Forecasting ISP for Rail Transit Drivers Based on Heart Rate Variability with Wearable Devices

Inappropriate stopping at platform (ISP) of metro train contributes to transit delays and may lead to dangerous accidents. Since the driving performance is highly related to the physiological responses of drivers, ECG metrics are selected to forecast ISPs of rail transit. There are two types of ISP, namely, platform overrun (PO) and platform early stop (PES). RR intervals of drivers during the 100 s driving before the stop are explored to unveil the differences: (1) between ISP and no-ISP (appropriate stopping at platform); and (2) among non-ISP, PO, and PES. To warn the driver of the ISP, the study selects two warning points that are 8th and 21st second prior to the stop according to the sum of the RR variation per second, respectively. Furthermore, the study extracts the heart rate variability (HRV) features from RR intervals recorded during the period prior to the warning point via either 30 s or 60 s time windows. Then, the study constructs a convolution neural network (CNN) with attention mechanism and several machine learning models to predict ISP/no-ISP and non-ISP/PES/PO before the train stops at platform. The results show that both RR intervals and most HRV features are significantly different among non-ISP, PES, and PO. The importance of HRV features is inconsistent if the HRV features are extracted by different time windows and warning points. The prediction results show that CNN models have a better performance than traditional machine learning models in predicting both binary and three-class categories. Introducing attention mechanism increases the model performance. Also, HRV features extracted by 60 s time windows prior to the 8th second warning point provides the best results in predicting binary and three-class categories.