Continual anomaly detection of offshore reciprocating natural-gas compressors under variable operating conditions

Natural gas compressors on offshore platforms operate under non-stationary conditions. Their data distributions change with load variation, pressure fluctuation, production decline, and equipment degradation. Static anomaly detection models often suffer from threshold drift, missed alarms, and catastrophic forgetting during long-term operation. To address these challenges, this study proposes an intelligent monitoring framework that integrates unsupervised anomaly detection, continual learning, and root-cause analysis using real data from the compressors. First, a variational autoencoder - extreme value theory (VAE-EVT) model is developed to learn normal operating patterns and determine adaptive anomaly thresholds through tail modeling of reconstruction errors. Second, a deep generative replay with fine-tuning strategy (DGR-FT), is proposed to update the model under new operating conditions without storing complete historical raw data. Third, SHAP is introduced to explain anomaly decisions and identify key fault-related variables. The framework was validated using field data collected from January 2023 to January 2025 on a single reciprocating compressor platform, with three real valve-related fault cases. In Case A, the VAE-EVT model achieved an F1-score of 97.77% and detected valve faults 58 days earlier than the PLC threshold alarm. In Case B, DGR-FT improved Recall and F1-score from 8.88% to 16.24% to 98.96% and 99.41%, respectively. In Case C, DGR-FT achieved a mean F1-score of 98.33% across 30 random seeds. The training time was only 4.05 s, which was much shorter than that required for joint training. SHAP analysis showed that discharge-temperature differences, pressure-ratio changes, and bearing temperatures were the main contributors to anomaly decisions. These results demonstrate that the proposed framework can improve early fault detection, mitigate catastrophic forgetting, and provide interpretable support for maintenance decisions.