Classification of Process Pipework Vibration Using Machine Learning

Abstract

This paper aims to use a deep convolution neural network (CNN) to classify pipework vibrations. Vibration levels are a good indicator of the risk of vibration-induced fatigue failures (VIF). Vibrations in pipework are very common in oil and gas and petrochemical plants which are a result of flow-induced forces from turbulent flows in the pipe, flow momentum change, or fluid pulsations. Strain measurements are usually used to identify and quantify the risk of VIF; however, that is not always easy due to the surface operations required before the installation of the strain gauge. Another alternative would be to rely on vibration measurements using accelerometers and signal analyzers. The dominant frequency of vibration and the root mean square of the velocity are currently used for the vibration acceptance criteria to categorize the vibration levels into three categories: OK, CONCERN, and PROBLEM. The motivation here is to use collected field data for both vibration and stress measurements for multiple pipework plants to train a deep CNN on the images of a continuous wavelet transform from the time-frequency features to be used to predict the classification of new data based on the vibration levels. The results show that the developed CNN can successfully classify the vibration data 93% of the time.