低信噪比环境下超声细微缺陷特征提取的协同增强网络方法OA

To address the challenge of extracting subtle ultrasonic defect features in low signal-to-noise ratio environments,this study proposes a gated residual and dual-attention collaborative enhancement network for low SNR ultrasonic signals.Based on convolutional neural network,the model adopts a'residual block squeeze-excitation(SE)module-pooling'cascaded structure:a standard SE module is embedded in the residual block for initial channel screening,a locally enhanced SE module is used at the end of network stages to focus on peak signals,and gated residual connections are employed to dynamically preserve original subtle features,thus realizing collaborative optimization of noise suppression and feature enhancement.Experimental results show that the improved model achieves a mean root mean square error(RMSE)of 0.068 3 and a mean absolute error(MAE)of 0.047 1,which are 49.7% and 41.7% lower than those of the baseline CNN,respectively.It also outperforms models with only a single attention mechanism or residual blocks,verifying the superiority of dual-mechanism collaboration,while exhibiting excellent training stability and maintaining high accuracy in low SNR environments.In conclusion,the proposed model effectively overcomes the bottlenecks of noise interference and subtle feature learning.Its prediction accuracy,anti-interference capability,and stability are significantly superior to traditional methods and existing models,providing an efficient technical solution for ultrasonic non-destructive testing of steel pipes with important industrial application value.