Deep Learning Techniques and Attention Mechanisms in Fault Diagnosis and Predictive Maintenance: A Review

Industrial equipment Fault Diagnosis (FD) and Remaining Useful Life (RUL) prediction are fundamental tasks in Predictive Maintenance (PdM) and Prognostics and Health Management (PHM). In recent years, deep learning has substantially advanced industrial PHM by enabling end-to-end representation learning from sensor signals and operational time series. Nevertheless, the practical deployment of these methods remains constrained by cross-condition distribution shift, limited labeled data, long-range temporal dependencies, noisy or incomplete observations, insufficient interpretability, and weak reproducibility. Attention mechanisms and Transformer-based architectures provide a promising pathway to address these challenges through global dependency modeling, multivariate interaction learning, and structured feature selection. This review systematically examines recent advances in deep learning and attention mechanisms for industrial PHM from an engineering-oriented perspective. For FD, we summarize end-to-end diagnostic models, attention-enhanced evidence aggregation, contrastive and self-supervised learning, and cross-condition transfer diagnosis. For PdM and RUL prediction, we review long-sequence forecasting, uncertainty-aware modeling, interpretable prediction, and physics-informed integration. Beyond model taxonomy, this review further discusses critical evaluation and deployment issues, including over-optimistic performance, data leakage risks, overlapping-window bias, out-of-distribution robustness, uncertainty-based decision support, and open reproducibility. Finally, a unified evaluation perspective and a closed-loop PHM framework are presented to connect data processing, model inference, trustworthy output, maintenance decision-making, and continuous feedback. This review aims to support more reliable model selection, evaluation design, and deployment-oriented PHM implementation.