A Multiscale Causal–Neural Framework for Robust Correction of Raman Signals in Confined Gas Detection

In this study, a cavity-enhanced Raman spectroscopy (CERS) detection system is designed and built to meet the long-term monitoring requirements of trace gases in confined spaces, and a dynamic signal correction framework for long-term stable operation is proposed. Nitrous oxide (N2O), which exhibits strong Raman activity and good indicative properties in closed environments, is selected as the primary target gas, and nitrogen (N2) is utilized as a long-term stable reference to construct a correction architecture. To address complex environmental interference and nonlinear signal drift, a causal–neural–Levenberg correction method (CNL-CM), integrating multiscale dynamic causal inference (MS-DCI) and nonlinear neural modeling, is developed. By extracting multiscale time-varying Granger causality index (TGCI) features and applying backpropagation neural networks (BPNNs) optimized via the Levenberg–Marquardt (LM) algorithm, real-time dynamic correction of N2O signals is achieved. Experimental results under system drift, temperature and humidity perturbations, and coupled environmental disturbances confirm that the CNL-CM method significantly outperforms traditional correction methods such as spectral band area normalization (BAN), extended Kalman filter (EKF), and long short-term memory (LSTM) networks. Furthermore, an additional validation using carbon dioxide (CO2) as an alternative target gas demonstrates the transferability and generalizability of the proposed framework. The corrected signals maintain stability throughout long-term operation, with a sliding standard deviation consistently below 0.35 counts/s. These results highlight the robustness and long-term stability of the CNL-CM method in dynamic, real-world confined gas detection scenarios.