Wind turbines are widely used in wind power systems, and have become the central pillar in the transition to renewable and low-carbon energy. As a key component, condition monitoring of mainshaft bearing (MB) is critical to ensure the safe and stable operation of wind turbines. Nevertheless, due to harsh operating environments and ultra-low rotating speeds, conventional diagnostics based on vibration analysis always perform poorly. More importantly, naturally damaged MB usually presents compound faults, which further increases the diagnostic difficulty. To resolve this issue, a generalized sparse spectral coherence using acoustic emission (AE) is proposed for compound fault diagnosis of MB. Firstly, to obtain synchronized rotating speeds, physics-inspired sparse learning is innovatively developed to estimate rotational speed from AE signals themselves. Subsequently, by introducing the lp/lq norm, a generalized sparse cyclic spectrum is constructed for compound fault diagnosis. On this basis, the data-driven optimization scheme is then designed to improve diagnostic accuracy. Finally, the superiority of the developed method is validated by simulation analysis and engineering applications. The results demonstrate that the proposed method can effectively perceive compound faults of MB operating under ultra-low speed, which may provide a useful tool for the health management of wind turbines.