An improved Bayesian network for deviation diagnosis in high-speed spinning winding chuck assembly

Abstract

The quality of mechanical assembly is essential for ensuring the overall performance and reliability of the final product. However, current methods for diagnosing assembly deviations often fail to meet the practical demands of assembly quality assessment. This paper presents a new approach for monitoring deviation sources during the assembly process by utilizing a Structure-Variable Dynamic Bayesian Network (SVDBN). By thoroughly analyzing the entire assembly process, mathematical models are developed to address different types of deviations. The SVDBN is composed of two stages: learning the prior probability network and performing posterior probability inference. The prior network learning involves both network structure and parameters, introducing an automated method for real-time updates of the network structure as it evolves. Cross-validation kernel density estimation (CV-KDE) is applied to learn the probability density functions of each feature, achieving a mean squared error as low as 0.001. This method was validated through the chuck shaft assembly process in a winding machine, where real-time probabilistic inference was conducted for the entire assembly, along with quantitative and sensitivity analyses of the deviation source nodes. The overshoot probability for deviation sources such as MDA, MDSC, and MDLS was concentrated between 0.80 and 0.95, while the range for DDTS was broader, spanning from 0.65 to 0.95, indicating its significant effect on dynamic balance response, especially during the third stage of long sleeve assembly. Sensitivity analysis showed that the long sleeve assembly had the highest overshoot rate, confirming its critical role in the chuck shaft assembly process.