Application of PSO-CNN Algorithm in Semiconductor Gas Sensors Based on Temperature Modulation

The temperature modulation detection method of gas sensors can effectively enhance the selectivity of sensors. However, the optimization of multimodal parameters in dynamic temperature modulation and the extraction of highly similar response signal features are crucial for improving the accuracy of gas identification. In this article, a low-frequency dynamic temperature modulation detection method is used, with flammable and explosive hazardous gases in the environment, specifically CH4, CO, C2H5OH, and H2S, as the test gases. By employing rectangular, sine, and triangle wave multimodal temperature modulation, 27 temperature modulation detection modes were established. An output response model based on amplitude ratio was constructed, revealing the theoretical basis for the enhanced selectivity through temperature modulation. The experiments employed fast Fourier transform (FFT) and discrete wavelet transform (DWT) for the extraction of output features from dynamic modulation parameters (waveform, bias voltage, frequency). Combined with principal component analysis (PCA) for comparative analysis, the feature modes were optimized. Based on feature extraction, a particle swarm optimization convolutional neural network (PSO-CNN) model was constructed for gas pattern recognition and classification. Experimental results show that using the FFT feature extraction method combined with the PSO-CNN temperature optimization method can more effectively distinguish similar dynamic signals. Moreover, the highest and most stable recognition rates for the four gases were achieved under the following conditions: rectangular wave modulation with a bias voltage of 4–5 V and a period of 30 s; sine wave modulation with a bias voltage of 2.5–5 V and a period of 30 s; and triangle wave modulation with a bias voltage of 4–5 V and a period of 20 s. Furthermore, under these three optimized modes, the random forest (RF) algorithm was used to achieve quantitative identification of the four flammable and explosive gases. This article combines artificial intelligence methods with dynamic temperature modulation multimodal features of gas sensors, providing an effective solution for improving the detection and identification of complex gas environments.