Classification of Compressors Stability Using an Image Recognition Machine Learning Model

Abstract This paper describes an application of Neural Networks for the classification of centrifugal compressor stability. The model was based on the pressure readings recorded at various positions of the throttling valve and for different locations inside the machine. Data were preprocessed and rearranged in the form of normalized grayscale images, representing a portion of consecutive pressure readings. This approach allowed for the implementation of image recognition techniques instead of standard artificial intelligence solutions. The models obtained were shown to be highly accurate, capable of predicting the opening area of the throttle valve. The best one achieved an error rate of around 1.6% and could operate on images created from 7569 consecutive readings. The paper also describes the results for multiple tested setups, presenting the metrics MSE (mean squared error) and R2 score (coefficient of determination) for different numbers of measurements, as well as for varying network architectures. The adoption of image recognition allowed for convenient visualization of the obtained results in the form of SHAP (SHapley Additive exPlanations). Their usage allowed for the investigation of the “reasoning” behind the AI model, determining which phenomena can be the most conclusively associated with each stability region.