Multi-Sensor Time-Series Classification for Activity Tracking Under Variable Length

Recent advances in mobile devices, wearable sensors and data analytics are enabling greater adoption of remote health monitoring systems, where human activity recognition (HAR) plays a key role. Most HAR systems assume that consecutive data segments are independent and aim at accurately classifying each individual data segment. However, because training and operating contexts are often different, and collecting labels can be time-consuming and expensive, the recognition accuracy is typically considerably lower in the new (operating) context. In this paper, we propose a framework for HAR that has several unique capabilities: 1) Turns predictions into useful information (e.g. daily distribution of activities for a user) that could be assessed by a clinician or health professional. Our goal is to mitigate the effects of variation between a training and an operating context. 2) Leverages the dependence between consecutive data samples to perform robust activity recognition. And 3) Performs activity classification for variable length set of activities. The framework we present encompasses two main modules. The first is a Recurrent Long Short-Term Memory deep neural network, the core component which gives us the capability to capture the dependence between consecutive sensor data samples and generate labels for every micro segment. The second is a post-processing step performed for uncertainty filtering and label smoothing of decisions. The filtering method enables the system to recognize activities in the presence of difference in the training and testing data distributions, and the smoothing method makes the system properly capture the distribution of activities. We demonstrate that our deep learning architecture can achieve 90% accuracy in standard training scenarios. Furthermore, we show that, in the case of inter-subject variation, the two methods allow us to still accurately capture the distribution of activities.