Patch Time-Series Transformer Model for Low Severity Blockage Fault Detection in Centrifugal Pumps

Blockage in centrifugal pumps can significantly impair their hydraulics. Although the present data-driven approaches for diagnosing blockage faults have advanced significantly, three key challenges still remain. First, most studies focus on moderate-to-severe level of blockage, with limited capability for capturing the early-stage blockages that produce weak fault signatures. Second, the existing approaches lack a unified framework that can diagnose and classify blockage faults across a wide spectrum of severity, ranging from no-blockage condition to severe blockage condition. Third, several deep-learning methods rely on handcrafted features or signal-transformation techniques, which increases their implementation complexity and limits generalizability. To address these challenges, a patch time-series transformer (PatchTST) framework is proposed for blockage fault diagnosis of a centrifugal pump using multimodal sensor data. An experimental test facility is developed to generate 13 different blockage conditions, including no-blockage operation, suction-side blockage, discharge-side blockage, and combined suction and discharge blockage. The pump's responses to 13 conditions are demarcated into three levels of blockage fault severity depending on the percentage blockage in the flow area. The proposed PatchTST framework learns the discriminative temporal features directly from the raw sensor data, thereby eliminating the need for feature engineering or signal transformations. The best-case results indicate that the model achieves a nominally high-test accuracy of 98.95%, a validation accuracy of 99.33%, and an average accuracy of 97.55% under five-fold GroupKFold cross-validation. Overall, the PatchTST framework proposed in this study demonstrates a high potential for incipient blockage detection and predictive maintenance of centrifugal pumps.