Compressors are the key assets of many industrial processes, where reliable operation underpins productivity and safety. Maintaining optimal performance and preventing disruptions requires diagnostic tools capable of accurately identifying both healthy and faulty operating conditions. Effective diagnostics support proactive maintenance planning and improve overall system reliability. Traditional diagnostic methods, such as threshold- and rule-based approaches, often struggle to generalize across different machines and operating environments. Physics-based models provide interpretability but require extensive system knowledge and may fail to capture real-world variability. Data-driven techniques offer scalability and automation, though their effectiveness is limited by scarce fault data and difficulties in generalizing to unseen conditions. This study develops a data-driven diagnostic framework that combines a synthetic fault dictionary, sequence-to-image signal encoding, and convolutional neural networks (CNNs). The framework generates balanced training datasets by embedding synthetic faults (bias, drift, and spike) into healthy field data, enabling the simulation of diverse fault conditions. By leveraging image-based representations, CNNs learn discriminative fault features even with limited faulty data. The framework is validated by using real-world data from three compressors of the oil and gas sector. Results demonstrate high diagnostic accuracy and low detection latency and thus confirm the potential of the proposed method as a scalable and robust solution for compressor fault diagnostics.
Compressors are the key assets of many industrial processes, where reliable operation underpins productivity and safety. Maintaining optimal performance and preventing disruptions requires diagnostic tools capable of accurately identifying both healthy and faulty operating conditions. Effective diagnostics support proactive maintenance planning and improve overall system reliability. Traditional diagnostic methods, such as threshold- and rule-based approaches, often struggle to generalize across different machines and operating environments. Physics-based models provide interpretability but require extensive system knowledge and may fail to capture real-world variability. Data-driven techniques offer scalability and automation, though their effectiveness is limited by scarce fault data and difficulties in generalizing to unseen conditions. This study develops a data-driven diagnostic framework that combines a synthetic fault dictionary, sequence-to-image signal encoding, and convolutional neural networks (CNNs). The framework generates balanced training datasets by embedding synthetic faults (bias, drift, and spike) into healthy field data, enabling the simulation of diverse fault conditions. By leveraging image-based representations, CNNs learn discriminative fault features even with limited faulty data. The framework is validated by using real-world data from three compressors of the oil and gas sector. Results demonstrate high diagnostic accuracy and low detection latency and thus confirm the potential of the proposed method as a scalable and robust solution for compressor fault diagnostics.
CNN-based Compressor Diagnostics by Means of Time Series Encoded as Images
Caputo C. A.
;Venturini M.;Manservigi L.;
2026
Abstract
Compressors are the key assets of many industrial processes, where reliable operation underpins productivity and safety. Maintaining optimal performance and preventing disruptions requires diagnostic tools capable of accurately identifying both healthy and faulty operating conditions. Effective diagnostics support proactive maintenance planning and improve overall system reliability. Traditional diagnostic methods, such as threshold- and rule-based approaches, often struggle to generalize across different machines and operating environments. Physics-based models provide interpretability but require extensive system knowledge and may fail to capture real-world variability. Data-driven techniques offer scalability and automation, though their effectiveness is limited by scarce fault data and difficulties in generalizing to unseen conditions. This study develops a data-driven diagnostic framework that combines a synthetic fault dictionary, sequence-to-image signal encoding, and convolutional neural networks (CNNs). The framework generates balanced training datasets by embedding synthetic faults (bias, drift, and spike) into healthy field data, enabling the simulation of diverse fault conditions. By leveraging image-based representations, CNNs learn discriminative fault features even with limited faulty data. The framework is validated by using real-world data from three compressors of the oil and gas sector. Results demonstrate high diagnostic accuracy and low detection latency and thus confirm the potential of the proposed method as a scalable and robust solution for compressor fault diagnostics.I documenti in SFERA sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
