PSI - Issue 64

Christoph Brenner et al. / Procedia Structural Integrity 64 (2024) 1240–1247 Christoph Brenner et al. / Structural Integrity Procedia 00 (2019) 000 – 000

1246

7

Subsequently, the OCTs trained with the model database are employed to predict the optimal parameters for the cracked bridge scenario using the corresponding measurement data. To measure the accuracy of these predictions, the numerical model is assessed using the predicted parameters. Following this, the root mean squared error (RMSE) is computed to quantify the deviation between the simulated strain values and the exact measured values. These results, along with those for the initial model, are depicted in Fig. 4. As illustrated in Fig. 4, there are notable discrepancies in the predicted parameters between the two classification methods. While the actual damage should predominantly affect parameter four, adjacent areas also experience a reduction in stiffness. Relative to the multi-classification tree approach, the individual OCTs tend to forecast lower parameter values and consequently predict smaller stiffness reductions. Despite this, both models exhibit a reduced RMSE in the SHM data compared to the initial model, with the multi-classification method performing notably better. 4. Discussion The preceding section detailed the application of the data-driven method to a large-scale steel bridge equipped with a global SHM system, focusing on a local damage scenario. In contrast to prior studies by Kapteyn et al. (2022) or Svendsen et al. (2023), this study does not assume extensive damage resulting in up to 80% reduction in stiffness or component failure. Instead, it examines the effects of a discrete crack of relatively small dimensions. The findings suggest that the resultant stiffness reduction is more likely to fall within the range of a hundredth of a percent rather than double-digit percentages. Using SHM data from the cracked model, the prediction capability of OCTs is evaluated. What sets this study apart from existing literature is the absence of identical or similar damage states within the model database used to train the OCTs. Instead, the model selector encounters a completely novel, unknown damage state for which there is no exact match in the database. Another novel aspect of this study is the investigation of a large-scale structure with a multitude of closely related parameters that may interact directly. Results obtained from a model with six parameters indicate that individually predicting each parameter, as done in previous studies e.g. by Kapteyn et al. (2022), is inappropriate due to the inability to account for parameter interactions. Better results are achieved through multi-classification trees, but the accuracy for both training and test datasets remains modest. A significant insight from the study is the impracticality of employing an "all-in-one" approach for large-scale structures with numerous parameters. For a model with a high two-digit number of parameters, very large and complex trees would be required to cover all suitable parameter combinations. This underscores the necessity of either adopting a staggered approach or integrating engineering expertise with semantic data to preselect sub-areas of interest. Neglecting parameter interactions within these selected areas is infeasible due to their spatial proximity and mutual influence. Furthermore, simultaneous prediction of multiple parameters proves computationally demanding. The study also observes a notable decline in the interpretability of OCTs with increasing tree depth, particularly in multi-classification tasks. Therefore, it is recommended to limit the tree depth to around five to six for better interpretability. However, despite these efforts, the overall accuracy of developed OCTs remains modest, prompting further investigation into potential limiting factors. Several factors contributing to this limitation are identified: • Limited dataset size: The relatively small size of the model database may constrain the model's ability to adequately capture complex patterns. • Feature selection considerations: The efficacy of a feature selection strategy in enhancing predictive performance needs to be investigated. • High parameter correlation: The high correlation among parameters could introduce redundancy or confounding effects, potentially complicating the prediction process. Addressing these limitations requires further attention. Expanding the dataset through additional data acquisition efforts or incorporating data from similar structures could improve model generalization. Additionally, exploring