LPDD-SSAN: Limited Physics Data Diffusion Guided Selection Sparsity-Aware Attention Network for Engine Valve Wear Prediction

In marine internal combustion engines, valve wear is a critical factor affecting engine performance. Accurately predicting valve wear can significantly enhance engine reliability and maintenance efficiency. However, frequent shutdowns for valve disassembly and measurement are impractical in marine engine, making data acquisition extremely challenging. This limitation restricts the fitting capability of traditional deep learning models, leading to a noticeable decline in detection accuracy. In this article, we propose a novel limited physics data diffusion guided selection sparsity-aware attention network (LPDD-SSAN) for engine valve wear prediction, which requires only small amount of physical data to effectively model valve wear. First, we develop a plug-and-play limited physical data diffusion (LPDD) module based on a triboelastodynamics model. This module effectively enriches the data space and enhances the model’s fitting performance by diffusing limited physical data into the vibration-wear domain. Second, we design an innovative selection sparsity-aware attention network (SSAN) to accurately fit the vibration-wear domain. Finally, we introduce a sequence weight selection module (SWSM) to enhance the feature representation of the vibration-wear domain, thereby improving the model’s predictive accuracy. In our experiments, the LPDD-SSAN demonstrates superior performance, achieving an $F1$ -score of 98.50%, representing a 7.59% improvement over the baseline. The results indicate that LPDD-SSAN can effectively predict valve wear under limited data conditions, offering excellent robustness and generalization capabilities.