Compound fault diagnosis of rolling mill main drive system based on UIO-TROA-CNN

Abstract

To address compound faults—including actuator faults, sensor faults, and unknown disturbances—in the main drive system (MDS) of a rolling mill, a fault diagnosis model is established based on the d-q model of a three-phase AC motor. To overcome the limitation of conventional model-based fault detection (MFD) in isolating faults under compound fault scenarios, a novel fault diagnosis framework is proposed by integrating model residuals with deep learning. An Unknown Input Observer (UIO) is designed to detect system faults, and its convergence is rigorously proven using Lyapunov theory and linear matrix inequalities (LMIs). The coupled residual signals generated by the observer are segmented into sequential subsequences and processed by a Convolutional Neural Network (CNN) for feature extraction and classification. To account for the temporal and dynamic nature of the residuals, a Tyrannosaurus Rex Optimization Algorithm (TROA) is adopted to optimize the CNN hyperparameters. Numerical simulations on a rolling mill demonstrate that the proposed UIO achieves superior state estimation performance, with a 16.48% reduction in the Root Mean Square Error (RMSE) of angular velocity difference of the motor compared to the Sliding Mode Observer (SMO). Furthermore, the proposed TROA-CNN outperforms Bayesian Optimization (Bayes)-CNN, Whale Optimization Algorithm (WOA)-CNN, and Grey Wolf Optimization (GWO)-CNN in terms of fault classification accuracy (99.76%) and noise robustness (96.27% under 5% noise). In scalability tests where the number of fault types or dataset size are doubled, the inference latency increases by approximately 10%, and the training time rises by about 40%. These results demonstrate that the UIO-TROA-CNN achieves high accuracy, strong robustness, and excellent scalability, making it well-suited for fault diagnosis in industrial environments with high noise and complex fault types.