PSI - Issue 38

L. Heindel et al. / Procedia Structural Integrity 38 (2022) 159–167 L. Heindel et al. / Structural Integrity Procedia 00 (2021) 000–000

166

8

10

0.20

0.5 1 2

0.15

RMS

0.10

0.3 0.5 0.7 1.0 0.5- 0.5 + Scaling FP Displacement FP Force VS Force Fig. 4: FRF, LSTM and hybrid models are compared on all serviceload testing data files, which are rescaled from two independent signal shapes and sorted by the respective scaling factor. A scaling of 0.5- denotes that a negative o ff set was applied to the data with scaling factor 0.5, while 0.5 + denotes a positive o ff set. The system sti ff ness changes significantly with an increased scaling factor or when an o ff set is used. The RMS errors are averaged over the predicted signal channels. 0.3 0.5 0.7 1.0 0.5- 0.5 + Scaling 0.1 Multi-Rain ratio 0.3 0.5 0.7 1.0 0.5- 0.5 + Scaling 0.05 0.10 0.15 0.20 0.25 RMS 0.3 0.5 0.7 1.0 0.5- 0.5 + Scaling 0.1 0.5 1 2 10 Multi-Rain ratio FRF Only LSTM Hybrid 1 Hybrid 2 0.3 0.5 0.7 1.0 0.5- 0.5 + Scaling RMS 0.1 0.2 0.3 0.3 0.5 0.7 1.0 0.5- 0.5 + Scaling 0.1 0.5 1 2 10 Multi-Rain ratio

4. Conclusion

This paper presented a novel hybrid approach for signal estimation in non-linear dynamic systems. It combines FRF models with LSTM networks by performing both network training and predictions on short subsequences of measurements, which are later recombined. Two di ff erent options for this hybrid combination were compared, where the first approach utilizes the LSTM network to reduce the error of the FRF model and the second approach enhances the input data of the LSTM network by the FRF prediction. Applications of both VS and FP were demonstrated using an experimental dataset, generated by a servo hydraulic fatigue test bench. In each study, the di ff erent hybrid models were compared to both a pure FRF and LSTM approach. An evaluation using both RMS and Multi-Rain error metrics shows, that LSTM and FRF models achieve a comparable overall per-