Convolution Neural Network for Defect Depth Estimation in Non-Linear Frequency Modulated Thermal Wave Imaging

Abstract Defect depth estimation in composite structures evaluated by active infrared thermography is a difficult task. Deep learning (DL) approaches are becoming more and more popular to automate the process and provide quicker insights. Since the performance of the DL model depends on the amount of noise contained in the raw thermal data under harsh inspection conditions, a proper denoising technique is necessary before feeding the raw data to depth estimation models. The one-dimensional convolution neural network (1D-CNN) for estimating the depth of defects in carbon fiber reinforced polymer (CFRP) examined using logarithmic frequency modulated thermal wave imaging is presented in this research. The noise is simulated by adding 5 and 10 dB of additive white Gaussian noise (AWGN) to the thermal profiles’ signal power. Furthermore, these thermal data are denoised using the recently developed stacked denoising autoencoder (SDCAE), which is thought to be a practical denoising method, prior to being fed into the 1D-CNN. The depth estimation performance of the suggested 1D-CNN is confirmed using a variety of performance indicators after it has been trained and evaluated on original, noisy, and SDCAE denoised thermal data separately. According to the results and performance measures, the 1D-CNN exhibits over 95% accuracy even in situations with high levels of noise, and it gets better in denoised cases and original thermal response. The slightly improved performance for denoised thermal data confirms that SDCAE is a good denoising method to use before estimating the depth of a defect.