Improving Noisy Student Training on Non-target Domain Data for Automatic Speech Recognition