The invention relates to a device which uses a fractal dimension to perform the online detecting and pre-warning the abnormal state of a sinking roller. The invention comprises a pair of vibration sensors (2) with each connected with a data acquisition module (3) and installed on a framework near to a sinking roller (1), the output terminal of the data acquisition module is connected with a computation controlling module (5) through connecting a signal storage (4), and the computation controlling module is also connected with an alarm (6). The method comprises the following steps: step one, a vibration signal time series is acquired to acquire a state vector Y, and a vibration waveform assembly is reconstructed; step two, the average value Ave (Y) and the peak-to-peak value PP (Y) of Y are accounted, to make a topological mapping; step three, a signal is reconstructed; step four, the length fractal dimension d of M is accounted; step five, the alarm values Alarm1(d) and Alarm2(d) of the fractal dimension of the sinking roller is set; step six, the length fractal dimension of a vibration waveform is monitored by adopting the method of the forward lead of a gliding window, when d is larger than or equal to Alarm2(d) and lower than or equal to Alarm1(d), the sinking roller state is normal, when that d is lower than Alarm2(d) or larger than Alarm1(d) presents continuously for five times or more, the sinking roller state is abnormal, and the system alarms.