Generative Adversarial Networks in Imbalanced Gas Samples

Deep neural networks have been widely applied for gas concentration estimation in low-cost gas sensor arrays; however, their dependency on sample distribution remains a significant challenge. Current research indicates that deep learning models are susceptible to sample imbalance, where their predictive accuracy is strongly influenced by the number of available samples. In sensor arrays used for monitoring indoor and outdoor harmful gas emissions, most response values remain within a normal range, while only a limited number exhibit high response values. Addressing this imbalance typically requires assigning weights to different classes or pruning datasets; however, the cross-sensitivity of sensors and the limited availability of datasets complicate this approach. In this study, we investigated the impact of sample imbalance on model performance and proposed a simulated sensor generative adversarial network (SSGAN) to generate synthetic sensor response values alongside their corresponding gas concentrations. A multiple-sensor generator was designed to produce sensor array response values paired with gas concentrations, while discriminators ensured that generated samples closely resembled real instances without being identical. Furthermore, a customized generative loss function was developed to optimize the training of the SSGAN. To validate our approach, experiments were conducted on the UCI Machine Air Quality dataset using a traditional convolutional neural network (CNN), a backpropagation neural network (BPNN), and a custom-designed attention block. The results demonstrated that SSGAN effectively reduced the average absolute error of the three target models by 4.45%, 12.06%, and 3.08%, respectively.