Diagnosis, Learning and Optimization of Smart and Connected Products

点击翻译按钮获取中文摘要

Today's smart product transmit real-time sensor data to a control centre that can monitor its condition, perform remote diagnosis, and adjust its configuration in real time to improve performance. Such products have revolutionized many industries, transforming traditional manufacturers into service providers. Tesla, for example, monitors its cars for malfunction via in-car sensors. The centre then advises when service is needed. Many issues can be rectified via “over-the-air" software updates. If on-site service is needed, the centre dispatches a service team with the right spare parts. Car configuration can also be adjusted via software so that the car can adapt to user behaviour to improve safety and efficiency. However, the imperfection of sensor data makes the operation of smart products complex. Many failure modes may share similar features and be hard to distinguish. To learn more about a product's true condition, the control centre can wait and monitor the product longer, but this invariably introduces delay. Faced with tension between learning and intervention, the centre must decide when to stop monitoring and what action to take in a timely and accurate fashion. This trade-off between learning and doing lies at the heart of the operations of smart and connected products. Existing models for the control of sensor-embedded systems fall mostly under the framework of partially observable Markov decision processes. It is known that multi-state problems suffer from the curse of dimensionality in dynamic programming and hence the optimal policy is difficult, if not impossible, to compute, and thus most existing work considers simple systems with binary states. A network of connected products requires one to generalize the optimization to multi-state systems, which is now possible due to recent advances in artificial intelligence that compresses the state space. Additionally, existing work often assumes model parameters are known, which is not always the case in reality. The proposed work will use a Bayes-adaptive approach to jointly learn model parameters and hidden system state while making sequential decisions. This research program will: develop the optimal monitoring and diagnostic algorithm for smart products by leveraging the full potential of imperfect sensor data; gain insight into how to optimally fine-tune a product in real time to maximize the performance; and develop a new framework that harnesses connectivity among products to aggregate and share information to improve efficiency of maintenance and operations. Research outcomes are important for manufacturers in the Internet of Things era and for academic researchers in industrial engineering and operations research. These topics also provide a strong opportunity to train HQP in data analytics and real-time optimization.