Abstract: Unanticipated wear of Tunnel Boring Machine (TBM) disc cutters is a critical factor causing project delays and cost overruns in tunneling engineering. Accurate, real-time prediction of the cutter's wear state is therefore essential for enabling predictive maintenance. Data-driven methods, particularly deep learning, have shown promise for this task, but their performance is constrained by the scarcity of high-quality labeled data in practical industrial settings. To address this challenge, we propose a novel, decoupled semi-supervised framework called PL-HLANet. The first component of this framework is a multi-view Pseudo-Labeling (PL) module, which mines high-confidence supervisory signals from massive unlabeled data by leveraging heterogeneous views derived from feature engineering and diverse model architectures; it is followed by a consistency check to ensure label quality. This process effectively augments the training set while correcting for sampling bias. Subsequently, a specialized Hierarchical Hybrid Attention Network (HLANet) is used to make predictions. The HLANet organically integrates a Temporal Convolutional Network (TCN) for local feature extraction, a Bidirectional Long Short-Term Memory (Bi-LSTM) network for capturing temporal dynamics, and a custom attention mechanism for focusing on critical information. Experiments on a real-world tunneling dataset show that PL-HLANet significantly outperforms both supervised and mainstream semi-supervised baselines, like Mean Teacher and FixMatch. The framework's effectiveness is further substantiated by validations of its architectural design and data-driven selection of hyperparameters. Moreover, PL-HLANet has a high inference speed, showcasing its practicality for real-world scenarios. Our work provides an effective solution for machining equipment monitoring in data-scarce industrial environments.