Electric Submersible Pumps (ESPs) are widely used in oil production systems but are frequently affected by electrical, thermal, and mechanical failures that lead to unplanned shutdowns and production losses. With the growing availability of high-frequency downhole monitoring data, data-driven methods have gained increasing attention for ESP condition monitoring. However, many existing machine-learning-based approaches operate as black boxes and provide limited physical interpretability, restricting their practical deployment. This study proposes an explainable artificial intelligence (XAI) framework for ESP failure diagnosis using real multivariate field data. The framework integrates Principal Component Analysis (PCA) for dimensionality reduction, a Random Forest (RF) classifier for multi-class operating state identification, and SHapley Additive exPlanations (SHAP) for transparent model interpretation. Multivariate ESP sensor data, including pressure, temperature, vibration, voltage, and current measurements, were preprocessed and labeled based on documented failure reports. PCA was applied to address multicollinearity, while the RF model classified operating conditions into stable, unstable, and distinct failure modes. The proposed framework achieved high classification accuracy and consistently detected failure conditions several days prior to shutdown events. SHAP-based analysis further provided feature-level explanations of model predictions, enabling identification of dominant physical drivers such as motor overloading and abnormal vibration behavior. The results demonstrate that combining predictive performance with explainability enhances the reliability and practical value of data-driven ESP diagnostic systems, offering an effective decision-support tool for proactive monitoring and maintenance in digital oilfield applications.