Data-Driven Fault Detection in Reciprocating Compressors: A Method Based on PCA and GLRT

This study introduces a novel failure detection algorithm using Principal Component Analysis (PCA) and the Generalized Likelihood Ratio Test (GLRT) for reciprocating compressor valves. Unlike typical machine learning methods requiring both normal and fault condition data, our approach only needs normal condition data. The process begins with acquiring vibration signals under various conditions, but only normal data is used for modeling. Features are extracted from these signals in the time domain, followed by training a PCA-based model. GLRT then helps in setting parameters for fault detection. Our model demonstrates 87% reliability, as indicated by the area under the ROC curve.