Reliable and power-efficient machine learning in wearable sensors

Wearable inertial sensors are widely used to monitor human movements. An important application of these sensors is to estimate metabolic equivalent of task (MET) values associated with physical activities. However, a major obstacle in widespread adoption of such technologies is that any changes in the configuration of the network requires new data collection and retraining of the underlying machine learning algorithms that are designed to estimate MET. Examples of such changes include sensor misplacement and sensor addition. Sensor misplacement refers to changes in wearing site or on-body location of the sensor during its operation. Sensor addition indicates expansion of the wearable network to accommodate new inertial sensor nodes. This chapter introduces algorithms that address the problems of sensor misplacement and sensor addition with consideration of power-efficiency in networked wearable systems. First, a machine learning approach is presented for power-efficient sensor localization to compensate for sensor misplacements by automatically identifying the wearing site of the wearable sensor. Second, a transfer learning algorithm is introduced to adopt the knowledge of existing sensors in the wearable network for autonomous training of a MET estimation algorithm in a newly added sensor. The combined algorithms discussed in this chapter provide a reliable, power-efficient, and reconfigurable MET estimation system for use with wearable sensors. This chapter also presents an evaluation of the described algorithms using real-data collected in two experiments involving both daily physical activities and fitness movements. The chapter concludes with a discussion of the results and insights into future directions.