PSI - Issue 38

L. Heindel et al. / Procedia Structural Integrity 38 (2022) 159–167

163 5

L. Heindel et al. / Structural Integrity Procedia 00 (2021) 000–000

(a) Fatigue test bench

(b) Hydro-mount

Fig. 2: Experimental data is measured using a 3-component servo hydraulic test bench (a) for the fatigue assessment of hydro-mounts (b).

prediction problem. The LSTM now learns to identify whether the prediction is valid and can be used directly for a given input data sample, or whether further adaptations need to be made.

3. Experiments

The proposed hybrid modeling strategies are demonstrated on a dataset from experimental fatigue analysis for both VS and FP tasks. Here, di ff erent measures of error are introduced to provide an overview of the respective strengths and weaknesses of each model.

3.1. Experimental dataset

A 3-component servo hydraulic fatigue test bench, depicted in Figure 2, is chosen as an example application. The system excitation is controlled by a drive signal, consisting of three inertia compensated force channels, which apply a load to an oil filled suspension hydro mount. For each channel, the inertia compensated force F comp is given by F comp = F measured − ζ · a measured , (7) where F measured is the measured force and a measured designates the measured acceleration. The factor ζ is determined from measurements at the shaking test bench without the test specimen, resulting in F comp = 0 for each time and fre quency. A parallel PID controller (Instron 8800) was used to realize the control loop. The force transmitting kinematic causes non-linear interactions between individual axis, while the oil filling introduces additional non-linear dampening to the system. Three force and three displacement sensors measure the system response. The static force-displacement relationship, shown in Figure 3, includes a friction based hysteresis. A dataset is created from a large number of measurements, using the presented test setup. This dataset is split into the categories training, validation and testing. Only the training dataset is used during the training process of the LSTM networks, while the model performance can be evaluated separately on the validation dataset. This enables the identi fication of suitable values for the hyper-parameters subsequence length L , overlap o , learning rate λ and the number of memory blocks and cells. When the parameterization is finished, the model performance is again evaluated using the testing dataset, which is completely independent from both training and validation. The training dataset consists only of measurements from synthetically generated random noise with varying o ff sets and amplitudes. While the validation and testing datasets still include some noise measurements, they are mainly com posed of serviceloads used during fatigue tests. These service loads consist of measured load profiles from di ff erent test components, which are adapted to the load capacity of the hydro-mounts in order to prevent damaging them. This