基於智慧環境之多層情境探勘識別與適應

Abstract

物聯網科技的普及與流行，使得低運算能力的裝置在嵌入感測器和控制器後，可以被大量地佈置在智慧環境中，同時結合在使用者的身上。這些裝置所產生的大資料可被進一步收集，並於以人為本的系統中被應用。在如此多的應用系統中，其中之一的代表便是智慧家庭。 在智慧家庭空間中，系統應用利用機器學習或資料勘探方式對環境中的大資料進行分析，並依照分析結果以辨認使用者的活動及對應的服務。但是，隨著智慧家庭中所安裝感測器的種類增加，以及可穿戴式設備的大規模使用，不同類型的感測器所表示的情境可能會互相影響，以致降低辨認活動的準確度。此外，隨著資料量不斷成長，非監督式學習中標記活動的成本越發之高。而在實際應用中，如何在僅標有一部分資料的情況下，使用已建立的活動模型辨識未出現過的新活動，通過與用戶的詢問交互以完善原有的模型，這也是一個挑戰。因此，我們提出多層次的非監督式學習分析架構用於在智慧空間中的情境探勘與未知情境之辨識；同時使用主動式學習與異常檢測，使系統適應新增情境以提升辨識的準確性。 本研究的主要貢獻有以下幾點：第一，為有效降低不同類型感測器資料之間可能產生的影響矛盾，我們提出了一個基於非監督式學習演算法的多層次結構的模型以探索家中使用者的活動。第二，我們基於多層次結構的模型檢測潛在的全新活動，同時辨識已有活動並反饋至使用者。第三，我們提出同時考量情境數量及所佔比的異常檢測機制，使系統適應新增情境更準確。

As we are moving towards the time of Internet of Things(IoT), a significant number of heterogeneous low-power sensors and actuators are deployed around the world. The data from them, which we called Big Data, are generated for further analy-sis in human-centric applications in IoT. One of the applications in IoT is in Smart Home. In smart home environment, the application provides context-aware services and context discovery after recognizing user activities through obtained sensory data. However, there are still some challenges in activity context discovery and recognizing. With the sensors on large number of types and places to locate, the negative effects from heterogeneous features will arise, and may lead to confusing results. Besides, it is difficult to label the large quantity of sensor data quickly generated. How the appli-cation recognizes existing activities and discovers potential ones within only part of sensor data is still challenging. So, we propose an unsupervised context discovery and recognition based on hierarchical fusion of heterogeneous features in real smart envi-ronment. Also, the active learning with feedbacks presents user’s activity and reinforce the existing models. There are three major contributions in this thesis: Firstly, a hierarchical structure of models of contexts from unsupervised algorithms is proposed to discover the poten-tial contexts within heterogeneous sensors. Secondly, we propose an online context recognizing to check whether it is unseen context hierarchically. Thirdly, the active learning as well as outlier detection with percentage and frequency of context help adapt new unseen context into recognition instead of random outlier.