An End-to-End Self-Supervised Contrast Learning Network Embedded With Prior Knowledge for Laser Doppler Signals Denoising

The laser Doppler signals are affected by environmental light, temperature, and mechanical vibrations during measurement, leading to significant baseline drift and random noise. Conventional laser Doppler vibrometer (LDV) signal denoising methods encounter several challenges, including intricate parameter selection, limited real-time performance, incomplete noise suppression, and inadequate generalization capability. To address these challenges, an end-to-end self-supervised contrast learning network embedded with prior knowledge (E2E-PKSCLN) is designed. First, the proposed method designs a prior knowledge embedding module based on the Akaike information criterion (AIC), which models baseline drift as a piecewise linear trend and removes it from the original signal through residual computation. Then, the denoising autoencoders with simple self-supervised contrast learning (SimCLR) module are constructed to optimize the signal feature expression through parallel data augment and contrast learning, which suppresses complex noise across frequency bands. Finally, an all-fiber laser Doppler vibration measurement verification platform is established, and 60 groups of LDV signals across 12 frequency bands are collected for testing. The experimental results demonstrate that the proposed method enhances the key indicators of signal-to-noise ratio (SNR) by 1–40 dB compared to the traditional methods, especially in the high-frequency band, with the structural similarity index measurement (SSIM) value of signal denoising approaching 1.