Pump stroke event-driven cross-cylinder transfer learning for fluid-end fault diagnosis in multi-cylinder drilling pumps

Abstract Cross-cylinder transfer diagnosis offers a promising solution to the issue of data scarcity and high acquisition costs for pump faults. However, numerous steady signal segments from the fluid end cannot extract differentiated condition features in noisy environments, impeding cross-cylinder fault characterization. Moreover, the cyclic-reciprocating operation of the fluid end causes considerable variation in fault signatures with stroke position, while traditional models fail to exploit the spatial structural information embedded in these stroke-related signals. To address these challenges, this paper proposes a pump stroke events-driven cross-cylinder transfer diagnosis method. First, pump stroke events are defined and segmented using phase information to filter out noise-corrupted segments and concentrate analysis on fault-sensitive regions. Second, a Residual Capsule Network with Wide Kernel structure (WRCN) is developed to exploit spatial structural information: capsule layers explicitly model spatial-part relationships to preserve stroke-position dependencies; wide-kernel residual convolutions suppress high-frequency noise and enhance salient low-frequency feature representation; and a domain classifier extracts domain-invariant features across cylinders. Experiments confirm that both pump stroke events and the WRCN independently improve cross-cylinder diagnostic performance, with the efficacy of the former depending on proper event selection. The discharge-event-driven method outperforms existing approaches, achieving a relative accuracy increase of more than 4% with enhanced feature-space separability, thereby establishing a reliable new paradigm for cross-cylinder transfer diagnosis.