Rebalancing Mel-frequency Cepstrum and parallel fusion model for surface hardness monitoring of laser shock peening using acoustic emission

Laser shock peening (LSP) technology is commonly used to strengthen and extend the life of components such as aero-engine compressor blades. Online monitoring of the LSP process using acoustic emission (AE) technology is crucial for improving its reliability and consistency. However, it is still challenging to obtain the accurate monitoring due to the high sampling rate of AE with wide frequency band, complex propagation behavior of AE elastic waves and difficulty in accurately characterizing subsurface plastic deformation. In this paper, we propose a new method for surface hardness monitoring of LSP based on AE and different degree of plastic deformation both for TC4 titanium alloy and 7075 aluminum alloy with experimental study. Firstly, interval bimodal difference D_iJ and interval bimodal slope K_iJ of the power spectral density (PSD) of AE signal was proposed to quantitatively characterize the degree of plastic deformation of those two types of materials after LSP before revealing the nonlinear correlation between LSP-AE and the variation of microscopic properties of LSP subsurface material's. Then, a new lightweight Transformer Parallel Fused Partial Convolutional Neural Networks (TPFPCNN) driven by Rebalancing Mel Frequency Cepstrum (RMFC) is proposed to identify six types of surface hardness of LSP based on AE. The proposed RMFC can adaptively rebalance the filter bank of the conventional linear Mel frequency domain to enhance the key information of LSP-AE while weakening the less-key time–frequency area. Moreover, careful comparisons and discussions were performed with Linear Frequency Spectrogram (LFS) and Mel Frequency Cepstrum (MFC) as well as TC4 titanium alloy and 7075 aluminum alloy under different LSP parameters. Finally, a strict quantitative comparison was made between TPFPCNN and four other state-of-the-art models in terms of the calculation complexity and mean test accuracy before feature visualization of the parallel fusion model and its explanation related to of the LSP process.