This article addresses the fault detection (FD) problem for traction drive systems with stochastic noises and deterministic disturbances. A traction drive system with sensor and actuator faults is first described as a dynamic process. Then, the disturbance-decoupling residual generator is developed by constructing a subspace for deterministic disturbances. Based on the generated residual signals, the corresponding ambiguity sets are constructed to characterize the distributional uncertainties of noises. Moreover, the design of the target FD system is formulated as a distributionally robust optimization (DRO) problem. By solving the DRO problem, a robust FD approach is developed. It is worth noting that this method not only delivers satisfactory detection performances, but also enhances robustness against both deterministic disturbances and distributional uncertainties of stochastic noises. The reliability and validity of the developed method are illustrated by a numerical simulation and an experimental study on an actual traction drive system.
14 October 2025 in IEEE Transactions on Instrumentation and Measurement
This paper addresses the fault detection (FD) problem for traction drive systems with the consideration of noises with inexact probability distributions. A traction drive system with sensor faults is first described as a dynamic process. Following the idea of distributionally robust optimization (DRO), an ambiguity set based on the Wasserstein metric is employed to tackle the distributional uncertainties associated with signals obtained in traction drive systems. By utilizing this ambiguity set and the DRO framework, a robust FD system is developed, ensuring reliable detection performance under the uncertainty of the noise distribution. The reliability and effectiveness of the developed approach are confirmed through an experimental study on an actual traction drive system.
The advancement of sensor technology has spurred the emergence of data-driven fault detection (FD) strategies. Particularly for many control systems with time-varying parameter characteristics, data-driven FD offers significant practical value and application value. This paper addresses the FD problem for linear time-varying (LTV) systems. Specifically, within the Sieve estimation framework, the stable kernel representation of LTV systems is designed using a novel data stacking method and B-spline basis functions. Subsequently, the coefficient function is constructed to provide a general solution to online FD tasks. Finally, this method is validated through experiments. The results demonstrate that the sieve estimation-based strategy exhibits excellent estimation and FD performance in both offline and online phases.
This article addresses the fault detection (FD) problem for traction drive systems in the presence of unknown noise covariances. The dynamic behavior of the traction drive system, affected by actuator and sensor faults, is first formulated. Following the philosophy of the subspace identification, the system matrices are identified directly from collected process data using QR decomposition and singular value decomposition. Based on the identified model, a robust Kalman filter (KF)-based FD scheme is developed. By exploiting the iterative interaction between the estimator and measurement data within the KF framework, the noise covariance matrices are adaptively estimated, which alleviates the adverse effects caused by empirical covariance selection in conventional KF-based FD methods. Experimental results obtained from a real traction drive system verify the effectiveness and reliability of the proposed approach.