Coupling object detection and digital twin: an algorithmic framework design for predictive maintenance of mechanical equipment

Abstract

This paper addresses the shortcomings of single-sensor modal representation capabilities and fragmented features in virtual model space during the operation and maintenance of complex mechanical equipment. A predictive maintenance algorithm framework coupling target detection and digital twins is proposed. By constructing a cross-modal spatiotemporal feature alignment module and an adaptive cross-attention fusion network, latent space interaction and noise filtering of highdimensional visual defect semantics and low-frequency temporal signals are achieved. Based on the digital twin statespace model and incorporating monotonically degrading constraints, an autoregressive evolution algorithm is designed to capture the global trend of nonlinear degradation of the equipment. Experimental evaluation in an edge-cloud collaborative hardware environment shows that this coupled framework improves the prediction accuracy of the remaining service life of the equipment by 4.2% and achieves an inference speed of 46 FPS and a system latency of 21.5 ms on edge computing nodes. The feature interaction and closed-loop evolution mechanism established in this study optimizes the feature perception accuracy of industrial systems in non-stationary service environments, providing a quantitative algorithm solution for predictive operation and maintenance scheduling of industrial equipment.