Feature selection is one of the most important aspects of condition monitoring. The condition or health of a machinery is manifested in the form of changes to its signal features. Traditional approaches for feature selection involve collecting an exhaustive list of features from different domains, all of which may not be significant. The objective then is to find a subset of features that reveal the changes in a significant way. The usual approach to select important features is by applying thresholding to loading scores of variables obtained from principal component analysis (PCA). However, there is not enough study as to the effectiveness of selected features in classifying multiple faults. In this paper, we show that by applying sparse principal component analysis (SPCA) to wavelet packet features, we obtain a subset of features that are interpretable and classification accuracy obtained in multiclass classification by using only those features is encouraging. This classification accuracy is compared to that of features obtained from conventional PCA by taking top four features with maximum absolute loading scores. It is observed that, in the particular application concerned, SPCA based features perform at least as good as PCA based features. The method has been applied to a real-world bearing data set that is widely used in condition monitoring.