[3F3] Towards uncertainty-aware physics-informed neural networks for prognostics in bushings
J Cuesta¹,², U Leturiondo¹, Y Vidal² and F Pozo²
¹Ikerlan Technology Research Centre, Spain
²Universitat Politècnica de Catalunya (UPC), Spain
Wear in mechanical joints represents a key factor influencing the reliability and operational lifespan of engineering systems. Revolute clearance joints, in particular, are susceptible to progressive degradation because of inherent gaps that induce dynamic contact and friction. While physical and artificial intelligence models exist, accurately predicting wear evolution under real-world conditions remains an open challenge. To address this, a hybrid modelling framework is introduced that integrates physical principles with deep learning techniques. The approach combines modified long short-term memory layers with a final Archard wear equation layer, enabling the modelling of non-linear temporal dynamics associated with wear processes. Monte Carlo dropout is incorporated for uncertainty quantification. The results obtained on a sensorised crank-slider system test-bench show that the hybrid methodology is capable of predicting the evolution of accumulated wear, offering a significant advance towards prognostics.
