Research on Ultrasonic NDT of Wire to Terminal Joints: Comparison of Combinations of Various CNNs and Signal Processing Technologies

Abstract

The wire to terminal joints are prepared using ultrasonic welding and find extensive application in various fields, such as new energy vehicles and aerospace. Traditionally, tensile strength tests have been employed for welding quality inspection. However, this study proposes an automatic nondestructive evaluation scheme to overcome the inefficiency and destructiveness associated with tensile testing. To achieve this, a 5 MHz/32-element array ultrasound probe is utilized for ultrasound detection and signal acquisition from two groups of joints categorized as OK (good quality) and NG (poor quality) based on their welding quality. Signal processing techniques including short-time Fourier transform, wavelet transform, and Gramian angular field are applied to convert one-dimensional time series into two-dimensional signal feature maps. Convolutional neural networks such as VGGNet, ResNet, DenseNet, and MobileNet are utilized for the classification of two-dimensional signal feature maps. The comprehensive evaluation of different feature maps and combinations of neural networks is conducted from various perspectives including network complexity, recognition accuracy, memory consumption, and inference time. The study findings indicate that wavelet transform feature maps achieve the highest accuracy across diverse neural networks, reaching up to 95% accuracy in VGGnet13 despite higher associated costs. In MobileNet-Small and ShuffleNet-V2 networks, the accuracy stands at approximately 85%, accompanied by faster inference times and lower costs. Considering all factors holistically, the combination of wavelet transforms feature maps with MobileNet and ShuffleNet demonstrates superior cost-effectiveness and suitability for ultimate deployment and application on mobile devices facilitating automated non-destructive assessment of wire to terminal joints quality.