Aiming at the problems of multi-modal signal interference, weak early fault characteristics and insufficient state evolution modeling in health monitoring of pumping system, this paper puts forward a diagnosis algorithm combining multi-scale time series characteristics and dynamic health assessment. Core methods include: (1) Synchronizing multi-source signals such as synchronous motor current, bare rod load, and vibration based on the pumping unit stroke cycle; extracting time-domain statistics, frequency-domain energy proportion, and multi-scale permutational entropy (MPE) to construct a 30-dimensional feature vector; (2) Design Attention-BiLSTM network, capture the long-term dependence of health state by bidirectional LSTM (BiLSTM), and adaptively weight the critical stroke period through attention mechanism to realize accurate state classification; (3) Build a continuous health index (HI) based on Mahalanobis distance, quantify the degree of equipment degradation, and divide it into three stages: health, degradation and failure criticality; (4) Introducing time series weighted support vector machine (SVM) to enhance the weight of recent samples and improve the robustness of early warning. The experiment uses the two-year operation data of 10 pumping units in an oil field in China, including 6800 stroke samples in four States, such as normal and wandering valve stuck. The results show that the accuracy rate of this algorithm is 95.8%, the macro F1 score is 0.941, and the recall rate of early pitting corrosion of bearings is 93.2%, which is nearly 30% higher than that of the traditional method. The early warning can be triggered about 2500 cycles before the fault occurs, and the single sample diagnosis takes seconds, which meets the requirements of real-time and interpretability in industry and provides an effective decision-making basis for predictive maintenance.
Jun Xia (Sun,) studied this question.