Abstract: In view of engineering applications, the separation of signal from noise by wavelet analysis is studied for the fault diagnosis system of low-speed heavy-duty roller bearings. By virtue of the features of multi-level and multi-frequency band of wavelet decomposition and wavelet restructuring technique, a simple, accurate and practical fault diagnosis method is developed. The new method is available to find out some troubles, which cannot be found by other ways, such as the impact/wear between rollers, and inner and/or outer race. Its advantage and availability have been proved by actual disassembling of a low-speed high-duty bearing used for a ladle turntable, during which three damaged bearing rollers were found in outer ring with the raceway worn out. An impact/wear fault characterized by low frequency was thus diagnosed successfully. The result indicates that the wavelet analysis can extract effectively weak signals or separate them from noise and that it can make up for the deficiency of frequency spectrum analysis.