Automatic multi-anomaly detection of pipelines with ensemble deep learning-based computer vision

Pipelines are critical infrastructure playing critical roles in ensuring public safety and economic development. Pipeline inspection is an important way to early detection of anomalies for proactive and preventive measures to improve safety while minimizing maintenance expenses. However, traditional methods that employ pipeline inspection gauges and manual data analysis have inefficiency in cost and time. This paper presents an approach for automatic detection of four classes of pipeline anomalies via analyzing videos captured by optical cameras in pipeline inspection toolset. The approach uses a framework that integrates advanced generative adversarial networks, image pre-preprocessing, and ensemble deep learning techniques to address data scarcity challenges, achieving high accuracy and high efficiency. The effects of dataset size and data balance on the performance of ensemble learning models are evaluated. The approach has been validated using outdoor testing data unseen to the ensemble learning model, demonstrating high generalizability in detecting corrosion, crack, deposit, and intrusion.