基于孪生残差多尺度特征融合网络的方孔锁松动识别

Recognition for Square⁃hole Lock Looseness Based on Siamese Residual Multi-scale Feature Fusion Network

以人工模式为主的方孔锁松动异常识别存在检修效率低且难以保证检修质量的问题。为了提高方孔锁松动检修效率和检修准确率,设计一种孪生残差多尺度特征融合网络用于方孔锁松动识别。针对孪生残差网络中无法充分利用浅层多尺度特征的问题,利用一种特征融合模块(feature fusion module,FFM),对不同尺度下的特征进行自适应融合。提出一种数据增广算法模拟方孔锁松动故障,解决方孔锁实际松动数据数量较少的问题。测试集上的试验结果表明,这种增广算法能够明显提高方孔锁松动识别准确率,F-Score评价指标提高。相比于孪生残差网络,孪生残差多尺度特征融合网络具有更高的识别准确率,在不同松动角度的测试集上,准确率提升最高可达2.66%。

The abnormal identification of square⁃hole lock loosening based on manual mode has the problems of low maintenance efficiency and difficult to ensure the quality of maintenance.In order to improve the maintenance effi⁃ciency and accuracy of square⁃hole lock looseness,this paper propose a Siamese residual multi-scale feature fusion network to identify square⁃hole lock looseness.Aiming at the problem that the multi-scale features are not fully uti⁃lized in the Siamese residual network,a Feature Fusion Module(FFM)is designed to adaptively fuse the features at different scales.In addition,this paper proposed an augmented algorithm to simulate the square hole lock loose fault,which solves the problem of the small number of real loose data.The experimental results on the test set show that the proposed augmentation algorithm can significantly improve the recognition accuracy of the model,and the F-score was improved.Moreover,compared with the Siamese residual network,the proposed Siamese residual multi-scale fea⁃ture fusion network has higher recognition accuracy,and the maximum recognition accuracy improvement gain reaches 2.66%on the test set with different loosening angles.