Adversarial unpaired disentanglement network (AUDNet): Precise analysis of severely overlapping CS2/SO2 signals in UV-DOAS.

CS₂ and SO₂ are typical characteristic gases for evaluating the insulation status of gas-insulated switchgear. Their significant absorption peaks in the ultraviolet band make ultraviolet differential optical absorption spectroscopy (UV-DOAS) an ideal detection method. However, severe spectral overlap limits the accuracy of traditional approaches. To resolve interference issues, this paper proposes a CS₂/SO₂ mixed gas detection system combining UV-DOAS with a generative adversarial network (GAN). This approach employs adversarial training to identify and separate single-gas features from overlapping spectra. The system first uses DOAS for preprocessing the mixed spectra. Then, the adversarial unpaired disentanglement network (AUDNet) is constructed based on wasserstein GAN with gradient penalty (WGAN-GP) for spectral disentangling. Its loss function ensures data consistency and smoothness of disentangled structures. Experiments show that within test ranges of CS₂ (3.81-179.2 ppb) and SO₂ (44.43-942.73 ppb), the mean absolute percentage errors (MAPE) for concentration prediction are 1.67% and 1.12%, respectively, with detection limits reaching 0.33 ppb and 7.04 ppb at a 50-cm optical path. Compared to existing deep learning methods, this system requires no strict pairing between mixed spectra and single pure-component spectra for training.