A Novel Multisensor Orthogonal Attention Fusion Network for Multibolt Looseness State Recognition Under Small Sample

Abstract

Bolts are common and effective fasteners for connecting steel structures, and bolt looseness detection is essential to ensure structural reliability and equipment safety. However, due to the complex spatial structure and multi-bolt of pipeline flange bolt connections, it is difficult for existing methods to implement fast and precise detection with single-sensor information and scarce data. To overcome these challenges, a novel multi-sensor orthogonal attention fusion network (MsOAFN) is proposed for multi-bolt looseness state recognition under a small sample in this paper. MsOAFN consists of feature extraction, feature fusion and state classification. A sensor-to-channel (STC) multi-scale convolution is used to extract features of different sensors to avoid data conflicts among sensors. In addition, a feature self-calibration attention module is built to expand the receptive field, alleviate information loss and enrich depth features during the down-sampling process. Then, a lightweight orthogonal attention fusion module is constructed, which fuses the shallow and depth features to further supplement and enhance feature information. Finally, the fusion features are input into the classification module to finish bolt loosening state recognition. Two case studies were conducted on a pair of typical flanged multi-bolt connection to verify the effectiveness of the proposed method, with an average accuracy of 99.77% and 98.84%, respectively. The superiority of this method in loosening detection of flanged bolt connection structure is compared with other algorithms.