Interpretable CNN Regression for Thermal Hotspot Localization Using Continuous Wavelet Transforms
Please login to view abstract download link
Structural Health Monitoring (SHM) is critical for ensuring the integrity of aerospace and civil structures. However, traditional damage localization methods often struggle with environ mental variations, such as thermal gradients, which alter ultrasonic wave velocity and cause signal misalignment. This paper proposes an interpretable Deep Learning framework for the precise lo calization of thermal hotspots in Al-7075 aluminum plates. Our approach utilizes the Continuous Wavelet Transform (CWT) to convert multi-sensor ultrasonic signals into 2D scalograms, cap turing non-stationary features in the time-frequency domain. A key novelty is the construction of ”panoramic scalograms”, which horizontally concatenate signals from 10 sensor channels to preserve global spatial context and enable the network to learn inter-sensor correlations. The re gression model employs a customized CNN architecture featuring asymmetric kernels and Global Average Pooling (GAP) to optimize feature extraction from these high-aspect-ratio inputs. To transition from a ”black-box” regressor to a physically consistent tool, we adapt the Grad-CAM framework for coordinate regression. Experimental results demonstrate high localization accuracy, achieving a Mean Absolute Error (MAE) of 1.86 mm for the X-coordinate and 1.43 mm for the Y-coordinate. Grad-CAM analysis confirms that the model’s predictions are driven by actual wave packets and reflections rather than numerical noise, validating the model’s adherence to elastody namic principles.