In light of the low efficiency of the bistable stochastic resonance model in modulated signal processing, an extraction and detection method for weak signals was proposed based on adaptive multi-stable cascaded stochastic resonance (AMCSR). First, the vibration signal was demodulated, and the multi-stable stochastic resonance model was adopted, according to the distribution feature of the upper and lower envelope signal. Second, the stochastic resonance model parameters were optimized adaptively through the ant colony algorithm, which adopted the fault feature frequency amplitude as the optimization index. Finally, the fault feature frequency of the envelope signal was progressively strengthened through the optimal cascaded stochastic resonance system with noise as the energy source, then the enhanced output signal was obtained. Experimental results of the bearing vibration signal analysis show that, through the AMCSR method, the fault feature frequency components can be strengthened, and the weak fault signal can be effectively extracted from the mixed signal.