On accounts of complex structures and changeable operating conditions of helicopter planetary gear train， the vibration signal is complicated and disturbed by fluctuating working environment. Moreover， various fault patterns and less samples increase the difficulty of the planetary gear train fault diagnosis. Aiming at these problems， a new fault diagnosis model named stacked contractive autoencoders-auxiliary classifier generative adversarial networks （SCAE-ACGAN） is proposed， which combines SCAE and ACGAN. The generator of ACGAN generates new samples which are similar to original samples， expanding diagnosis samples. Then， new samples are transformed to discriminator with original samples. SCAE is used as the discriminator of ACGAN to extract good robustness fault features from input samples and discriminate its authenticity and categories. Through adversarial learning， the generator and discriminator are alternatively optimized to improve the abilities of generation and discrimination. We evaluate the proposed methods on diagnosis of helicopter planetary gear tooth crack fault. From the experimental results， the SCAE-ACGAN can improve anti-noise ability and get better diagnosis performance in less samples cases. The experimental results show that SCAE-ACGAN exhibits the excellent performance of fault diagnosis， and has good robustness and adaptability in the case of small number of samples and fluctuation in working conditions.