Traditional overdamped stochastic resonance (SR) methods are difficult to match with complicated and variable input signals due to single stable-state types. Moreover, their performance depends on the parameter selection of highpass filters. To further explore the potential of SR, this paper studies the behavior of underdamped SR in a multistable nonlinear system by analyzing its output frequency responses, and presents a promising underdamped multistable SR method for weak signal detection and further incipient fault diagnosis of machinery. Numerical analyses indicate that the proposed method is supposed to possess two advantages: 1) the stable-state diversity of the multistable potential makes it easily match with input signals and 2) underdamped multistable SR is equivalent to a bandpass filter as the rescaling ratio varies, which is able to suppress the interference from multiscale noise. Simulated and experimental data of rolling element bearings demonstrate the effectiveness of the proposed method. For comparison, ensemble empirical mode decomposition (EEMD) method and traditional overdamped bistable SR method are also employed to process the data. The comparison results show that the proposed method can effectively detect incipient fault characteristics and perform better than traditional SR and EEMD methods.