Assessing the risk of vibration-induced fatigue in process pipework using convolutional neural networks

This study develops convolutional neural networks (CNNs) to classify pipework vibration states in process plants, aiming to assess the risk of vibration-induced fatigue (VIF). A major challenge in VIF assessment is the need for strain measurements which, while ideal for assessing the risk of VIF, require direct installation of strain gauges on the pipework, deployment of a data acquisition system, and postprocessing and analysis of the measured data. In contrast, vibration data can be efficiently collected using accelerometers and single-channel data loggers, providing a more feasible solution for initial screening. Current vibration acceptance criteria rely on predefined thresholds to classify vibration levels into three categories: OK, CONCERN, and PROBLEM, based on the dominant vibration frequency and root mean square (RMS) velocity. However, these criteria may not fully capture the complexity of VIF mechanisms. Strain measurements are typically performed when vibration levels fall into the CONCERN or PROBLEM categories. Stress levels calculated from strain measurements are also classified into OK, CONCERN, and PROBLEM categories according to industry standards. To address these challenges, this study explores using CNN-based models for automated VIF risk assessment. Several CNN architectures using time-domain acceleration data and continuous wavelet transform (CWT) images are evaluated to identify the most effective approach for classifying vibration states. The tested two-dimensional (2D) CNN architectures include a 2D-CNN trained on CWT images, a 2D-CNN trained on CWT, RMS, and Kurtosis features combined into red–green–blue (RGB) images, and a ResNet-50 model. These approaches are benchmarked against a one-dimensional (1D) CNN trained directly on raw acceleration data. The (1D- and 2D-) CNN models were trained on data obtained from multiple operational plants across different countries, ensuring a diverse dataset representative of real-world conditions. The results show that the 1D-CNN, trained on field data and tested on unseen samples, achieves the highest accuracy (98%) with minimal training and inference times, outperforming the other models. These findings demonstrate the capability of CNNs to autonomously learn discriminative vibration patterns, eliminating the need for predefined criteria or additional strain-based measurements. The proposed 1D-CNN model offers a robust and computationally efficient solution for real-time VIF risk assessment, enabling automated structural health monitoring in industrial settings.