Abstract: Due to the continuous operation of mechanical equipment, the emergence of unknown faults is inevitable. To address the challenges of limited source domain knowledge and difficulties in identifying new fault patterns in target domains for mechanical equipment fault diagnosis using traditional transfer learning methods, this paper proposes a Contrastive Transfer Learning-Based Diagnostic Method for Unknown Faults in Mechanical Equipment (CTL-MEDM). First, a coarse-to-fine feature extraction network is designed to comprehensively leverage global and local characteristics, thereby enhancing the feature representation capability of raw vibration signals. Second, a Mutual Information Noise Contrastive Estimation (MI-NCE) loss function is introduced in conjunction with cosine similarity metrics to optimize feature space distribution, significantly improving inter-class discriminability between known and unknown faults. Building upon this foundation, a dynamic feature weight adjustment mechanism is proposed to align marginal distribution discrepancies between source and target domains through domain adversarial training, while adaptively optimizing critical feature weights via backpropagation. Experimental results demonstrate that the proposed method effectively achieves unknown fault diagnosis for bearings and gears under variable operating conditions.