Session: 19-03 Compressors and pumps

Paper Number: 154137

Classification of Compressors Stability Using an Image Recognition Machine Learning Model 

This paper describes an application of neural networks for classification of a centrifugal compressor stability. The model was based on the pressure readings recorded at several throttling valve positions, and for different locations inside the machine. Data was preprocessed and re-arranged to the form of normalised grayscale images, representing a chosen number of consequtive pressure readings. Such approach allowed for implementation of image recognition as the chosen Machine Learning technology. Convolutional Neural Network (CNN) were chosen as this approach is known to perform well the
image recognition tasks. The obtained models were proved to be highly accurate, being able to predict the throttle valve opening area, with an average error around 1.6%, using 7569 consequitive readings. The paper describes also the results for multiple tested setups, presenting the Meas Square Error (MSE) and R2 score metrics for different numbers of measurements, as well as for varying network architectures. The adoption of image recognition allowed for convinient visualization of the obtained results in the form of SHAP (SHapley Additive exPlanations) maps. Their usage allowed to investigate the ”reasoning” behind the AI model, finding out which phenomena can the most conclusively be associated with each stability region. It was observed that point located above the blade was especially efficient to identify current impeller operating range, and the tip pressure variations served as a good reference for the network.

Presenting Author: Paweł Płuciennik Lodz University of Technology

Presenting Author Biography: Paweł is a student at Modeling and Data Science course in Lodz University of Technology. He finished his bachelor with distinctions, as one of the best students in the year.

Authors:

Paweł Płuciennik Lodz University of Technology
Magdalena Jaśkiewicz Digica
Grzegorz Liskiewicz Texas A&M