Physics-Guided Soft Sensing-Aided Fault Diagnosis Framework of Sucker Rod Pumping System

Real-time monitoring of oil production status is crucial to its reliability and safety. The conventional solutions aim to diagnose the sucker rod pumping system (SRPS) by the indicator diagram (ID) directly. However, it is challenging to obtain in real-world production, restricting the applications. To tackle the issue, we propose a physics-guided soft sensing (PGSS)-aided fault diagnosis framework to realize efficient and explainable diagnosis without relying on hard-to-measure mechanical parameters online. The framework includes the ID inversion, dual-channel feature extraction, and fault diagnosis tasks. First, a soft sensing (SS) technique is applied to estimate the ID based on easily measured information. Then, we propose a dual-channel feature extraction with curvature analysis for the estimated IDs. Finally, we conduct a neural network-based diagnostic model with a combination of cross-entropy, self-entropy, and regularization terms as optimization strategies, which can improve the generalization and reduce the impact of uncertainty caused by the above SS work. To validate our method, we apply real-world SRPS data collected from oilfields in Northeast China and achieve excellent estimation and diagnosis results. It can give a novel idea to achieve high practicality and explanation for fault diagnosis tasks of SRPS.