A two-stage data quality improvement strategy for deep neural networks in fault severity estimation