Tooth root crack is a common fault in the gear system, it’s of significance to detect the crack fault during the operation process of the gearbox. However, the gearbox usually working in a non-stationary condition, namely speed or load are time-varying, which increases the difficulty of fault diagnosis since the statistic features and spectrum vary by time. In this paper, an order envelope analysis method is proposed for fault diagnosis of a two-stage gearbox with local crack fault under non-stationary working conditions. Firstly, an accelerometer and an encoder are mounted on the gearbox to acquire non-stationary vibration data and rotating speed, respectively. After that, the angle domain signals are derived from the non-stationary vibration data from the interpolation algorithm. Then envelope analysis is employed for the angle domain signals to detect the crack characteristic frequency components in an envelope spectrum. Finally, the proposed approach is assessed by seeded tooth root crack fault in different working conditions, results show that the new method can effectively detect the tooth cracks in various non-stationary working conditions.