EfficientNet-b0-Based 3D Quantification Algorithm for Rectangular Defects in Pipelines

Abstract

Quantitative analysis of the magnetic leakage signal is crucial for evaluating the magnitude of pipeline damage after fault detection. To address these issues, it is necessary to establish a correlation between the leakage signal and the magnitude of the defect to make accurate predictions through regression analysis. This research introduces EffiTriDimNet (ETDN), a multi-task convolutional neural network that combines one-dimensional pipeline defect leakage detection data into a unified feature map while simultaneously measuring the three-dimensional characteristics of the defects. This multi-task learning model utilizes a shared feature extraction layer and numerous branching networks to create individual predictions for each task: length, width, and depth. The network uses the EfficientNet structure for feature extraction, while the training and validation process utilizes the PyTorch Lightning framework. Experimental results demonstrate that when compared to other algorithms, ETDN outperforms them on the identical test set, exhibiting average absolute errors of 0.1920, 0.1809, and 0.2904 for length, width, and depth, respectively.

Keywords

• Pipeline defect prediction
• non-destructive testing
• EFFICIENTNET-b0
• feature extraction
• regression analysis
• PyTorch lightning framework