Improving Performance and Explainability of Sensor-Based Human Activity Recognition

Significant progress has been made in developing systems that can automatically recognize human activities, largely thanks to deep learning models and sensor data. However, high performance and comprehensibility are often at odds in these systems. The aim of this research is to enhance both the accuracy and understandability of models that use sensors for human activity recognition (HAR). We present a new convolutional neural network design which takes input from accelerometers and gyroscopes across multiple data streams. This allows efficient capture of time and frequency characteristics. Additionally, we put forward an attention module that identifies informative patterns in the sensor signals to shed light on the model’s predictions. Experiments were conducted on a benchmark dataset of common daily movements gathered from ubiquitous sensors. Our model achieves state-of-the-art accuracy exceeding prior deep learning techniques. The attention visuals provide transparency into which sensor segments are most relevant for accurate activity classification. This validates improved model interpretability. Our multi-channel architecture pushes sensor-based HAR forward through heightened predictive performance. We offer original perspectives to create reliable, precise systems for HAR.