ИСТИНА

One of the most important problems of humanity's existence in the industrial age is the problem of maintaining its comfortable living environment. Environmental monitoring is a component of the safety management system. The development of pulsed terahertz and submillimeter spectroscopy methods makes it possible to monitor the state of the environment by measuring and analyzing the absorption spectra of the atmosphere. The proposed analytical approach integrates machine learning methods with high-precision spectroscopic measurements. A ResNet neural network with self-awareness mechanisms has been created, which identifies gases in a multicomponent atmospheric mixture and determines their concentrations through the analysis of absorption spectra in the terahertz range. The network architecture includes convolutional blocks with residual connections and a specialized Focal Loss function. Experimental studies on model sets of spectra demonstrate the potential of the technique in identifying six gas components with an accuracy of 0.01 ppm. The neural network has achieved 90-95% accuracy in detecting gas concentrations. A series of experiments based on real experimental data were conducted to refine the architecture of the model and improve its characteristics. The scientific significance of the study lies in a new methodological approach that expands the possibilities of spectroscopic analysis. Research is in an active stage: noise models are being refined, data augmentation is underway, and prediction mechanisms are being adjusted. The prospect is to create a universal gas analysis platform with high sensitivity to microconcentrations of atmospheric gases.