Dynamic operation friction temperature model for bearings based on the SKF method and intelligent algorithms

Abstract This study addresses the complex challenge of predicting bearing temperature by developing a specialized experimental setup designed for bearing temperature testing. The experimental rig enables the measurement of outer ring temperatures under various operational conditions, including variations in bearing speed, radial and axial loads, grease content, and initial temperatures. A range of models is employed, including the SKF friction temperature model, along with advanced machine learning algorithms such as the Backpropagation Neural Network (BP), Genetic Algorithm Back Propagation (GABP), and Particle Swarm Optimization Back Propagation (PSOBP). These models aim to predict bearing temperatures based on the aforementioned experimental parameters. A comprehensive comparative analysis highlights the superior performance of the PSOBP algorithm in terms of both efficiency and stability, outperforming the BP and GABP models across a range of performance metrics. Furthermore, by calculating the SHAP values of the machine learning models, the study quantifies and ranks the contributions of various factors influencing bearing temperature. However, it is important to note that while the PSOBP model demonstrates strong performance, its predictive accuracy for the nine steady-state conditions in this study is slightly lower than that of the GABP model. This discrepancy may be attributed to the limited sample data available for training. The intelligent algorithm features a streamlined integrated modeling framework that obviates the need for separate kinetic, thermodynamic, and lubrication sub-models. Within the range of training input parameters, it facilitates the modeling of both transient and steady-state thermal behaviors without the need to define complex coupling relationships between boundary conditions and variables explicitly. Additionally, the algorithm demonstrates adaptability to different bearing types and possesses basic learning capabilities, which lay a foundation for further generalization to broader operating conditions.