Research on Intelligent Pump Fault Diagnosis Hinges on Cross-Domain Attentional Bias Diagnosis Transfer Learning Network

Deep learning-based fault diagnosis is a fundamental component in pump health monitoring. However, the efficiency of feature extraction and classification in cross-domain transfer learning remains a challenge. This paper proposes a novel Cross-domain Attentional Bias Diagnosis (CABD) model to enhance training efficiency for fault identification under variable working conditions. Our key contributions include (1) integrating enhanced adversarial training with spatiotemporal invariant features for superior domain adaptation (DA); (2) constructing a feature extractor with a hybrid CNN and Bi-LSTM model equipped with an attention mechanism to improve feature space alignment; (3) employing conditional entropy-based sample weighting to optimize feature transfer, thereby improving robustness in real-world applications. Experimental results on pump datasets demonstrate that the CABD model achieves at least 10% higher accuracy with higher computational efficiency compared to single-model-based neural networks, with an outstanding diagnostic accuracy of 99.448% in transfer tasks.