Development of AI framework for the prediction of future anomalous events

这笔催化剂赠款旨在多伦多大学和阿拉伯联合酋长国(阿联酋)之间建立伙伴关系，开发用于预测异常事件的人工智能(AI)框架。异常事件会严重影响健康、安全和经济状况。这些事件通常可以在许多应用中找到，包括自然灾害(例如地震)、生物医学工程(例如癫痫发作)、机器故障诊断(例如部件故障)或其他(例如房地产崩盘)。这些事件不仅很少发生，而且可能非常多样化和罕见。一般的有监督机器/深度学习方法可能不能直接应用于这个问题。处理这一问题的传统方法是，要么在这些事件发生后‘检测’它们，要么开发用于预测未来总体趋势的‘预测’模型。考虑到异常事件的后果，我们希望开发出人工智能模型，不仅可以在事件发生之前对其进行预测，还可以给出足够的准备时间来采取预防措施，减少后遗症。我们的主要假设是，在许多情况下，异常事件(异常前)之前的短暂时间段可能包含重要的信号，这些信号可能不同于正常数据的其余部分，表明其开始。因此，我们提出了一种新的深度学习框架，该框架可以利用正常数据或结合异常前数据来检测这些变化并预测关注事件的发生。加拿大团队可以访问由各种生理模式组成的两个大型多模式传感器数据集。与我们的阿联酋合作伙伴一起，除了信号处理和领域专业知识外，我们在构建分类和异常检测模型方面拥有独特的专业知识。该项目将动员双方研究人员相互访问，组织研讨会，集思广益，会见各自的高级领导层，以探索未来的合作和资金，发表研究论文，以及交换学生的机会。加拿大PI和我们的一名研究生将参观阿联酋的大数据分析中心。该项目将为加拿大方面的两个总部基地提供部分资金。

This Catalyst grant is aimed at building partnership between the Universities of Toronto and United Arab Emirate (UAE) on developing an artificial intelligence (AI) framework for predicting anomalous events. Anomalous events can severely impact the health, safety and economical situations. These events can be commonly found in many applications, including natural disasters (e.g., earthquakes), biomedical engineering (e.g., seizures), machine fault diagnosis (e.g., malfunction of components) or others (e.g., real estate meltdown). These events not only occur rarely but can be very diverse and infrequent. The generic supervised machine/deep learning approaches may not be directly applied to this problem. The traditional approach to handle this problem is to either 'detect' these events after their occurrence or develop 'forecasting' models that are meant for predicting general future trends. Considering the consequences of anomalous events, we want to develop AI models that can not only predict these events before they occur but also give sufficient lead time to take preventive measures and reduce the aftereffects. Our main hypothesis is that in many situations, a brief time period before an anomalous event (pre-anomalous) can contain important signatures that may differ from rest of the normal data, indicating its onset. Therefore, we are proposing a novel deep learning framework that can leverage either the normal data or combined with pre-anomalous data to detect those changes and predict the occurrence of events of concerns. The Canadian team has access to two large multimodal sensor datasets comprising of various physiological modalities. Along with our UAE collaborator, we have a unique blend of expertise in building classification and anomaly detection models besides signal processing and domain expertise. This project will mobilize both the investigators to visit each other, organize a workshop, brainstorm and meet respective senior leadership to explore future collaborations and funding, publish research papers, and student exchange opportunities. The Canadian PI and one of our graduate students will visit the Big Data Analytics Center at UAE. This project will partially fund two HQPs on the Canadian side.