A comprehensive review of capsule neural networks for modern predictive maintenance

Abstract There is a significant evolution in modern predictive maintenance (PdM) using advanced deep learning architectures. Among these, Capsule networks (CapsNets) have emerged as a promising alternative to CNNs and their variants, given their ability to preserve hierarchical spatial relationships and to improve feature interpretability naturally. Although there is an increasing trend of CapsNet-based PdM studies, there are very few reviews that consolidate CapsNets’ integration within PdM pipelines for diagnosis and prognosis tasks. This review aims to address this gap through a PRISMA(Preferred Reporting Items for Systematic Reviews and Meta-Analyses)-based systematic review where a final set of 97 peer-reviewed studies published from January 2019 to January 2026 were included. The review examines CapsNets for fault diagnosis, remaining useful life estimation, hybrid architecture design with industrial domain adoption, and model explainability. A conceptual integration framework is proposed to incorporate CapsNets for their distinct role while remaining architecturally compatible with the existing maintenance workflows. Moreover, 5 dominant hybridisation strategies, including convolutional neural networks, attention mechanisms, domain adaptation, temporal CapsNets and system-level architectures, governing CapsNet integration for the PdM pipeline are identified, with domain-skewed industrial applications concentrated in mechanical and automotive systems. Further, the experimental realism of the surveyed studies, benchmark datasets and associated challenges is critically examined. The review concludes that the CapsNets paradigm for PdM is structurally promising but operationally immature, due to challenges with computational complexity, real-time deployment, and limited validation in large-scale industrial environments. This work thus provides a foundational reference to researchers and practitioners to understand current advancements in CapsNet-based PdM and its associated open challenges, with potential for developing interpretable PdM workflows.