Research on Vibration Signal Processing and Fault Diagnosis Algorithms for High-Pressure Quintuplex Pumps

Abstract

High-pressure quintuplex pumps are key components in oil and gas drilling systems, but their fault diagnosis remains challenging due to severe inter-cylinder vibration crosstalk, operating-condition variability, and limited labeled samples. To address these issues, this paper proposes an integrated vibration-based diagnostic framework, termed VICA-DA-ViT. First, a VMD–FastICA-based blind source separation strategy is used to suppress inter-cylinder crosstalk and extract more discriminative vibration components from mixed signals. Then, angular-domain resampling and normalization are applied to reduce speed-induced non-stationarity and improve cycle consistency. The processed one-dimensional signals are further converted into two-dimensional GAF–RP representations, where GAF captures global temporal correlations and RP provides complementary recurrence information. Finally, a ViT-based domain-adaptive diagnostic framework is constructed for unsupervised cross-condition diagnosis by integrating domain-adversarial pretraining, weighted bidirectional sample pairing, and Gaussian-mixture-based confidence filtering around a standard Vision Transformer backbone. Rather than modifying the internal architecture of ViT, the proposed framework improves cross-condition generalization by combining signal decoupling, angular-domain alignment, and domain-adaptation strategies for high-pressure quintuplex pump fault diagnosis. Experiments on a quintuplex pump dataset under multiple operating conditions show that the proposed framework achieves an average cross-condition accuracy of 87.42%, outperforming CD-Trans by 11.94%. These results demonstrate that VICA-DA-ViT can effectively improve the robustness of small-sample cross-condition fault diagnosis for high-pressure quintuplex pumps.