A Noise-Robust Self-supervised Pre-training Model Based Speech Representation Learning for Automatic Speech Recognition.