Phased array ultrasonic inspection and automated identification of wrinkles in laminated composites

The wrinkles in composite structures degrade the mechanical properties and shorten the service life. They are difficult to be identified with no interlayer separation inside. This paper proposed to detect and localize wrinkles in laminated composites combining phased array ultrasonic inspection and deep learning. A convolutional neural network model was established to recognize them at different depths with the short-time Fourier transform of waveforms. The model training could be accelerated through simulation in PZFlex. The training and testing accuracy of 99.6% and 94.3% were achieved respectively. Relying on this model, both the simulation and experimental data can be predicted successfully with high efficiency. The reliability was further discussed based on confusion matrix. Such methodology simplifies the tedious analysis concerning low signal to noise ratio, non resonance matching and near field effect often confronted in ultrasonic inspection of composites. It gives a sound basis for efficient wrinkle detection automatically.

Automated summary

This paper proposes a method that combines phased array ultrasonic inspection and deep learning to detect and localize wrinkles in laminated composites. A convolutional neural network model is established to recognize wrinkles at different depths using the short-time Fourier transform of waveforms. The model training is accelerated through simulation in PZFlex, and achieves high accuracy in training and testing. This method simplifies the analysis of low signal to noise ratio, non resonance matching, and near field effect in ultrasonic inspection of composites, providing a reliable basis for efficient wrinkle detection automatically.