A new bearing fault detection method based on the signal sparse decomposition theory was developed. An over-complete dictionary on which the bearing vibration signals in normal state can be represented sparsely was trained by the dictionary learning method. According to the fact that this dictionary just can sparsely represent the signals in normal state, the bearing vibration signal in unknown state was decomposed on this dictionary. The bearing state was determined by comparing the representation error of the signal on the dictionary with the given error threshold, and then the bearing fault detection was achieved. Experimental tests validate the effectiveness of the proposed method in bearing fault detection when setting an appropriate error threshold.