Enhanced Neural Network Filtering for High-Temperature H2O Concentration Detection Based on U-Net

To address severe noise interference in high-temperature H2O absorption spectra in metallurgical environments and the limited performance of traditional filtering algorithms, this study proposes an enhanced U-Net neural-network filtering method based on residual connections and smoothness constraints, termed Residual & Smooth U-Net (RSU-Net). Based on the classical U-Net architecture, the proposed model incorporates dilated convolutions, residual connections, a smoothness constraint in the loss function, and endpoint smoothing to suppress boundary fluctuations while preserving spectral line profiles. For simulated spectra, RSU-Net increased the signal-to-noise ratio (SNR) from 16.35 dB to 50.43 dB, outperforming traditional filters and exceeding a feedforward neural network-assisted Savitzky–Golay (S-G) filtering method by 6.19 dB. Quantitative spectral-fidelity evaluation showed that RSU-Net achieved an integrated absorbance relative error of 0.11% and a peak position shift of 1.2 × 10−4 cm−1, indicating effective preservation of the main absorption peak characteristics. In experiments at 800–1010 °C, RSU-Net effectively reduced noise. At 1000 °C, it reduced the concentration standard deviation by 8.83 ppm compared with wavelet filtering. Allan deviation analysis was used as a relative stability indicator and showed improved stability over wavelet filtering.