Research on Quantitative Detection of Industrial Mixed Gases Based on Improved BP Neural Network

To address the cross-sensitivity and non-linear coupling issues caused by the coexistence of hydrogen, carbon monoxide, ammonia, and nitrogen dioxide in industrial environments, a flow-through quantitative detection system based on a MEMS gas sensor array was designed and constructed. The steady-state peak sampling method was employed for feature extraction from high-dimensional time-series data, and regression prediction models were developed using a traditional BP neural network and BP neural networks optimized by four swarm intelligence algorithms (ALA, AOO, SFOA, and SDO). The experimental results indicate that the intelligent optimization algorithms excel in decoupling the "cross-response" phenomenon, with all optimized models outperforming the traditional BP network. Among them, the SDOBP (Sledge Dog Optimizer-BP) model demonstrated the best overall performance, achieving the highest accuracy in carbon monoxide and hydrogen detection, with the Root Mean Square Error for hydrogen reduced to 2.17, an 84.2% improvement over the traditional model. The system achieves high-precision quantitative inversion of multi-component gases in complex environments, providing an effective means for industrial environmental safety monitoring.