Ultrasonic Metal Recognition Based on a Time-Frequency Feature Multi-Scale Fusion Network

When identifying metals with the same appearance and material non - destructively, the existing ultrasonic metal identification methods often overlook unknown - class identification, have poorly constructed data sets and limited practicality. This paper proposes an ultrasonic metal identification method based on the multi - scale fusion network of time - frequency features. It uses the multi - head multi - scale residual convolution module to extract and fuse time - frequency features, and combines the attention mechanism to process the fused features, which overcomes the over - fitting problem of a single time - domain feature and improves the generalization ability. An independent data set including training, validation and test sets is constructed. The training and validation sets each contain ultrasonic data of 12 test blocks, and the test set contains 13 classes (the 13th class is 5 test blocks of unknown categories). The test - set data are not used in model training to simulate actual scenarios. Experiments show that this method can efficiently identify metals in the self - built data set, and its effectiveness in classifying long sequences is verified on the UCR data set.