Carbon fibre reinforced plastics (CFRP) are often used in civil engineering to strengthen steel structures,but various defects can affect the mechanical properties of CFRP-steel structures. Therefore,it is of great research significance to detect and identify various defects in CFRP-steel structures quickly and accurately. This study uses wavelet packet analysis to process the ultrasonic A-scan signals of steel specimens to identify the three most common types of defects in engineering,including,delamination and the steel plate cracking. The eight eigenvalues of the approximate coefficients and detail coefficients were extracted to construct the feature vectors;then,three algorithms,namely gradient descent optimization,quasi-Newton method and conjugate gradient descent optimization,were used to train the neural networks;finally,these neural networks were used to intelligently identify the three defects and the best neural network was selected. The results showed that the BP neural network model trained by the conjugate gradient algorithm has the best recognition accuracy of 93.75%.