Condition-Adaptive Phase Alignment and Load-Assisted Fusion for robust fault diagnosis of drilling pumps

Abstract

Abstract

Abstract Drilling pumps operate under highly dynamic conditions involving varying stroke rates, fluctuating loads, and sensor measurements corrupted by operational noise. In addition, structural coupling among multiple cylinders, flow-induced pressure pulsations, and local temporal drift can make fault signatures highly non-stationary. These factors reduce the reliability and robustness of conventional steady-state diagnostic methods and generic deep learning models under the investigated drilling-pump operating conditions. To address these challenges, this study proposes a Condition-Adaptive Phase Alignment and Load-Assisted Fusion Network (CAPA-LAFNet) for robust fault diagnosis of drilling pumps. First, based on angular-domain preprocessing, a residual phase alignment module is introduced to correct the small-scale local misalignment that remains after fixed mechanical phase alignment. This module is conditioned on stroke rate, cylinder position, and lightweight structural-response statistics, thereby reducing temporal dispersion. Second, a stroke-rate-conditioned dynamic analytic filtering front-end is constructed to adapt low-level filtering responses to changes in local waveform patterns and normalized frequency-band distributions. Third, a pressure-assisted branch is integrated through a gated fusion mechanism to introduce fluid-load-related context into the stress-vibration main branch. Experimental results over five independently regenerated leakage-controlled partitions show that CAPA-LAFNet achieves an average diagnostic accuracy of 99.43 ± 0.35% under the adopted evaluation protocol and exhibits improved robustness under the tested noisy conditions. Ablation and physical-consistency analyses further indicate that residual phase compensation, stroke-rate-conditioned filtering, and pressure-assisted fusion contribute to fault discrimination under variable operating conditions.