Sample Generation Based on Guided Spatial Constraints for Industrial Safety Condition Recognition

Abstract Real-time and accurate identification of industrial conditions is crucial for ensuring safe production and enhancing operational efficiency. However, in practical applications, it is often difficult to obtain sufficient samples of abnormal or specific conditions, which severely limits the performance of traditional data-driven models. To address this issue, we propose a sample synthesis method based on Guided Spatially Constrained Sample Generation (GSC-SG). The core idea of this method is to define clear spatial constraint conditions based on the inherent spatial prior knowledge of industrial scenarios, limiting the scene spatial range for sample generation, and ensuring that the generated augmented samples align with real working conditions. The proposed method yields high-quality augmented samples that adhere to scene logic and industrial norms. Experiments conducted on two dedicated industrial datasets demonstrate that the proposed GSC-SG significantly improves key performance indicators such as mean Average Precision (mAP) and precision compared to various baseline data augmentation methods.