In complex engineering environments, the majority of collected data predominantly comprises samples from normal operating conditions, leading to highly imbalanced datasets characterized by a scarcity of minority-class samples. Meanwhile, these data are inevitably contaminated by noise that masks subtle characteristic patterns, further complicating classification and recognition tasks. Traditional Transformer models exhibit limitations in capturing local features, sensitivity to data volume, difficulty in extracting hierarchical features, and susceptibility to overfitting. To overcome these limitations, this work proposes a novel Neutrosophic Transformer. By integrating a neutrosophic input representation and a neutrosophic feature extraction head, the model facilitates effective information exchange among diverse data components, including noise versus signal and minority vs majority samples. Consequently, the proposed architecture not only captures global contextual features but also preserves local and fine-grained details, enabling precise feature extraction. Furthermore, we introduce a Kernel Neutrosophic K-Nearest Neighbors (KNKNN) model designed to handle nonlinear data distributions in the extracted feature space. It is integrated with the proposed Neutrosophic transformer to form a unified framework that delivers robust and accurate classification. Comparative experiments on three real-world imbalanced engineering datasets demonstrate that the proposed method surpasses comparative approaches in feature extraction capability, classification accuracy, robustness, and generalization. Moreover, it achieves significant improvements in overall analytical performance.
Wu et al. (Mon,) studied this question.