Improving Class Separability of Frequency-domain Features for EMG Pattern Recognition