285

Aron Zahradka1, Owen Bunce2, Chad Sutherland3, Jared Kowis4, Canada.2, Canada.3, Canada.4

Aron Zahradka1, Sarah Newton¹, Owen Bunce2, Chad Sutherland3, Jared Kowis4
1Cambio Earth, Victoria, Canada. 2BGC Engineering, Calgary, Canada. 3Cambio Earth, Calgary, Canada. 4Enbridge, Houston, USA

Inertial Measurement Unit (IMU) bending-strain data play a critical role in detecting and assessing landslide impacts on pipelines. In line inspection (ILI) vendors often flag numerous pipeline bending-strain features based on IMU data; however, most identified features (>90–95%) are not associated with geohazards such as landslides. Manually reviewing these strain indications is labour-intensive and relies heavily on the experience and expertise of the analyst. This study describes a machine learning (ML) method developed to automatically screen and prioritize IMU bending-strain anomalies linked to landslide-related deformation. A Convolutional Neural Network (CNN) deep-learning classifier model was trained using IMU data from North American transmission pipelines, primarily in the Appalachian region of the USA. The dataset included over 30,000 strain feature training examples segmented into 200-meter windows, with 478 labelled as landslide impacted by pipeline geohazard experts. Model performance was evaluated using a hold-out test dataset consisting of 25% of the available dataset, including 112 landslide impacted strains. The model achieved 90% recall, 95% specificity, 95% accuracy, and an ROC-AUC of 0.96. In other words, it filtered out approximately 95% of benign features while capturing over 90% of landslide-related deformations. This research demonstrates the potential of ML models applied to IMU data which already has the potential to reduce analysis time per ILI run and improve consistency in identifying high-risk features. By rapidly identifying landslide related pipeline deformation, it helps operators allocate resources more effectively, respond faster to emerging threats, and maximize the value gained from their IMU data. The future integration of additional contextual data such as lidar and landslide mapping, along with increases in the training dataset size and diversity, are expected to yield substantial future improvements to model performance.

Read/Download PaperGradeAvailable after the event