Capsule networks for intelligent fault diagnosis: a roadmap of recent advancements and challenges

Abstract

Against the backdrop of the rapid development of intelligent fault diagnosis (IFD) technology, Capsule Networks (CapsNets), owing to their ability to preserve spatial hierarchical information and enhance interpretability, have emerged as a promising alternative to traditional convolutional neural networks (CNNs) and thus become a research hotspot in the field of IFD. However, there is currently a lack of systematic reviews on the application status and challenges of CapsNets in IFD. To bridge this gap, this paper views the literature published in the Web of Science Core Collection from January 2018 to June 2025 and conducts an analysis of CapsNets from three dimensions, namely theory, application fields, and model improvements. The findings show that CapsNets are primarily applied in mechanical fault diagnosis (MFD) (75.2%, with bearing faults accounting for 35.6%) and electrical equipment (16.8%), while their applications in energy, healthcare, and other fields remain limited. Existing studies have improved the model by enhancing training sample quality, optimizing network architectures, and strengthening interpretability. Nevertheless, several challenges persist, including low training efficiency, poor sample quality, insufficient model accuracy and robustness, inefficient dynamic routing mechanisms, and weak interpretability depth. Looking ahead, the development of CapsNets in IFD can focus on six directions: lightweight and efficient algorithm design; multimodal data fusion and cross-domain transfer learning; enhancement of anti-noise ability and robustness; compound fault decoupling and interpretability enhancement; edge computing and embedded deployment and integration with digital twins. This paper provides systematic guidance for the application and research of CapsNets in IFD.

Topics & Concepts