Development of an LSTM-based Filtering Technique to handle Unacceptable Oscillations in Dynamic Tensile Testing at Intermediate Strain Rates

Kurzfassung

Dynamic tensile testing at intermediate strain rates poses a persistent challenge in accurately assessing force signals. Measured force signals contain not only intrinsic material properties but also inertial effects that are superimposed onto the force signals. The superimposed oscillations complicate their interpretation and, therefore, hinder the precise identification of material parameters. Thus, there is a demand for a reliable technique to handle undesired oscillations linked to inertial effects. This study presents a novel adaptive filtering technique based on Long Short-Term Memory (LSTM) networks for signal decomposition and prediction as shown in Fig. 1. The model is initially trained on experimental data of metallic materials recorded in a DLR research data management system (RDMS), shepard [1]. Furthermore, a model-based data augmentation is implemented to create synthetic data, allowing the variability in the LSTM model. This implementation is utilised with a dynamic model based on a 4-degree-of-freedom (DOF) mass-spring-damper model representing the entire testing setup [2]. After that, the trained LSTM model is used to decompose the force signals into desired and undesired components of force signals. Dynamic tensile testing of fibre-reinforced polymer (FRP) composites at intermediate strain rates up to 200 s-1 is performed to validate the model, while experimental data intentionally contains force oscillations. The findings reveal that the pre-trained model enables the identification of unacceptable oscillations from newly measured force signals. Further validation successfully filtered out unacceptable oscillations from force signals, significantly enhancing signal clarity on true material properties. Finally, the limitations of this technique are discussed. This data-driven approach is efficient and independent of the user’s expertise, unlike traditional filtering methods that rely entirely on the expertise. This technique will allow us to gain deeper insights into the dynamic behaviour of composite materials at intermediate strain rates, thus improving material characterisation methodologies for accuracy and reliability.