STP-Model: A semi-supervised framework with self-supervised learning capabilities for downhole fault diagnosis in sucker rod pumping systems

Abstract

In oilfield crude oil production, failing to detect and address downhole faults promptly can have severe consequences, including production interruptions, equipment damage, and resource wastage. Employing artificial intelligence to diagnose the operational status of oil wells is crucial. However, the high costs and specialized expertise required for data labeling make collecting sufficient training data challenging, limiting supervised learning approaches’ applicability. To address this, this paper introduces a semi-supervised framework with self-supervised learning capabilities for downhole fault diagnosis (STP-Model). It utilizes supervised learning modules for the limited labeled data available and constructs temporal association and prediction self-supervised modules for the extensive volume of unlabeled downhole fault data. These modules aim to learn the data’s intrinsic temporal relationship structure. The temporal association module splits and combines time-series data segments to construct positive/negative samples based on temporal consistency, learning the semantic information of time-series data. The temporal prediction module divides a continuous time series into two fixed-length sequences, aiming to minimize the gap between predicted and actual sequences and capture time series trend changes. These modules jointly compute the loss and update a shared-weight encoding network. Experimental results show the STP-Model achieves a 33.25% improvement in accuracy over supervised methods and an accuracy improvement of over 7.41% compared to other semi-supervised methods.