A Multisensor Feature Fusion Attention Convolutional Neural Network for Complex Magnetic Leakage
Xianming Lang, Ze Wang
Abstract
Among complex defects, small defects in oil and gas pipelines are easily submerged and difficult to detect. To improve the detection accuracy for small defects in complex magnetic flux leakage (MFL) signals, we propose a weak supervision method called multisensor feature fusion attention convolutional neural network (FACNN). First, an improved conditional adversarial generation network is presented, which introduces a supervised loss function at the analog signal level to reduce the number of parameter iterations for sample generation. Second, the feature extraction module uses the decoupled fully connected (DFC) attention mechanism and a convolutional neural network parallel structure to aggregate the features gathered at the center of the image and the features of the convolutional neural network, from which the small defect features can be fully extracted. Third, the feature fusion module uses the proposed loss function to guide the fusion of axial, radial, and circumferential signal feature maps, which enhances the effective propagation among small defect features. Finally, the experimental results show that the average detection accuracy of the proposed method for detecting small defects reaches 96.7%, which is 5.5% higher than the best detection accuracy of the existing methods.