Ultrasonic Signal Recognition Method Based on Single Class Attention Loss Function

Accurate identification of ultrasonic signals can effectively improve the accuracy of defect detection and inversion. However, ultrasonic signals with the different numbers of echoes in ultrasonic NDT data show extremely high similarity in high-dimensional space due to the influence of echo feature time shift and noise interference, thus affecting the classification accuracy. To address the above issue, we introduce a method of ultrasonic signal recognition based on Single Class Attention (SCA) loss function. First, SCA loss function gathers the feature vectors of positive class to an anchor point in the embedding space, and moves the feature vectors of negative class away from the anchor point. It makes the neural network focus on a single class of data in the feature extraction process, and enhances the feature extraction ability of the network for a specific class of data. In addition, we design a feature extraction network composed of three subnetworks, and use the proposed SCA loss function to optimize the network to extract more distinguishable features. The ultrasonic signal collected in real environment is adopted in the experiment, and the experimental results show that our method outperforms the other baseline methods in ultrasonic signal recognition.