Manifold-aware diffusion method for multi-dimensional unbalanced data enhancement of complex electromechanical systems

The problem of extreme sample imbalance in industrial field monitoring data seriously restricts the decision-making reliability of diagnostic models. When dealing with high-dimensional monitoring data, the existing generation model is difficult to effectively decouple the complex observation space, resulting in the data generation model easily falling into mode collapse and a lack of fidelity. This paper proposes a new paradigm of conditional diffusion generation with manifold learning (MC-CESDM), which integrates flow matching theory and feature decoupling mechanism. It effectively overcomes the limitations of the existing models: stream matching avoids lengthy iterative noise prediction by constructing a deterministic probability path, which greatly improves the training stability and sampling speed; feature decoupling ensures that the generated samples can be accurately controlled by continuous physical attributes such as variable operating conditions. In this study, a typical wind turbine monitoring data is used as a case for experimental verification, and compared with the mainstream generation algorithm. The results show that the model has achieved a significant lead in key indicators such as feature correlation and distribution distance (Fréchet distance), which proves its excellent generation fidelity. This method overcomes the bottleneck of efficiency and quality of data generation under complex operating conditions, and provides an efficient and reliable technical tool for sample enhancement of unbalanced data of electromechanical systems. It lays a theoretical foundation for improving the level of intelligent operation and maintenance of industrial equipment.