Research on Gas Recognition Method Based on Cosine Similarity of Amplitude Ratio Features under Dynamic Detection

Temperature modulation detection methods can significantly enhance the selectivity of semiconductor gas sensors. However, the complexity of similar dynamic response signals in sensor arrays and the optimization of temperature modulation parameters are crucial to the extraction of gas feature information. In this paper, the temperature modulation detection method using amplitude-ratio ΔR_a/ΔR_g output is adopted to dynamically test a four-array configuration of commercial metal oxide semiconductor gas sensors. Nine kinds of temperature modulation detection modes are established by using the sine-wave modulation waveforms as the excitation signals, constructing a resistance response model based on ΔR_a/ΔR_g, revealing the theoretical basis for the improvement of the selectivity of the temperature modulation. The classification model based on cosine similarity (CS) algorithm was designed to simultaneously realize the qualitative and quantitative detection function of four test gases (CH₄, CO, C₂H₅OH, and H₂S). The Optimized experimental results indicate that the classification accuracy reaches 100% under modulation mode of sinusoidal 3-5 V and 4-5 V bias with a period of 30 s. Compared with the classical algorithms of K-Nearest Neighbor (KNN), Random Forest (RF), and Radial Basis Function Neural Network (RBF), the simple CS algorithm demonstrates excellent recognition performance, indicating that the amplitude-ratio feature not only controls the scale of data processing but also reduces the dependence on complex pattern recognition algorithms. Under this optimized mode (4-5 V, 30 s), the concentration prediction results were corrected by linear regression algorithm, and the R² was improved from 0.8895 to 0.9762. The good linear output under the amplitude-ratio model contributes to the accurate prediction of gas concentration.