Conv-TssNet: A time-domain signal separation network based on cepstral analysis for direct absorption spectroscopy with interference features

This work presents Conv-TssNet, a fully convolutional neural network that integrates cepstral analysis for end-to-end separation of overlapping absorption signals directly in the optical time domain. Trained on a large-scale HITRAN-based dataset augmented with realistic baseline perturbations and spectroscopic variability, Conv-TssNet demonstrates robust generalization across diverse gas mixtures and measurement conditions, achieving high-fidelity reconstruction with mean scale-invariant signal-to-noise ratio (SI-SNR) values of 37.2 dB and 40.5 dB for target and background components, respectively. Experimental validation in a premixed CH 4 /air flat-flame burner confirms accurate and repeatable retrievals of temperature and CO concentration, with uncertainties as low as 14 K and 0.14%, respectively. Compared to the interference-feature baseline transform and background subtraction (IFBTBS) method, Conv-TssNet significantly reduces retrieval uncertainties and demonstrates superior noise immunity under turbulence-induced distortions. It maintains temperature deviations below 3.33% with 2.30% uncertainty at a high distortion level ( c 1 / 2 = 3 . 0 ), compared to 7.25% deviation and 5.10% uncertainty with IFBTBS, and accelerates computational speed by approximately 900-fold, enabling real-time, in situ applications. This study represents the first cross-domain adaptation of deep time-domain speech separation architectures to laser absorption spectroscopy, establishing a novel framework for applying deep learning to complex optical diagnostics. Future work will focus on extending the proposed network to multi-species spectral separation, developing integrated retrieval-inference frameworks, and incorporating uncertainty-aware modeling to enhance robustness for practical field deployments. • A fully convolutional neural network, Conv-TssNet, that integrates cepstral analysis for end-to-end separation of overlapping absorption signals in the optical time domain, is proposed. • Trained on a large-scale HITRAN-based dataset with baseline perturbations and spectroscopic variability, Conv-TssNet achieves high-fidelity reconstruction (SI-SNR up to 40.5 dB). • Experimental validation in CH 4 /air flat-flame burner confirms accurate retrievals of temperature (±18 K) and CO concentration (±0.18 %). • Compared with IFBTBS, Conv-TssNet reduces retrieval uncertainties, improves noise robustness under turbulence (3.33 % vs 7.25 % deviation), and accelerates computation 900×.