This paper proposes the Contribution by Gradients (C-grad) method for interpreting neural network models applied to regression problems on ultrasonic data. As machine learning applications continue to expand in Non-Destructive Testing (NDT), the “black-box” nature of neural networks raises concerns about the consistency and interpretability of AI-generated solutions in industrial applications. The C-grad method addresses these challenges by quantifying input contributions through their rate of change across layers, providing detailed layer-by-layer insights into model decisions. Unlike traditional gradient algorithms focused solely on final outputs, C-grad backward propagates activation changes from intermediate layers to input arrays, enabling a more detailed breakdown of feature importance at each layer and enhancing interpretability. Applied to 1-dimensional convolutional neural networks (1D-CNNs) for ultrasonic A-scan time-traces data in corrosion profiling, the method shows superior stability and consistency over the traditional Gradient-based Class Activation Map method (Grad-CAM). By combining explanation infidelity testing with a peak-to-peak distance metric that correlates model explanations with physical echo features, C-grad offers a comprehensive framework for assessing model robustness to corrupted inputs. C-grad has proven successful in identifying crucial physical features in the A-scan time-traces for the 1D-CNNs trained on two corrosion profiling metrics: mean material thickness and roughness. These case studies under C-grad investigation provide detailed insights into model trustworthiness and guide architecture optimisation in ML-driven NDT.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号