NEW EXPLAINABILITY METRICS TO ASSESS AI-BASED RECOGNITION OF TYPICAL DEFECTS IN EXISTING RC BRIDGES

The paper presents a simple methodology to assess the effectiveness of different machine-learning techniques to automatically detect typical defects in existing reinforced concrete (RC) bridges. The work is framed in the topic of infrastructures safety, for which in the last years a continuous interest and different methodologies have been developed towards the monitoring of the health state of existing bridges and viaducts. In this view, one of the most important phases to develop consists in the visual inspections of the defectology of each structural element composing bridges (e.g., piers, deck), which requires the work of several expert engineers that periodically assess these structures and surveys the evolution of existing defects. To support this time-consuming task, systematic and cost-saving protocols are required. To this scope, new technologies as machine- and deep-learning approaches can be employed for a continuous control of defects in existing bridges and their evolution, paying attention to the effectiveness of the automated prediction. With this goal in mind, this work investigates the reliability of automated predictions of typical defects in RC bridges, by means of artificial intelligence-based (AI-based) methods. First, a database of photos reporting defects was collected, and the most recurrent defect typologies were classified by domain experts. After, several existing architectures of convolution neural networks (CNNs) were trained and tested on the dataset. Still, in order to physically interpret the results, class activation maps (CAMs) techniques were applied on the output, in order to visually assess the reliability of each CNN. From the obtained activation zones, two novel metrics were proposed, aimed at quantifying the capability of each CNN to predict defects. The results show that in some cases, even CNNs with high accuracy can fail the prediction, which raises the necessity of involving eXplainability methods, especially when working on sensitive topics as infrastructure safety.