Abstract: Fault features set can be acquired by processing acquired data in machine diagnosis. Unfortunately many of these features are either partially or completely irrelevant or redundant to the fault state usually exist in the fault feature set. These features often decrease diagnosis accuracy and reduce the computational efficiency. Feature selection can remove those redundant features to avoid the influence. A new feature selection method on the basis of feature sensitivity analysis as selection evaluation is designed for support vector machine (SVM). Feature sensitivity is defined as relative feature importance for the classification decision on the basis of a single SVM training run. And genetic algorithm is used for select optimization in the method. According to results of simulated data and diesel engine fault feature selection experiment, it is proved that this method possesses excellent optimization feature selection property, and acquires higher accuracy.

Key words: Genetic algorithm, Feature selection, Support vector machine (SVM) Sensitivity analysis