Computational study demonstrates accurate fault detection in industrial compressors using image-encoded sensor data, highlighting a scalable approach to proactive industrial maintenance.
Key Points
Develop and evaluate a data-driven diagnostic framework that pairs synthetic fault generation with sequence-to-image encoding and convolutional neural networks to address the scarcity of industrial compressor fault data.
Generated balanced training datasets by injecting synthetic fault profiles (drift, bias, and spikes) into field-recorded baseline operational data.
Encoded one-dimensional time series sensor readings into two-dimensional image representations to extract discriminative features via convolutional neural networks.
Validated diagnostic accuracy and detection latency using real-world operational datasets from three separate compressors operating in the oil and gas industry.
Demonstrated high diagnostic accuracy in differentiating healthy machine behavior from multiple simulated and real fault conditions.
Achieved low detection latency across diverse operating environments without requiring extensive prior physics-based system modeling.