Fouling fault early warning in centrifugal compressors based on adaptive Deep Forest

Abstract

Fouling in centrifugal compressors can significantly degrade its performance, leading to increased operational costs and reduced equipment reliability. Traditional methods for detecting and diagnosing fouling often struggle to provide early warnings across varying operating conditions. To address this challenge, a data-driven approach is proposed in this research. A comprehensive quantify of the effects of surface roughness on the compressor performance is conducted using the equivalent sand grain roughness model, and generate a complete performance database under various fouling scenarios. This database is then utilized to train several machine learning models, and a novel physical evaluation metric based on performance curves is introduced to assess the model’s predictive capabilities, further enhancing its interpretability. The results demonstrate that the adaptive deep forest model significantly outperforms other models in predicting compressor performance, particularly in handling nonlinear data. The newly developed physical evaluation metric provides a clear and quantifiable standard for assessing the impact of fouling on performance, offering a robust basis for fault diagnosis. This integrated approach not only enhances the early detection and monitoring capabilities for fouling faults in centrifugal compressors, but also provides a more reliable and interpretable tool to ensure operational safety and efficiency.