HVD (Hilbert vibration decomposition) algorithm is often used for signal decomposition under frequency coupling due to its strong anti-modal aliasing performance, but the selection of starting points is the key. In view of this, an autoregressive HVD algorithm was proposed to ensure that the boundary effect was inhibited and the precise decomposition of frequency coupled signals was solved. In this method, HVD was taken as the decomposition basis function and combined with the adaptive boundary continuation capability of the autoregression model, the HVD algorithm was optimized, and the accurate decomposition of signals with frequency coupling characteristics in composite faults was finally completed. Taking the two-stage parallel shaft gearbox of the wind turbine experimental system as the verification object, the compound information at different speeds was analyzed.The identification effect proves that autoregressive HVD has a significant advantage in the compound fault diagnosis with frequency coupling characteristics.