Electrical Engineering Department, University of California, Los Angeles, Los Angeles, CA 90095, USA.
IEEE Trans Biomed Circuits Syst. 2012 Apr;6(2):167-78. doi: 10.1109/TBCAS.2011.2166073.
Energy efficiency has been a longstanding design challenge for wearable sensor systems. It is especially crucial in continuous subject state monitoring due to the ongoing need for compact sizes and better sensors. This paper presents an energy-efficient classification algorithm, based on partially observable Markov decision process (POMDP). In every time step, POMDP dynamically selects sensors for classification via a sensor selection policy. The sensor selection problem is formalized as an optimization problem, where the objective is to minimize misclassification cost given some energy budget. State transitions are modeled as a hidden Markov model (HMM), and the corresponding sensor selection policy is represented using a finite-state controller (FSC). To evaluate this framework, sensor data were collected from multiple subjects in their free-living conditions. Relative accuracies and energy reductions from the proposed method are compared against naïve Bayes (always-on) and simple random strategies to validate the relative performance of the algorithm. When the objective is to maintain the same classification accuracy, significant energy reduction is achieved.
能量效率一直是可穿戴传感器系统的一个长期设计挑战。由于需要不断缩小尺寸和提高传感器性能，在持续的主体状态监测中，这一点尤为关键。本文提出了一种基于部分可观察马尔可夫决策过程（POMDP）的能量高效分类算法。在每个时间步，POMDP 通过传感器选择策略动态选择用于分类的传感器。传感器选择问题被形式化为一个优化问题，其目标是在给定能量预算的情况下最小化分类错误成本。状态转移建模为隐马尔可夫模型（HMM），相应的传感器选择策略使用有限状态控制器（FSC）表示。为了评估这个框架，从多个主体的自由生活条件下收集了传感器数据。与朴素贝叶斯（始终开启）和简单随机策略相比，提出的方法在相对准确性和能量减少方面的性能得到了验证，以验证算法的相对性能。当目标是保持相同的分类准确性时，可以显著减少能量消耗。
