多功能仿人手指传感器建模、制备及触觉图像识别

Tactile sensor is the key way to perceive the environment for robot that can help the aged and the physically disabled. However, the traditional tactile sensors have the disadvantages of single function, low integration and large size. Therefore a new humanoid finger sensor is proposed under the methods of nonlinear modeling, manufacturing and tactile image recognition, which adopts liquid conduction and is able to sense three-dimensional force, temperature and micro-vibration. In order to explore the perception mechanism and physical characterization of the finger sensor, three scientific problems need to be studied in this project. Firstly, according to the research for the propagation mechanism of electric, heat and vibration in liquid medium, the nonlinear models would be established to reveal the relationship between the physical properties of liquid medium and the conduction of electric, heat and vibration, respectively. Secondly, the high reliable manufacture technology would be developed with structure and circuit integration, while it needs to solve those problems such as size and layout constraints, liquid leakage, etc., which can improve the reliability of the sensor and meet the needs of commercialized application. Finally, facing the problem of large and complex output signals of the finger sensor, a fast-tactile-recognition method based on deep learning theory would be proposed to improve the robustness of the learning model, reduce the effect of noise signals interference, and to achieve the intelligent recognition and classification for tactile features. This research provides a theoretical basis for the multi-physical properties-measured sensor and its signal processing technology, which has important significance to the development and application of sensing technology for intelligent robots.

触觉传感器是助老助残机器人感知环境的关键途径，而传统触觉传感器功能单一、集成化程度低、尺寸大。为此提出基于液体介质传导，可感知三维力、温度、微振动等物理特性的新型仿人手指传感器的非线性建模、制备及触觉图像识别方法。为了探索手指传感器的感知机理及性能表征规律，本项目对三个科学问题进行研究。首先，研究电、热以及振动在介质中的传播机理，建立介质的物理属性与电、热和振动传导的非线性关系模型。其次，研究传感器结构与电路一体化的高可靠制备工艺，克服敏感元件结构尺寸和电路布局的约束以及易漏液等问题，提高传感器的使用可靠性，便于商业化应用；最后，针对传感器输出信号多且复杂的问题，提出基于深度学习理论的快速触觉图像识别方法，提高学习模型的鲁棒性、减小噪声信号干扰的影响，实现触觉特征的智能识别与分类。本研究为集成多物理量测量的柔性传感器及信号处理技术提供理论依据，对智能机器人传感技术的发展和应用具有重要意义。

触觉传感器是服务操作机器人感知物体的关键途径之一，而传统触觉传感器功能单一、集成化程度低，存在操作不灵活、环境认知能力差等问题，为此提出了基于液体介质传导，可感知三维力、温度、微振动等物理特性的新型仿人手指传感器，并开展了手指传感器介质传导特性的非线性建模、手指传感器结构优化及可靠性制备、基于深度学习的触觉图像识别算法及试验验证等研究。构建了传感器的电、热、微振动在导电液体介质中传播、传导的非线性模型，建立了导电液体的湍流模型，采用多因素正交试验方法，对流体的粘度、管径、管道弧度、内部压力等关键参数进行了优化设计；采用集成度高且功能强大的MCU芯片、小封装元器件、小型连接器件等，实现了手指传感器内嵌控制单元的小型化；开发了基于逆模型的温度预估补偿算法，实现了对温度滞后的有效补偿；采用真空模具浇注的方法完成了手指皮肤的制备，采用3D打印实现了指骨和指甲盖等零件的加工，完成了手指传感器组件的集成装配，包络尺寸不超过38mm×20mm×7mm；研究了基于压阻效应的柔性触觉点元设计和感知方法，开发了空间分辨率为4×4的柔性阵列触觉传感器；研制了多自由度标定试验台，通过控制目标物体与传感器的相对运动，获得稳定的输入信号，可快速、准确地标定手指传感器的多类型（温度、压力、微振动等）采集数据；基于CAFFE的深度学习工具构建了CNN网络的学习训练模型，从仿人手指传感器的触觉图像中提取出了物体硬度、热传导、表面凸凹感、纹理等触觉特征，实现了物体分类的识别，识别的准确率、精确率和召回率均大于95%；针对柔性触觉传感阵列，提出了基于CNN卷积神经网络和LSTM长短期记忆神经网络融合模型的触觉序列识别方法，通过Python的深度学习库Keras对触觉图像进行训练，达到了90%以上的识别正确率。.本项目研究揭示了手指传感器介质属性对传导电、热、振动的影响关系，构建了多维触觉图像的智能识别算法，研发了仿人手指传感器样机并进行测试验证，解决了手指传感器输出信号繁杂、特征不明确、耦合干扰大的问题；获得了狭小空间约束下机器人多源信息融合感知的具体应用，在航天在轨服务与维护、工业自动化分拣、助老助残服务等领域具有巨大的市场应用前景。