Intelligent Assembly Action Recognition for Next Generation Manufacturing

制造业是加拿大最重要的经济部门之一。然而，由于制造环境的高度动态化，大多数生产过程都是由人类操作员完成的，因为机器人仍然无法复制人类的认知和灵活性。为了实现所需的质量控制并减少生产时间和成本，下一代制造非常需要可靠的人类行为识别系统。该项目建议开发新的机器学习算法，这些算法可以处理未标记数据或有限的训练数据，同时获得与监督学习模型相竞争的甚至更好的性能。具体地说，我们建议研究(I)自我监督学习策略；(Ii)多模式学习模型；(Iii)高效的在线动作识别管道；以及(Iv)终身学习范式，用于使用未标记数据或有限的训练数据进行动作检测和识别，同时获得与监督学习模型竞争甚至更好的性能。该项目将与I-5o合作，I-5o是一家使用人工智能视觉系统进行制造的技术创新者。除现金捐助外，I-5O还承诺为该项目提供大量实物捐助。该项目直接与I-5O在人工智能领域的战略计划以及改进其算法以便于部署和实施的活动保持一致。提出的解决方案在技术上是新颖的，并专门为解决装配动作识别问题而设计。所提出的解决方案将显著降低模型对训练数据的依赖，提高其装配动作识别的效率和准确性。从而大大降低了数据标注的成本，缩短了模型的训练和开发过程。该项目开发的系统将帮助加拿大制造商通过对其生产运营的准确实时监控来提高劳动效率，并保持在全球制造业格局中的竞争力。

Manufacturing is one of Canada's most important economic sectors. However, due to the highly dynamic nature of the manufacturing environment, most production processes are done by human operators as robots still cannot replicate human cognition and dexterity. To achieve desired quality control and reduce production time and cost, a reliable human action recognition system is highly desired for next-generation manufacturing. The project proposes to develop new machine learning algorithms that can work with unlabelled data or limited training data while achieving competing or even better performance than supervised learning models. Specifically, we propose to investigate (i) a self-supervised learning strategy; (ii) a multimodal learning model; (iii) an efficient online action recognition pipeline; and (iv) a lifelong learning paradigm for action detection and recognition with unlabelled data or limited training data while achieving competing or even better performance than supervised learning models.The project will collaborate with i-5O, a technology innovator for manufacturing using AI-powered vision systems. In addition to the cash contribution, i-5O has committed substantial in-kind contributions to this project. This project directly aligns with i-5O's strategic plan in AI and activities on improving their algorithms for ease of deployment and implementation. The proposed solution is technically novel and specifically designed to solve the problem of assembly action recognition. The proposed solutions will significantly reduce the model's dependency on training data and increase its efficiency and accuracy in assembly action recognition. Thus, it will greatly reduce the cost of data annotation and shorten the process of model training and development. The system developed in this project will help Canadian manufacturers improve their labor efficiency through accurate real-time monitoring of their production operations and stay competitive in the global manufacturing landscape.