Recognizing defects in stainless steel welds based on multi-domain feature expression and self-optimization

The recognition of different welding defects is important for the assessment of the safety of welded structures. This study proposes a method to recognize defects in stainless steel welds based on multi-domain feature expression and self-optimization. This is because of the poor detection of feature-related information from signals, inadequate feature extraction by the convolutional network, the limited capability of intelligent techniques, network redundancy, and a lack of self-adaptive capability in prevalent methods, A 1D ultrasonic detection dataset of austenitic stainless steel welds in the time domain (TD) is first constructed and the 1D TD signals are rendered in the time–frequency domain, Gramain angular field, and the Markov transition field. The aim is to enrich the feature expression of 1D ultrasonic echo data of the weld defects. A comparison among a variety of lightweight convolutional neural networks on multi-spatial domain datasets is used to identify a combination of network and spatial domain datasets that are suitable for recognizing welding defects. Finally, a multi-scale depthwise separable convolution is designed, and is subjected to adaptive compression and parameter-adaptive optimization based on the sparrow search algorithm to construct the self-optimizing lightweight multi-scale MobileNetV3 (SLM-MobileNetV3) model. The results of experiments showed that the SLM-MobileNetV3 model has an accuracy of 97.26% for the recognition of five types of defects: inclusion, crack, porosity, incomplete penetration, and a lack of fusion. The time required for testing a single image was 2 ms. Experimental analysis showed that the proposed method can improve the accuracy of defect recognition while using few parameters and computational cost, and is generalizable.