Adversarial domain adaptation with norm constraints for enhanced fault diagnosis in pumping units via surface motor power

Abstract

Timely downhole fault diagnosis of the pumping unit is a critical task in minimizing oilwell downtime and reducing energy consumption. Existing work has been exploring using low-cost, real-time surface motor power instead of installing dynamometer sensors at the wellhead for diagnosing faults. However, the high similarity of motor power samples across pumping condition categories leads to time-consuming and laborious manual labeling. Additionally, variations between wells cause shifting in the motor power feature distribution, reducing the diagnosis accuracy of traditional deep learning-based diagnoses. To address these challenges, this paper proposes a novel adversarial domain adaptation network with norm constraints (ADANN) for diagnosing faults in pumping units. First, the approach redefines feature extraction by incorporating modulated deformable convolution layers in place of traditional convolution modules within the deep residual network, thereby allowing for more precise and adaptive capture of geometric variations in motor power features. During domain adaptation, we innovatively introduce a norm-constrained alignment strategy into the domain adversarial training. The norm constraint, by maximizing the output variance of the batch normalization layer, encourages the model to learn more dispersed feature representations. This further enhances the ability of domain adversarial training to learn domain-invariant features, thereby improving generalization performance on the unlabeled target domain. Finally, comparative experiments on the collected dataset from real oil wells demonstrate the superior performance of ADANN.