CyclicalNet: A Dense-Connected Architecture for Multiscale Cyclic Feature Learning in Planetary Gearbox Fault Diagnosis

Abstract Planetary gearboxes, crucial for industrial applications like wind turbines, automobiles, and robotics, are susceptible to failures due to complex vibrations and harsh conditions, making the development of effective fault diagnosis methods essential. While general-purpose deep learning models—such as Transformers, ResNets, DenseNets, and generative adversarial networks (GANs)—have shown promise in identifying intricate patterns within time series data for fault diagnosis, conventional one-dimensional convolutional kernels have inherent limitations, as they can only capture features from adjacent time series points. Additionally, methods that utilize attention mechanisms often struggle to uncover consistent dependencies among sparse data points that may be concealed within complex patterns. This paper introduces CyclicalNet, an innovative deep learning architecture designed to address these fundamental limitations in planetary gearbox fault diagnosis. The architecture uniquely integrates cyclical blocks, encoder blocks, and dense connections to process high amplitude–frequency components, traditionally overlooked in conventional models. A strategic reshape operation transforms one-dimensional sequential data into two-dimensional representations, enabling more structured and efficient feature extraction while facilitating simultaneous modeling of temporal and periodic features. The incorporation of dense connections promotes multiscale feature learning and enhances information flow across the network, resulting in richer and more discriminative feature representations. Through comprehensive experimental evaluation on planetary gearbox datasets, CyclicalNet demonstrates superior performance in fault diagnosis accuracy compared to existing methods, establishing a new benchmark in predictive maintenance technology.