Pipeline infrastructure monitoring is critical for maintaining safety, efficiency, and operational integrity. Due to the high and rising value of energy products, external interferences such as acts of vandalism, damage, and theft are commonly observed in oil and gas pipelines. These actions are carried out through tapping, knocking, or drilling. The main objective is to identify the optimal combination of feature extraction techniques and machine learning based classification models for accurately recognizing different types of external interferences observed in oil and gas pipelines.
An experimental study was conducted on pipelines to assess classification performance across multiple external interferences. A comprehensive evaluation was conducted using various feature extraction techniques, which were subsequently applied to train five different machine learning classifiers.
The findings indicate that the model performance significantly varies with the choice of feature extraction technique. Spectrogram features demonstrated consistently high performance across different models. Feature extraction through Empirical Mode Decomposition (EMD) showed strong results in identifying interferences, except with the Support Vector Machine (SVM) model. K-Nearest Neighbors and Random Forest models exhibited superior performance, achieving high accuracy in most cases, specifically when trained with spectrogram and EMD features.
This highlights the importance of selecting appropriate features and models for effective pipeline monitoring. This work emphasizes the critical role of robust feature selection and model adaptation for accurate monitoring systems.
