A innovation independent component analysis (ICA) method is presented for the time-frequency image blind source separation (BSS). Because there are common and potent correlations between similar images, the efficiency of traditional ICA applied into the time-frequency image is not satisfied, so the value of mutual information and kurtosis is studied the correlation and non-gauss of the image’s sub-bands severally, and make choice of some sub-band as the input parameter of ICA. After application of simulated time-frequency images, it is validated that the new method can improve the limitation of traditional ICA, finally the new method is applied into time-frequency image of rotor’s fault signals (loose & misalignment), the time-frequency image of respective fault is separated successfully, so the information of the characteristic is method is recognized.