Multi-projection support matrix regression for remaining useful life prediction of roller bearing

Support vector regression (SVR) benefits from strong statistical foundations for roller bearing remaining useful life (RUL) prediction, but it requires disrupting the data structure for feature extraction. Support matrix regression (SMR) is a novel data regression method that preserves the global features of the matrix-form data but ignores local structural state information. To addressing the aforementioned structural limitation, a matrix-form regression model is proposed, termed multi-projection support matrix regression (MPSMR). In MPSMR, a novel multiple projection modeling mechanism is designed to capture local manifold geometric properties, which can preserve the intrinsic data structure while effectively capturing potential inter dimensional correlations. Meanwhile, in the multi-hyperplane regression model ensemble, an innovative dynamic weighting allocation criterion is proposed to enhance the flexibility and adaptability, making the model more accurate in predicting data under different operating conditions. Experimental validation on two accelerated degradation roller bearing datasets conclusively establishes the superior performance.