Time-frequency Spectra Recognition Based on Sparse Non-negative Matrix Factorization and Support Vector Machine
基于稀疏性非负矩阵分解和支持向量机的时频图像识别

In the field of mechanical fault diagnosis, it is difficult to recognize the running condition of machines by human based on images corresponding to the condition such as time-frequency spectra, obit, power spectra, and so on. The most meaningful features required by learning machines, which can reorganize running condition of machines automatically, are always difficult to select and extract from the images. In this paper, the problem of machine running condition recognition based on images is treated purely as image recognition problem, so the procedure of meaningful features selection and extraction can be avoided. Sparse non-negative matrix factorization (SNMF) and support vector machine (SVM) are introduced to recognize the time-frequency spectra and therefore the corresponding running condition of machine automatically. After applying SNMF to image, the dimension is reduced obviously while the connotative and main features of image are reserved, therefore the computation cost of image recognition with SVM is saved and the recognition accuracy is possibly improved. Experimental results show that the proposed method can obtain higher recognition accuracy than conventional method and is dependent only weakly on the time-frequency analysis method.