Everyday life problems, armed conflicts, pandemics, and catastrophes – these are situations that are always accompanied by stress. Its chronic form can lead to so-called stress-related illnesses. Despite the development of health prevention, many people still get sick due to stress. Therefore, it is important to seek detective and classificatory solutions for stress, which may influence its reduction or control in the future. The example of this can be the thermographic stress registration presented in this article, combined with classification using lightweight CNN and Choquet fuzzy ensemble learning. The article proposed new ensemble frameworks for stress classification based on Choquet fuzzy integral, serving as an aggregation function. In the study, three pre-trained lightweight CNN models were used: MobileNetV2, Xception, and EfficientNet. The proposed fuzzy ensemble model achieves a classification accuracy above 90%. This work is of a prospective nature, with the possibility of implementing solutions in biomedical-psychological activities.