Fault diagnosis from visualization perspective using stream statistics

This paper proposed a concept called stream statistics for fault diagnosis. Its idea is to count the obtained signal's distribution in all intervals and transform them to statistical features. This idea differs from the conventional time and frequency domain methods and offers promising advantages, e.g., no need to select parameters and less calculation, over the conventional ones. To cope with the accompanied high dimensional problem, we apply the linear discriminant analysis (LDA) method for projecting statistical features to 2D or 3D space, which is feasible for visualization with the purpose of fault diagnosis. Other dimensionality reduction method, principle component analysis (PCA), is took into comparison in order for demonstrating the advantages of the proposed method. The visualization results of motor bearing data & hard disk drive (HDD) data prove the effectiveness of the proposed method. Moreover, the relationship between visualization results and condition monitoring is established and a modification for LDA based on original criterion function is given.