Digital Twin-Driven Transfer Learning for Predictive Maintenance in Industry 4.0: Current Frameworks, Challenges, and Research Directions

Predictive maintenance is a central component of Industry 4.0, but the development of reliable artificial intelligence models for fault diagnosis remains limited by the scarcity of real failure data, class imbalance, and the high cost of data annotation. The integration of Digital Twins and Transfer Learning provides a promising approach to overcoming these limitations by generating synthetic degradation and fault data and transferring diagnostic knowledge from simulated environments to real industrial assets. This study presents a systematic review of peer-reviewed literature indexed in Scopus and conducted in accordance with the PRISMA guidelines. The selected studies were analyzed and classified according to Digital Twin architecture, synthetic-data generation strategy, transfer-learning mechanism, industrial application, and reported implementation challenges. The reviewed frameworks mainly combine physics-based, data-driven, or hybrid Digital Twins with fine-tuning, parameter transfer, feature alignment, discrepancy minimization, and adversarial domain adaptation. Digital Twins are primarily used to reproduce rare fault conditions, support model pretraining, and reduce dependence on labeled target-domain data. Nevertheless, several barriers continue to limit industrial adoption, including the simulation-to-reality domain gap, insufficient model fidelity, negative transfer, computational overhead, and difficulties in maintaining synchronization under changing and non-stationary operating conditions. The absence of standardized datasets, common benchmarks, and comparable evaluation protocols also restricts reproducibility and generalization across industrial systems. Future research should focus on uncertainty-aware Digital Twins, adaptive and explainable transfer mechanisms, Generative AI-assisted data synthesis, standardized validation procedures, and edge-enabled deployment to support scalable and real-time predictive maintenance applications.