Understanding the Incident Wave Errors in Split Hopkinson Pressure Bar Test with Machine Learning Method

Abstract

Background

In Split Hopkinson Pressure Bar (SHPB) test, the misalignment of the striker bar leads to waveform errors in the incident wave, which results in inaccurate material mechanical property parameters.

Objective

The goal of this paper is to apply machine learning (ML) method to understand waveform errors in incident waves (error peak-valley features) and investigate the impact of imperfect striker bar on the incident wave.

Methods

ML projects were constructed by developing numerical models to establish waveform databases based on experimental data, and the continuous optimization of ML projects advances the application of a dual-output average curve (DOAC) method simulating the use of two strain gauges for error processing.

Results

The waveform errors were categorized into two types: non-parallel impact and parallel non-coaxial impact by continuously optimizing the ML model through error analysis, successfully understanding up to 24 types of waveforms. DOAC effectively eliminated the bending effect, and the error effects were decomposed into bending effects and other effects.

Conclusion

The high-accuracy ML results provide simple and real-time automatic correction solutions for waveform errors and quantify the errors, closing the loop between numerical simulation and experiments. The error and dispersion coupling effects can be successfully decoupled using DOAC, suggesting that bending waves are the main cause of error effects with the dominant bending effects.