MDCP-CPL: Multi-domain contrastive pretraining with Confidence-Aware curriculum Pseudo-Labeling for Semi-Supervised fault diagnosis

Deep learning-based automated fault diagnosis of rotating machinery facilitates early fault detection, reducing equipment downtime and maintenance costs. Nevertheless, diagnosis under limited labels remains a significant challenge in industrial settings. Existing Contrastive Learning (CL) methods are limited by: (a) reliance solely on time-domain features and (b) simplistic fine-tuning that underperforms with scarce labels. To address this, we propose MDCP-CPL, a novel semi-supervised framework that integrates Multidomain Contrastive Pretraining (MDCP) and Confidence-aware Curriculum Pseudo-Labeling (CPL). MDCP leverages time–frequency representations through intra-domain and inter-domain CL to learn robust and generalized features from scarce labels. An adaptive gated-fusion-projector dynamically balances the contribution of time and frequency encoders’ outputs, yielding an optimized multidomain latent representation. During fine-tuning, CPL combines curriculum learning with confidence-aware adaptive thresholding to progressively select pseudo-labels, stabilizing the training and suppressing label noise. Extensive experiments on the nuclear pump experimental dataset and public planetary gearbox datasets show that MDCP-CPL consistently outperforms state-of-the-art CL baselines, pseudo-labeling methods, and recent time–frequency contrastive approaches, achieving performance gains of up to 24%, 26%, 29%, respectively, under different label ratios. This validates that integrating multidomain contrastive learning, adaptive fusion, and curriculum pseudo-labeling enables highly effective fault diagnosis with minimal supervision, enhancing real-world industrial applicability.