Although rough set theory has been widely applied to remove redundancy features of dataset, it just works successfully in nominal domian. While dealing with hybrid data, fuzzy rough model can be more suitable for feature reduction. This paper makes a comprehensive study on two kinds of state of art fuzzy rough reduction algorithms, and theoretically proves that compared to positive region based fuzzy rough reduct, information entropy based fuzzy rough reduct can achieve better decision performance in most cases. For further comparison, a modified version of a representative information entropy based fuzzy rough reduction algorithm is proposed. Experimental results on several UCI datasets and a real problem case of ultrasonic flaw signal classification show that our proposed algorithm outperforms the conventional fuzzy rough based and crisp rough set based feature reduction algorithms.