Vibration-Based Fault Diagnosis in Pipelines Using AI

This work presents a comprehensive simulation-based approach for detecting defects in steel pipelines using vibration analysis, combining two methodologies: linear regression and artificial neural networks (ANNs). In both studies, pipe models were developed in ANSYS, featuring 121–130 variations with single-hole defects at different longitudinal and circumferential positions, along with a defect-free control pipe. Modal analysis was conducted to extract the first 15 natural frequencies for each model, which were all below 5 kHz to ensure compatibility with real-world sensor systems. In the first approach, a linear regression model was developed using MATLAB, where defect localization was based on the natural frequency differences between the control pipe and the defective models. The model achieved excellent performance across 70/30 and 80/20 training-test data splits, with R2 values exceeding 0.998 and mean absolute percentage error (MAPE) below 5%. Generalization tests on unseen pipe configurations confirmed the robustness and reliability of the method. In the second approach, three ANN architectures—feedforward, cascade-forward, and fitnet—were trained using the same frequency signature inputs. All models employed two hidden layers with 10 neurons each and were optimized using the Levenberg-Marquardt algorithm. Among them, the feedforward network showed the highest accuracy, especially in predicting fault positions on external test sets, with errors below 1% in some cases. The combined findings demonstrate that both linear regression and ANN models can accurately identify defect locations in pipelines using modal frequency data. These results support the use of low-frequency vibration analysis in developing real-time, cost-effective, and non-destructive structural health monitoring (SHM) systems. Future work should focus on physical validation using embedded sensors to extend these promising simulation results into practical industrial applications.