Condition diagnosis method for hydraulic-system based on random forest-XGBoost stacking model

Abstract Hydraulic system condition monitoring and diagnosis are critical for ensuring construction machinery reliability. To improve the accuracy, robustness, and interpretability of fault diagnosis for multiple components of a hydraulic system through the same monitoring data, this study proposed a novel data-driven integrated tree model fault-diagnosis method for hydraulic system. The method utilizes multi-sensor state data and operating condition indicators of the hydraulic system as the driving dataset. A random forest model was first employed as the base learner, with out-of-bag errors used to measure the accuracy of the individual tree models, along with weighted measures incorporating temperature smoothing to enhance them. Then the diagnostic results from the improved random forest model were combined with the original dataset to construct a meta-feature dataset, which served as the input for an extreme gradient boosting (XGBoost) meta-model to achieve the final diagnosis model and results. The Bayesian optimization was employed to determine the optimal hyperparameters for both the random forest and XGBoost models. The Shapley additive explanations framework was utilized for quantitative analysis of the influence weights of various features on the prediction results, revealing patterns for the contributions of different operational parameters to the diagnostic condition of the hydraulic system. Experimental results indicated that this method strikes a commendable balance between prediction accuracy and interpretability, achieving an overall accuracy of up to 99.4%. The analysis revealed that pressure, flow, and motor power play crucial roles in the fault diagnosis of the hydraulic system, providing effective data-driven support for the condition diagnosis of hydraulic systems.