Ultrasonic nondestructive testing for composite bonded structures based on convolutional neural network and bidirectional gated recurrent unit (CNN-BiGRU) optimized by attention mechanism.

Abstract

New ceramic matrix composites (CMCs) are commonly used as thermal protection materials bonded to the surfaces of aircraft substrates. Defects in composite bonded structures can cause the protective layer to detach from the airframe, seriously endangering aircraft safety. Ultrasonic nondestructive testing is a promising method for detecting defects in multilayer bonded structures. However, the porous nature and strong sound absorption of CMC result in severe attenuation and scattering of ultrasonic detection signals, reducing the signal-to-noise ratio. This makes it challenging to accurately identify real defect signals. Therefore, a novel method combining a convolutional neural network and a bidirectional gated recurrent unit (CNN-BiGRU) optimized by an attention mechanism, along with ultrasonic inspection, is proposed to identify defects in composite bonded structures. The method learns time and frequency domain features of original signals through convolution, applies an attention mechanism to determine the importance of these features, and delivers weighted results to the bidirectional gated recurrent unit network. Then, time and frequency domain features are fused, and a one-dimensional global average pooling layer is employed to reduce model parameters and prevent network overfitting. A nonlinear support vector machine is utilized as the final classifier instead of the traditional softmax classifier. The results indicated that the proposed CNN-BiGRU model surpasses traditional classifiers that require manual feature extraction, achieving an accuracy of 97.70%. The method addresses the limitations of traditional techniques and provides a valuable reference for defect identification in composite bonded structures for practical engineering applications.