Due to the over-reliance on prior knowledge and ignoring the influence of filter length on the deconvolution results, the traditional blind deconvolution method has many difficulties in extracting the weak fault characteristics of the bearing and estimating the signal impact period, resulting in poor robustness. Based on the powerful impact feature localization capability of envelope product spectroscopy (EHPS) and the excellent dynamic search performance of Alpha-Beta pruning strategy, this paper proposes a blind deconvolution method for adaptive fault cycle detection for fault diagnosis of intelligent bearings in rail transit. Guided by the envelope product spectrum, the maximum second-order cyclic stationary blind deconvolution without prior knowledge is realized, and at the same time, combined with the distribution characteristics of the fault-sensitive components in the frequency band, the fault harmonic energy ratio (PFHE) is constructed, which is expressed as the strength of the fault impact characteristics. Subsequently, the Alpha-Beta pruning strategy and the third-way method were combined to search for the optimal filter length, so as to realize the adaptive selection of multi-scale filter parameters. Simulation and experimental results show that the proposed method has better robustness and superiority than the other two blind deconvolutions. In addition, compared with the traditional traversal filter length method, the proposed method significantly improves the computational efficiency while ensuring the diagnostic accuracy.