A bimodal framework for nonstationary process monitoring via collaborative contrastive and adversarial unsupervised learning

Recognizing nonstationarity is pivotal for trustworthy industrial process monitoring. Existing methods address this issue from a unimodal perspective, which struggles to capture intrinsic heterogeneity. To resolve this, we introduce a novel unsupervised multimodal nonstationary monitoring framework (UMNMF), integrating a bimodal paradigm with contrastive and adversarial schemes. Initially, the knowledge labeling unit (KLU) is established to generate pseudo-labels augmented with prior knowledge for semantic guidance. Subsequently, the dynamic alignment and encoding unit (DAEU) exploits contrastive language-image pre-training (CLIP) and the Vision Transformer (ViT) for modality-aware alignment through a pseudo-supervised contrastive mechanism. Furthermore, the association alignment and distillation unit (AADU) is devised to achieve decoupling through self-adversarial distribution regularization within a variational graph autoencoder (VGAE). The superior performance is substantiated by extensive experiments on three industrial processes, where the UMNMF attains an average fault detection rate exceeding 94 % and maintains a false alarm rate below 2.5 %. Additional ablation studies further confirm the contribution of each module to overall performance improvement.