Issue 46

F. Bazzucchi et alii, Frattura ed Integrità Strutturale, 46 (2018) 400-421; DOI: 10.3221/IGF-ESIS.46.37 417 bonded condition. At the contrary, red-brow stains (Fig. 22(a)) suggest unprotected cables and rust transport by continuous water gradient. This latter condition represents of course the worst scenario, with maximum damage and lower robustness. The presence of this particular damage has been catalogued as Lv.4 (red) for the entire bridge. Extended and joint-diffused white stains have been catalogued as Lv.3 (orange), while local or isolated as Lv.2 (yellow). Lastly, a clear external surface of the bridge deck has been assigned to a Lv.1 degree of damage (green). After the first photographic campaign, an image classification (labeling) has been carried out manually by experts. Results of this classification are shown in Fig. 22c. These results have been used as ground truth for the building of the image layer of a DCNN CV detector Fig. 23(a). The set has been splitted in 65%-35% for training and validation. Before training, each picture has been filtered for red-contrast enhancement. A random set of 200 images from Kaggle has been used as false test. Training lasted 21 hours on a 2014 MacBook Pro. DCNN has been implemented with Pytorch and Anaconda 3.5.2.0. Other two layers have been implemented in the net, one based on the results of the load test and one on the dynamic properties of the bridge, both computed by a FEM tuned model Fig. 23(b). To date, other two photographic campaigns (tablet-based) have been carried out by A.N.A.S. technicians, and Fig. 24(a) shows the results of the application of the DCNN on the bridge n. 48. The system correctly spotted a red stain and classified the bridge as a Lv.3. Analogously, for bridge n. 52 (Fig. 24(b)), a Lv.4 classification has been successfully matched. Regarding instead span n. 85, classified Lv.3 manually, the DCNN detected a Lv.2. damage degree Fig. 25(a). Figure 23 : DCNN scheme (a) ; FEM tuned model (b) . Figure 24 : Detection of a Lv.3 damage (a) ; detection of a Lv.4 damage (b) .