Improving Accuracy in Fatigue Detection/Prediction by Dynamic Weighted Moving Average of Heart-Rate Variabilities

Instead of only real-time detections of driver fatigue based on the driver's heart rate variability (HRV) data, this work targets at early and highly accurate predictions of driver fatigue. A back-propagation neural network (BPNN) model is used to predict and a novel Dynamic Weighted Moving Average (DWMA) is used to increase accuracy and thus make predictions early enough. In DWMA, we propose modeling of data quality, designing of weights using data quality, and using an exponential error model to dynamically reduce prediction error propagations. The proposed driver fatigue monitoring system design is realized with a BPNN prediction model and DWMA technique, which results in not only predictions being performed by up to 15 to 17 minutes before being actually fatigued, but accuracy of predictions is also improved by 11% compared to the traditional WMA. For detections, BPNN/DWMA can achieve 95% accuracy and for predictions in the first five future time slots, an accuracy of 90% can be achieved, with a True Positive Rate of at least 80%. When DWMA is applied with other prediction methods such as ARIMA, predictions are increased by as much as 15% for the first three future time slots, which shows that DWMA is a general technique that can be used in combination with different prediction models.