[IMC 2023 ] Automated Explanations To Enhance Deep Learning Models in Power Grid Monitoring 

By Giovanni Floreale, Politecnico di Milano
Drones are emerging as a viable technology for monitoring infrastructure conditions. They allow cost-effective and safe condition monitoring, eliminating the need for personnel to access dangerous and remote areas.
However, their effectiveness depends on the quality of automatic image processing algorithms. Deep Learning (DL) models, have demonstrated high performance in detecting defects and deteriorating conditions in images of infrastructure components. Yet, these models can exhibit bias and achieve accurate results through non-causal shortcuts. Various explainability techniques have been proposed to identify the elements of the input that contribute the most to the DL model’s output. Nonetheless, evaluations of explanations often rely on manual assessment by domain experts. As the volume of captured data increases, this manual burden can become significantly overwhelming.
In this study, we propose a novel framework for the automated processing of explanations from a supervised DL model. The proposed framework demonstrates its ability to identify misclassifications, shortcuts and non-convincing explanations. This capability proves valuable in several ways: pinpointing directions for improving DL model’s by suggesting increased data collection under certain external conditions, recommending modifications to the model itself to mitigate non-causal classification strategies, and guiding experts in reclassifying incorrect outcomes to enhance the accuracy of the DL model.