Quantitative assessment of pipeline defects utilizing a dual-stage deep learning framework: Integration of pretrained YOLO network and multi-input parallel convolution architectures on magnetic flux leakage data

As long-distance oil pipelines near the end of their operational tenure, the propensity for leakage due to localized defects markedly increases, necessitating the imperative for systematic inspection and sustained maintenance efforts. Magnetic Flux Leakage (MFL) inspection, a mainstream non-destructive testing methodology, has been extensively adopted. In light of the voluminous nature of monitoring data, deep learning and computer vision technologies play a pivotal role in enhancing the efficiency and accuracy of detection. This study introduces an innovative cascading detection technique that amalgamates the advanced visual recognition network You Only Look Once (YOLO) v8 with a novel multi-input parallel convolution structure. Through channel fusion-based image preprocessing techniques, it adeptly utilizes tri-axial MFL experimental data to precisely localize pipeline defects, while concurrently predicting the axial length and depths of pipeline defects. This research meticulously investigates the impact of various data processing techniques and model architectures on the accuracy of defect recognition and their quantifiable prediction. Following stringent experimental validation, our data preprocessing method has demonstrated superiority over conventional approaches in defect detection and quantitative assessment tasks. Moreover, the proposed multi-input parallel convolution significantly outperforms conventional single-input prediction networks in quantifiable predictive accuracy concerning defects, demonstrating its potential in enhancing the predictive maintenance of long-distance oil and gas pipelines by enabling more precise, timely, and cost-effective interventions.