<p>Unsupervised clustering and domain adaptation for gearbox fault detection under variable speed conditions typically require labeled data from the source domain and prior knowledge of the number of classes. This paper proposes a novel fully unsupervised clustering and domain adaptation framework for gearbox fault detection under variable speed conditions. An engineering knowledge-guided processing approach integrates fault vibration signatures with a deep convolutional autoencoder, enhancing the effectiveness of feature extraction and computational efficiency by focusing on fault-relevant information. The number of classes and their corresponding clusters within each domain are determined using the K-means algorithm. A straightforward yet effective correlation-based strategy is applied to clusters from different rotational speeds, enabling accurate class alignment and domain adaptation across variable operating conditions. The proposed methodology eliminates the need for source domain labeled data and prior knowledge of the number of fault classes, addressing critical limitations of current unsupervised diagnostic systems. The approach is validated using two experimental gearbox datasets involving both single and compound faults. Results demonstrate its superior performance in unsupervised clustering, particularly in complex compound fault scenarios, and its effectiveness in domain adaptation across different operating conditions. These findings underscore the potential of the proposed framework for efficient and fully unsupervised fault detection in rotating machinery.</p>