Improving fault tolerance of wearable wearable sensor-based activity recognition techniques

• Authors: Ryoma Uchida;Hiroto Horino;Ren Ohmura
• Venue: Proceedings of the 2013 ACM conference on Pervasive and ubiquitous computing adjunct publication
• Year: 2013

Existing wearable sensor-based activity recognition techniques lack fault tolerance in the case of sensors data loss, such as communication disconnection and sensor failure. Compensating for missing data is one method to improve robustness and can be done by three levels in activity recognition: raw data level, feature value level, and classifier level. Our study proposes a method to compensate for the missing sensor data using an ARAR algorithm and compares this method with a previous method for compensating for the feature value using kernel regression in the feature value level. The ARAR algorithm method predicts future data from existing sequence data. We conducted some experiments to verify the usefulness of the proposed methods. Specifically, the prediction performance was evaluated by applying the ARAR algorithm to compensate for one to five successive windows. As a result of our test data, the F-measure rate was 73.4% in the case of sensor data loss. The ARAR algorithm compensation for one and two successive windows increased the F-measure to 76.8%. Overall, the ARAR algorithm method effectively compensates for instantaneous communication disconnection. On the other hand, the kernel regression method is especially compensates for burst communication disconnection. Therefore, we need to change the compensation method depending on sensor error patterns. Thus, we improved robustness of the activity recognition system by compensating for sensor data loss.