition machine learning model,this thesis establishes a finite element model for collecting tactile response signal.To optimize this finite element model,this thesis then analyses the influence of geometry and elastic modulus of the contactor on the tactile signal.At the same time,some parameters of the finite element model are randomized,and a highthroughput simulation script is programmed to calculate and obtain the tactile signal of the contactor when it touches objects of different materials and surface shapes.This method can obtain accurate tactile response signals without restriction of real-world conditions and influence of human factors.In order to establish a high-performance machine learning model for material and texture recognition,this thesis establishes a variety of machine learning models including support vector machine,decision tree,perceptron,multi-layer perceptron and convolutional neural network.After co MParative analysis,this thesis selects convolutional neural network to recognize,then optimizes the network structure and sample feature,and expands the sample set by using the methods of sample amplification and artificial data set.The optimized convolutional neural network can recognize objects whose material is metal or rubber and whose surface shape is point or line texture with97.5% accuracy,so the model has good performance in material recognition.In order to verify the recognition performance of material recognition machine learning model in real contact scenes,a tactile recognition verification system is designed and built in this thesis.The core of the system is the Raspberry Pi development board,and the signal acquisition unit is membrane pressure sensor which is arranged in 6×6array.The tactile response signal is generated in the sensor and transmitted to Raspberry Pi through analog-to-digital conversion chip and data selection chip.The driver and machine learning model of the system are run in Raspberry Pi,which can visually output the tactile response signal and recognition results.The result shows that the tactile recognition verification system equipped with material recognition convolution neural network can recognize object materials and texture with accuracy of 93.75%.更多还原