Early fault detection (EFD) in run-to-failure processes plays a crucial role in condition monitoring of modern industrial rotating facilities, which are more and more demanding for safety, energy and ecology saving and efficiency. To enable effective protecting measures, the evolving faults have to be recognized and identified as early as possible. The major challenge is to distill discriminative features on the basis of only the ‘health’ signal, which is uniquely available from various possible sensors before damage sets in and before the signatures of incipient damage become obvious and well-understood in the signal. Acoustic emission (AE) signal has been frequently reported to be able to deliver the early diagnostic information due to its inherently high sensitivity to the incipient fault activities, offering great potential of the AE technique for EFD, which may outperform the traditional vibration-based analysis in many situations. Up to date, the ‘feature-based’ multivariate analysis dominates the interpretation of AE waveforms. In this way, the decision making relies heavily on expert’s knowledge and experience, which is often a weak link in the entire EFD chain. With the advent of artificial intelligence, people are seeking for intelligent method to tackle this challenge. In this paper, we introduce a versatile intelligent analysis method for AE signals. A new architecture of convolutional generative adversarial network (GAN) is designed to extract deep information embedded in AE signals. In order to improve the robustness of the proposed EFD framework, a novel ensemble technique referred to as ‘history-state ensemble’ (HSE) is introduced in this paper. Primary merits of HSE are highlighted as: (1) it does not require extra computing time to obtain the base models; (2) it does not require special design on the network architecture and can be applied to different networks. To evaluate the proposed method, a rolling contact fatigue test monitored by AE sensors was performed, and experimental results have demonstrated the proposed ensemble method largely improves the robustness of GAN.