Comparative Analysis of Autoencoder and Contrastive One-Class Anomaly Detection in Reciprocating Compressors

Fault detection in reciprocating compressors is crucial for ensuring the reliability and efficiency of industrial operations, as failures in these systems can lead to costly downtimes. However, the lack of faulty data challenges the application of supervised approaches. Therefore, many one-class learning-based proposals have been introduced to address this task. This study presents a comparative analysis of two advanced models for fault detection under a one-class scenario: the autoencoder and the Contrastive One-Class Anomaly detection (COCA) model. Both models were evaluated on their ability to detect anomalies in high-resolution time series data from a two-stage reciprocating compressor under varying operational conditions. The autoencoder, trained solely on healthy condition data, demonstrated superior performance with higher and more consistent balanced accuracy across all test conditions compared to the COCA model, which showed more significant variability and the presence of outliers. The findings suggest that the autoencoder approach is more reliable for early fault detection in industrial applications, offering better generalization and robustness.