Identification of pipeline threat events in distributed fiber optic using MTF-PCA-based feature extraction techniques

Pipelines are vital in global energy industries, providing efficient methods for transporting natural gas and oil. Pipeline monitoring is crucial for detecting third-party intrusions and potential threats. Several approaches have been investigated in the field of pipeline threat detection, each offering distinct advantages and disadvantages. In contrast to conventional techniques, this work employs, to our knowledge, a novel strategy that uses both spatial and temporal data obtained from OTDR signals to identify third-party events in long-distance oil and gas pipelines. The proposed approach combines empirical mode decomposition (EMD), Markov transition field (MTF), and principal component analysis (PCA) to improve classification performance through machine learning techniques. Vibration signals of five typical threat events, including Hammer Digging, Hoe Digging, Excavator Digging, Human Jumping, and Human Walking are processed and transformed into two-dimensional MTF images by utilizing this approach, which are subsequently classified using machine learning models like backpropagation neural network (BPNN), random forest (RF), adaptive boosting (AB), and decision tree (DT). By integrating temporal and spatial characteristics, this approach improves the ability to distinguish between different third-party activities even if signals display slight deviations. Results indicate that all the classifiers perform well, achieving an accuracy of more than 95% while the backpropagation neural network demonstrated the highest classification accuracy across all event types, accomplishing an overall accuracy and F1 score of 99.10% and 99.49%, respectively. This shows the effectiveness of the proposed method for detecting complex events in long-distance pipeline safety monitoring, offering a swift and precise solution for pipeline surveillance applications.