Using one dimensional convolutional neural networks for classifying the vibration of process pipework

Abstract

Pipework in process facilities such as oil and gas or petrochemical plants is subjected to various dynamic excitations. Such excitations may include flow-induced forces due to turbulence or due to change in flow momentum (e.g., at reducers, ex-panders, elbows, etc.), shaking due to imbalances of rotating machinery, gas pulsations in closed branches and gas pulsations due to reciprocating machinery. Since process pipework usually conveys hot/cooled fluids, they cannot be rigidly fixed, and design codes (e.g., ASME B31.1) dictate a degree of flexibility of process pipework. The ensuing vibrations from these dynamic loads (overtime) may cause vibration-induced fatigue failures (VIF). It is in the operator's interest to quantify the risk of such vibrations as part of a risk-based inspection regime to ensure the plant's optimal and safe operations. Ideally, the risk of VIF is quantified by measuring strain. However, measuring strain in process plants can be challenging due to multiple reasons: the installation of strain gauges requires paint removal and surface preparation; pipework is often hot or insulated; the installation itself may be challenging physically if the pipework is excessively shaking. A popular alternative to strain measurement is to measure the vibration using widely available accelerometers and commercially available single channel analyzers. The motive here is that the velocity of vibration is correlated to the dynamic stress, and thus, vibration levels can be used as an indication of the risk of VIF. Current vibration acceptance criteria use the root-mean-square of the velocity and the dominant frequency of vibration to classify pipework vibrations into three categories: OK, CONCERN and PROBLEM levels. This paper uses one-dimensional (1D) convolutional neural networks (CNNs) to classify pipework vibrations. We use an experimental setup to generate vibration and stress signals in a pipe/shaker setup, and we pre-classify the stress signals into the three categories mentioned earlier. Wethen use the vibration time history signals to train the 1D CNN. Preliminary results show that the proposed CNN can classify the vibration data 95% of the time.