Detection of Valve Stiction in Industrial Control Loops through Continuous Wavelet Transformation with a CNN

Control valve stiction is a common equipment problem where the valve exhibits delayed response to control output and becomes stuck due to static friction, which can lead to undesired nonlinear behavior and oscillations. It is critical to identify and correct this problem to ensure consistent operation in control loops. This paper introduces the novel technique continuous wavelet transform-convolutional neural network (CWT-CNN) for non-intrusive valve stiction detection. Industrial Process data is converted to an image with continuous wavelet transformation and then fed into a deep convolutional neural network to classify stiction behavior. The CWT-CNN is fine-tuned from pre-trained models like GoogleNet and ResNet via transfer learning for better classification and faster training while requiring less data. This work uses control loops from various chemical plants for training. The best performing CWT-CNN using GoogleNet can accurately predict 95.62% loops in the validation set, and has a true positive rate of 83.9% on the test set.