PSI - Issue 64

Philipp Kähler et al. / Procedia Structural Integrity 64 (2024) 1248–1255 Kähler / Petryna / Structural Integrity Procedia 00 (2019) 000 – 000

1251

4

2.2. Testing and training the CNNs To test this cluster structure, a numerical example was created. For this purpose, the FE model of a two-span highway bridge with a length of 30 m and two traffic lanes from Fig. 3 is subjected to several virtual vehicle crossings.

(b)

(a)

Fig. 3. (a) FE-model of the highway bridge with moving loads, (b) artificial acceleration sensor positions.

The entire loading history is divided into different loading scenarios, e.g. single or multiple vehicles and loads on the first lane, the second one or on the both lanes. Each crossing is simulated dynamically as a moving vertical load with constant magnitude and constant horizontal velocity. A sample of 5000 randomly generated vehicle crossings was created for each load case. The load magnitude represents large trucks and was sampled out of a gaussian distribution with btr = 34 t, btr = 0,05 ⋅ btr . The load velocity is assumed to be uniformly distributed between ̇ btr = 80 km/h and ̇ btr = 110 km/h . Acceleration data, gathered from twelve different places on the bridge, is considered as artificial measurement data and then used to train the cluster structure. The data portion for training, validation and testing purposes is equal to 80 % , 10 % and 10 % , respectively. The output from the decision-making CNNs comprises simple integer values. By classifying the measured response of the bridge into predefined groups, the determination of the number of vehicles on the bridge and the differentiation between occupied traffic lanes was accomplished, aiding in navigating to the correct feature-extracting CNNs within the cluster structure. Training these feature-extracting CNNs to extract the load characteristics is considerably intricate, necessitating a greater number of epochs, particularly for handling output vectors involving multiple vehicles on the bridge. Nonetheless, a notably strong correlation between the reference data and the predicted data is achieved, which leads to maximum root mean square errors for the test dataset of u̇ =5 % in the load velocity and mag = 6.8 % in the load magnitude. The impact of different levels of measurement noise was also investigated, which led to a significant loss in the accuracy of the CNNs and consequently necessitated the application of noise-reducing filters. 3. Building a digital twin In Grieves (2015), a digital twin (DT) is defined as a combination of a real structure, whose behavior is measurable, a numerical model capable of making predictions for this behavior, and a set of methods capable of reconciling the model predictions and the real system behavior. Thus, the DT should be able to reproduce the real behavior of the physical system and reflect any system changes through changes of model parameters. Data assimilation techniques are suitable to handle this task, even when the measurement data is affected by noise and the numerical model is not perfect but rather subject to assumptions. In this contribution, a combination of two ensemble-based KF will be used for the continuous data assimilation.