Lightweight and Accurate Identification Model for Precursor Chemicals Based on Semiconductor Gas Sensor

Precursor chemical identification is of great significance to ensure public security and human health. Although the related identification models have achieved high accuracy, most models require a huge number of parameters. To solve the problem, an identification scheme of constructing a lightweight and accurate identification model is proposed based on linear discriminant analysis (LDA), class activation mapping (CAM), binary coding representation, and self-attention mechanism. First, the response data were obtained based on a semiconductor gas sensor and dynamic temperature modulation. Second, LDA was used to reduce the dimension of response data. Inspired by CAM, a novel feature selection method was proposed to extract the optimum feature combination, which can further reduce the dimension of the input feature. Subsequently, the response data were labeled by binary coding representation to effectively reduce the label dimension. Finally, the data-driven model was constructed based on a self-attention mechanism, only containing 129 parameters. The test results indicate that the qualitative and quantitative identification accuracies of the proposed model can achieve 100% and 98.61%, respectively. Such a lightweight and accurate identification model is a promising approach for integrating with embedded systems and portable equipment.