CNN-transformer hybrid deep learning method for suppressing interference fringe noise

Tunable Diode Laser Absorption Spectroscopy (TDLAS) technology often faces the problems of interference fringe overlap and phase shift, leading to target signal distortion. Traditional digital filtering and existing deep learning models are often limited to local smoothing when dealing with dynamic distorted noise that has frequency aliasing with real signals, which easily causes signal feature loss or absorption peak clipping. To address this challenge, a novel hybrid denoising architecture (New CNT) based on Convolutional Neural Network (CNN) and Transformer is proposed in this paper. This model uses CNN to accurately extract local waveform distortion features and combines the multi-head self-attention mechanism of Transformer to filter global background noise. To accelerate model convergence and avoid invalid training, an equivalent data optimization strategy is adopted, which relies on a smaller-scale training set to enable the model to obtain an additional performance gain of 4.42 dB while significantly reducing the total training overhead by 19%. In addition, the Monte Carlo phase perturbation technology is introduced into the training data to simulate actual working conditions, and a joint loss function (including weighted mean square error, Pearson correlation coefficient, and peak missing penalty) is innovatively designed, which effectively overcomes the peak shaving phenomenon during denoising and maintains the similarity of physical waveforms. The simulation results show that the average signal-to-noise ratio (SNR) of the mixed signal with envelope distortion noise is increased by 44.60 dB with the New CNT model, respectively, and exhibits extremely high accuracy and stability in the evaluation indicators of Peak Height Error (PHE) and Relative Standard Deviation (RSD). This study provides a reliable deep learning solution for TDLAS systems with high timeliness, high precision, and strong robustness.