Wind turbine gearboxes operate in harsh environments under varying operational conditions, making them prone to faults that can lead to significant maintenance costs and potential failures. Although unsupervised fault detection methods show promise in monitoring these critical components, they often struggle to maintain performance when faced with distribution shifts caused by varying operational conditions. This paper proposes a novel unsupervised method that leverages domain adversarial training to achieve robust fault detection across different working conditions. The method combines squared envelope order spectrum for feature extraction and domain adversarial training to learn domain-invariant features. Experiments conducted on a gearbox dataset demonstrate the effectiveness of our approach under varying rotational speeds, with consistent performance observed in both in-distribution and out-of-distribution scenarios. The results show that the proposed method significantly outperforms conventional fault detection approaches across different operational conditions.