Remaining Useful Life Estimation for Key Components of Manufacturing Equipment With Individual Differences

Methods of accurate and reliable remaining useful life (RUL) estimation play a vital role in reducing proportion of both overuse and underuse of key components during the machining process, and have important engineering significance for ensuring machining accuracy. However, stochastic process model-based methods usually have limitations in modeling such complex machining processes due to the lack of nonlinear processing capabilities, and data-driven methods are insufficient to quantify the prediction uncertainty of degradation processes. In addition, owing to the complexity of the structure, materials, and usage environment, there will inevitably be a distinct degradation trajectory among different instances of the same type of key component. Therefore, a data-model interaction mechanism is proposed to compensate for the shortcomings of both composite degradation index (CDI) construction and stochastic degradation modeling so as to improve the accuracy of RUL estimation. Specifically, k-nearest neighbor (KNN) is employed to establish a topology for capturing the interdependencies among sensor nodes. Then, the CDI for characterizing the degradation trends of key components is constructed by a graph Kolmogorov-Arnold network (GKAN), which enhances the nonlinear characterization ability of the method. Moreover, the drift coefficients within the stochastic degradation model are characterized as random variables following a Gaussian distribution, aiming to quantify the inherent individual variability during the modeling of the stochastic degradation process. More importantly, the data of the intermediate degradation process are taken into account by the improved stochastic degradation model, allowing for greater flexibility in the initial degradation level. Finally, the effectiveness and superiority of the proposed method are verified by the PHM2010 and PHM2012 datasets.