Objective Address the challenges posed by multi-sensor devices in terms of recognition rate and prediction accuracy of gas mixtures to enhance the detection accuracy of flammable gas, thereby ensuring the safety of personnel exposed to hazardous gas.

Methods The public dataset "Gas Sensor Arrays for Dynamic Gas Mixtures" from the University of California, Irvine(UCI) was employed for simulation tests using MATLAB 2021b. The sensor array in the dataset comprised 16 sensors. A total of 300 gas samples were used for training, including 70 carbon monoxide samples, 80 ethylene samples, 70 air samples, and 80 carbon monoxide-ethylene mixture samples. A flammable gas mixture classification method was proposed based on machine learning and sensor array technology. Principal Component Analysis (PCA) was first utilized to reduce the dimensionality of the input data, then Particle Swarm Optimization (PSO) was employed to optimize the hyperparameters of the Support Vector Machine(SVM). Finally, the gas recognition and classification results were compared with those obtained using unoptimized SVM and unoptimized BP neural network methods.

Results After PCA processing, the dimensionality of the original data was reduced from 16 to 4, with a cumulative contribution rate exceeding 99%. This algorithm achieved 100% accuracy(50/50) in the qualitative identification of air, carbon monoxide, mixed gas, and ethylene, which was 10% and 2% higher than the accuracy of unoptimized SVM (90%, 45/50) and BP neural network (98%, 49/50), respectively.

Conclusions The proposed method based on PCA, PSO, and SVM can accurately identify and classify gas mixtures, improving the speed and accuracy of sensor alarms and facilitating the timely detection of potential hazards in production. However, further research is needed to study the efficiency and accuracy of identifying gases with more complex compositions.