A case study: effect of wrist sensor displacement on HAR performance using LSTM and attention mechanism

Loose wearing or self-placement usually causes sensor displacement, which can deteriorate the performance of classifiers in real use. As a case study, this paper focuses on investigating the effect of wrist-worn sensor displacement on human activity recognition. We construct a new HAR dataset from different positions of the wrist. We create a LSTM model and an multi-stage attention model for the evaluation of our three designed scenarios. Experimental results show that the classification accuracies are affected by sensor positions and the worst performance occurs when test data are from a new position for a model. In addition, the results also indicate the superior performance of the attention model on all the scenarios compared with the LSTM model.