An Adaptive Sensor Mining Framework for Pervasive Computing Applications

Abstract

Considering the wealth of sensor data in pervasive computing applications, mining sequences of sensor events brings unique challenges to the KDD community. The challenge is heightened when the underlying data source is dynamic and the patterns change. In this work, we introduce a new adaptive mining framework that detects patterns in sensor data, and more importantly, adapts to the changes in the underlying model. In our framework, the frequent and periodic patterns of data are first discovered by the Frequent and Periodic Pattern Miner (FPPM) algorithm; and then any changes in the discovered patterns over the lifetime of the system are discovered by the Pattern Adaptation Miner (PAM) algorithm, in order to adapt to the changing environment. This framework also captures vital context information present in pervasive computing applications, such as the startup triggers and temporal information. In this paper, we present a description of our mining framework and validate the approach using data collected in the CASAS smart home testbed. Categories and Subject Descriptors General TermsAlgorithms, Design, Experimentation, Human Factors. KeywordsSensor data mining, Sequential mining, Pervasive computing applications, Smart environments, Adaptation. INTODUCTIONWith remarkable recent progress in computing power, networking equipment, sensors, and various data mining methods, we are steadily moving towards ubiquitous and pervasive computing, into a world entangled with abundant sensors and actuators. As a result, there is a wealth of sensor data that can be analyzed with the goal of identifying interesting patterns. For example, by discovering repetitive sequences (frequent or periodic), modeling their temporal constraints and learning their expected utilities, we can intelligently automate a variety of tasks such as repetitive daily tasks in homes, or assembly sequences in manufacturing floors. Mining sequences of sensor events brings unique challenges to the KDD community, and the challenge is even heightened more when the underlying data source is dynamic and the patterns change.In this work, we introduce a new adaptive mining framework for use with pervasive computing applications, which will detect and adapt to changing patterns in the sensor data. These patterns can be expressed as a set of time ordered sequences, and discovering and adapting to the changes in such sequences can be achieved by using a sequence mining algorithm tailored to the special domain requirements of pervasive computing applications. The pervasive computing special requirements include utilizing context information such as startup triggers and temporal information, a unified framework for discovering periodic and frequent patterns, and most importantly adaptation over the lifetime of the system. Startup triggers are events that can trigger another action, e.g. a person entering a room can act as a potential startup trigger for the light in the room to be turned on. Triggers, which are absent in most traditional data mining methods, ...

Activity knowledge transfer in smart environments

“…), and more complex activities occurring in different locations can be composed of those basic activities. Notice that as we only have access to limited input data (perhaps a few days or even a few hours), we cannot use conventional activity discovery methods such as frequent or periodic sequence mining methods [40] to find activity patterns in the data. Therefore, exploiting the spatial closure can be a way to overcome this problem.…”

Designing Lightweight Software Architectures for Smart Environments

“…Here we also suggest that middleware can be evaluated in light of these same requirements and offer an evaluation of the CLM middleware along these performance dimensions. We evaluate our CLM middleware via our implementation that supports the CASAS smart home environment [9]. The goal of the CASAS project is to design smart environments that act as intelligent agents, perceiving the state of the environment using sensors and acting on the environment to achieve user or project goals [10].…”

Action Prediction in Smart Home Based on Reinforcement Learning

“…Machine learning techniques could be offline where created patterns are fixed like in [9,10] or online learning where patterns are dynamic and may change over time due to environment dynamicity and inhabitant's behavior changes. Few works targeted the online learning like in [11]; they detect the changes in the behavior of the inhabitants from the user himself or by a smart detection method. Our work differs by its online and continuous learning.…”