轮式云服务机器人健康状态自主认知模型及其进化机制研究

The research on the independent health cognition model and its evolutionary mechanism of service robot is the key to guarantee the long-term safe and reliable operation of the robot. On the one hand, the traditional robot health assessment method does not consider the influence of dynamic working environment and service task on its health assessment. On the other hand, the limited hardware resources of a robot can hardly satisfy the computing and storage need of complex health assessment algorithm. Therefore, the study of this problem faces great challenges. This proposal presents a creative research plan by combining big data, cloud computing and deep learning technique. Firstly, the semantic knowledge mining method of service robot health context in complex situation is studied and the context semantic description of unsteady factors such as operating environment and service task is established, which could reduce the influence on the accuracy and robustness of the robot's health evaluation. Secondly, with the help of robot cloud platform and attention focused mechanism, a cooperative self-adaptation spatial-temporal feature extraction method based on deep neural network and multilevel fusion strategy is proposed. The separation problem of the temporal and spatial relationship of dynamic multimode health data is solved by the proposed feature extraction method. Lastly, the evolutionary mechanism of DNN model is learned, which could realize the incremental learning of health knowledge and the efficient compression of cognitive model. This project will establish the basic theories and methods of independent healthy cognition and evolution of wheeled cloud service robot. Besides, this project will provide important theoretical foundation and key technical support for the popularization and application of service robots.

对服务机器人健康状态自主认知模型及其进化机制进行研究是保证机器人长期安全可靠运行的关键。传统的健康评估方法未考虑服务环境的动态变化和服务任务的多样性，同时服务机器人本身计算和存储资源也非常有限，使得该问题研究面临极大的挑战。本项目另辟蹊径，基于大数据、云计算及深度学习技术，首先研究复杂工况下服务机器人健康上下文语义知识挖掘方法，实现运行环境及服务任务等非稳态因素的上下文语义描述，降低其对机器人健康评估精度和鲁棒性的影响；然后，借助于服务机器人云平台，研究注意力集中机制下基于深度神经网络和多级融合策略的深层时空特征协同自适应提取方法，突破动态多模时序健康数据深层特征时空关系常被割裂的难题；最后，研究健康认知深度神经网络模型进化方法，实现健康知识的增量学习及认知模型的高效压缩。初步建立起轮式云服务机器人自主健康认知与进化的基本理论与方法，为服务机器人的推广应用提供重要理论基础和关键技术支撑。

安全性和可靠性低是限制当前服务机器人应用的关键因素之一，服务机器人健康状态自主认知与智能故障诊断方法研究是提升其安全可靠性的有效手段。服务机器人工作环境和服务任务的多样性以及有限的机器人本体计算和存储资源，使得该问题研究面临极大的挑战。本项目围绕机器人健康状态自主认知过程中的关键科学问题展开了系统深入的研究。基于迁移学习结合胶囊神经网络等方法，研究了复杂噪声工况下服务机器人健康状态知识挖掘模型，构建了非稳态因素的语义模型，实现了变工况复杂环境下机器人机械子系统关键部件的故障诊断与健康状态识别；针对传感器数据时空特征协同提取困难，利用卷积神经网络、稀疏自编码机以及图神经网络等技术，提出了多通道注意力机制下的时空特征协同自适应提取方法；基于麦克纳姆轮服务机器人以及机器人云平台搭建了实验验证系统，利用神经网络剪枝、群智能优化算法等方法研究了神经网络模型的优化与压缩问题，揭示了非关键信息的遗忘和健康评估模型的云端进化机制。通过本项目的研究，项目团队共发表SCI论文10篇，EI论文4篇，授权发明专利4项，申请软件著作权3项。培养或联合培养2人获得博士学位，8人获得硕士学位，课题组在机器人智能故障诊断与健康状态评估方向组建了一支可持续的研究力量，目前，联合培养在读博士1名，硕士生5名。本项目研究的相关理论和方法将为服务机器人健康状态自主认知与进化、智能故障自诊断、时空特征融合等研究提供新的思路，具有重要的理论意义和实用价值。