Unsupervised Learning with Counterfactual Explanations for Sucker Rod Pump Fault Diagnosis

Sucker rod pump fault diagnosis faces several challenges, including the reliance of supervised learning on highquality labeled data, data imbalance, and limited model interpretability. To address these issues, this paper proposes a novel fault diagnosis framework that integrates unsupervised learning with counterfactual explanations. The framework first employs a Transformer-based autoencoder to learn the normal operational patterns of the equipment from unlabeled data. Subsequently, a Gaussian Mixture Model is applied for probabilistic modeling in the latent feature space to quantify anomaly risk. The key component of this framework is the introduction of a generative model-based counterfactual explanation mechanism, which generates “minimal-change” to identify key factors causing anomalies and guide repairs, thereby addressing the “black box” problem of deep learning models. Experimental results on a real-world dataset from a shale oil field show that the proposed unsupervised model demonstrates superior performance compared to supervised benchmarks across key evaluation metrics. This performance, achieved without relying on fault labels, validates the method’s effectiveness for interpretable and data-efficient predictive maintenance in practical industrial environments.