PSI - Issue 64

Marco Pirrò et al. / Procedia Structural Integrity 64 (2024) 661–668 Author name / Structural Integrity Procedia 00 (2019) 000–000

667

7

4.4. Results and discussions

Fig. 6 shows the evolution in time of the MAE for all the four accelerometers, along with the 99 th percentile of the training MAE values as the threshold for outlier occurrence. It is clear that the structural change due to retrofitting are detected by the proposed methodology since non-negligible outliers occurrence is present before and during the retrofitting. The same results are confirmed by comparing the statistics of the MAE values (i.e., average and standard deviations) before and after the retrofitting (Table 2): the statistics are higher when the structural condition is different from the one accounted during training. Furthermore, Table 3 reports the coefficient of determinations computed between the reconstruction error and the environmental parameters (i.e., temperature and relative humidity): the R 2 values are very low (less than 0.1), meaning that EOVs are implicitly learned by the AE model during training.

Fig. 6. Mean Absolute Error (MAE) over the monitoring period.

Table 2. Average (MAE av ) and deviation standard ( s MAE ) of MAE.

Period

Index MAE av

Ch1

Ch2

Ch3

Ch4

1.290 0.532 1.503 0.559 2.728 1.129 2.762 1.045

1.315 0.535 1.476 0.647 2.643 0.981 2.378 0.798

1.348 0.528 1.547 0.644 2.764 0.953 2.393 0.805

1.240 0.491 1.392 0.522 2.651 1.096 2.761 0.981

Post-retrofitting (Training)

s MAE

MAE av

Post-retrofitting (Validation)

s MAE

MAE av

Retrofitting

s MAE

MAE av

Pre-retrofitting

s MAE

Table 3. Coefficients of determination R 2 between MAE and surface temperature, relative humidity.