{"title":"Attention and deformable convolution-based dual-task high-precision fault recognition","authors":"Zhen Peng , Danping Cao , Huiqun Xu , Dan Zhu","doi":"10.1016/j.acags.2025.100267","DOIUrl":null,"url":null,"abstract":"<div><div>Deep learning has been widely applied in fault recognition task. However, current two-dimensional (2D) deep learning-based training methods fail to adequately consider the overall spatial characteristics of faults, resulting in discontinuous fault recognition results and unable to achieve the effect of three-dimensional (3D) deep learning training methods. To address this issue, we propose an attention and deformable convolution-based dual-task high-precision fault recognition method (ADTFM), which introduces a dual-task deep learning network architecture within the 2D training framework, effectively improving the fault recognition accuracy and reliability. ADTFM consists of two tasks (Main task and Auxiliary task) with the same network structure based on the deformable convolution operators and U-Net. The main task uses the Inline direction for training, and uses the deformable convolution operator to capture more accurate fault feature. At the same time, the auxiliary task is trained in the Time-slice direction, and the features generated by auxiliary task direction are transferred to the main task in training process. The two tasks are connected through the attention mechanism, so as to increase the spatial characteristics of faults in 2D training process, and effectively compensate for the spatial limitations of 2D training. By testing the public 3D datasets and the field 3D datasets, and comparing with the current high-precision FaultSeg3D fault recognition method, the results show that our method can improve the accuracy of fault recognition. Moreover, through the quantitative evaluation of computing consumption time and memory, it is shown that the proposed method effectively reduces the computational complexity and decreases the consumption of computational resources, and provide a more efficient solution for fault recognition task.</div></div>","PeriodicalId":33804,"journal":{"name":"Applied Computing and Geosciences","volume":"27 ","pages":"Article 100267"},"PeriodicalIF":3.2000,"publicationDate":"2025-07-23","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":null,"platform":"Semanticscholar","paperid":null,"PeriodicalName":"Applied Computing and Geosciences","FirstCategoryId":"1085","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S2590197425000497","RegionNum":0,"RegionCategory":null,"ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"COMPUTER SCIENCE, INTERDISCIPLINARY APPLICATIONS","Score":null,"Total":0}
引用次数: 0
Abstract
Deep learning has been widely applied in fault recognition task. However, current two-dimensional (2D) deep learning-based training methods fail to adequately consider the overall spatial characteristics of faults, resulting in discontinuous fault recognition results and unable to achieve the effect of three-dimensional (3D) deep learning training methods. To address this issue, we propose an attention and deformable convolution-based dual-task high-precision fault recognition method (ADTFM), which introduces a dual-task deep learning network architecture within the 2D training framework, effectively improving the fault recognition accuracy and reliability. ADTFM consists of two tasks (Main task and Auxiliary task) with the same network structure based on the deformable convolution operators and U-Net. The main task uses the Inline direction for training, and uses the deformable convolution operator to capture more accurate fault feature. At the same time, the auxiliary task is trained in the Time-slice direction, and the features generated by auxiliary task direction are transferred to the main task in training process. The two tasks are connected through the attention mechanism, so as to increase the spatial characteristics of faults in 2D training process, and effectively compensate for the spatial limitations of 2D training. By testing the public 3D datasets and the field 3D datasets, and comparing with the current high-precision FaultSeg3D fault recognition method, the results show that our method can improve the accuracy of fault recognition. Moreover, through the quantitative evaluation of computing consumption time and memory, it is shown that the proposed method effectively reduces the computational complexity and decreases the consumption of computational resources, and provide a more efficient solution for fault recognition task.