Time-series data play a critical role in early prediction problems using machine learning, particularly in modern applications involving electronic noses for measuring gas concentrations via chemical sensors. However, a drawback of previous studies is they extremely focus on time-series processing models but ignore preprocessing steps to generalize data, leading to the model being able to learn and predict accurately only a number of particular learned gas concentration values. To address this issue, a novel approach is proposed that transforms the original time-series data into square images, where each image corresponds to a pair of gas concentration values, while preserving the time-series characteristics of the data. The achieved results are comparable to those of previous studies but demonstrate a significantly improved capacity for generalization in prediction tasks.