
In the presence of an actuator fault, the robotic manipulator may lose its original control accuracy and may therefore fail to achieve the desired control objective. In order to overcome this problem, this paper proposes a dual event-triggered fuzzy adaptive fault-tolerant control method based on monitoring functions. Firstly, a set of monitoring functions is designed to monitor the error variables. When an actuator fault occurs, if the error variables violate the constraints imposed by the preset monitoring functions, the negative impact of the faulty actuator on the control performance is eliminated by switching to a healthy actuator. Secondly, the monitoring functions constrain the error variables and improve the transient and steady-state performance of the system. Based on the above ideas, an adjustable fault-tolerant band is designed to tolerate actuator faults of a certain severity when the error variables do not violate the prescribed performance constraints. Meanwhile, to effectively save the communication resources of the system, a dual event-triggered control strategy is designed in this paper for the adaptive law and the controller, and Zeno behavior is avoided. Finally, the superiority of the proposed control algorithm is illustrated by experiments on a two-link robotic manipulator.
Federated Learning (FL) enables collaborative model training without centralising user data, but in cross-device settings the large and volatile client population makes training slow and unreliable. Two factors dominate: (i) many devices are offline when polled, and (ii) the online devices differ widely in compute and network speed. We introduce TiFL–MDA, a composite client-selection strategy that jointly tackles these hurdles. First, Tier-based FL (TiFL) groups clients into latency tiers that reflect their resource profiles; then the Must-Detect-Availability (MDA) filter chooses the most responsive clients within each tier. By coupling resource awareness and availability detection, TiFL–MDA maximises parallelism without sacrificing fairness. Across five datasets and six heterogeneity settings, TiFL–MDA reaches target accuracy in up to 6.5 % fewer communication rounds than vanilla FL and improves final accuracy by up to 23 %. Unlike 'fast-only' selection, our tier-aware design prevents bias towards high-end devices.
Accurate prediction of remaining useful life (RUL) and uncertainty quantification are crucial for the prognostics and health management of degrading equipment. However, achieving high-precision predictions under data-scarce conditions and complex state-dependent degradation dynamics remains challenging. The combination of physical mechanisms with stochastic differential equation (SDE)-based degradation models reduces data dependency and captures state-dependent dynamics. Nevertheless, directly obtaining tractable RUL distributions from complex state-dependent SDEs remains challenging. To address these limitations, this paper proposes a physics-informed backward Kolmogorov neural operator (PI-BKNO) for adaptive few-shot RUL prediction and distribution-free uncertainty quantification. Firstly, a state-dependent degradation model is developed using physics-informed SDEs. To characterize the first-hitting-time distribution, a survival analysis framework is established based on the backward Kolmogorov equation (BKE) for characterizing the degradation dynamics with conditional survival probabilities. Secondly, the PI-BKNO architecture is proposed to learn a continuous operator mapping from dynamic degradation states to survival probabilities, thereby enabling efficient solution of the BKE. Afterwards, a synergistic online adaptive strategy is developed that combines physical parameter update with branch network fine-tuning for the real-time update of the underlying parameters. In addition, the RUL probability density is derived based on the predicted survival functions, providing distribution-free uncertainty quantification. Finally, a numerical simulation and a real-world fatigue crack growth case study are conducted. The obtained results demonstrate the superior performance of the proposed framework in predicting accurate RUL and quantifying uncertainties under few-shot scenarios.
Early detection of incipient demagnetization of permanent magnet synchronous motors is critical to ensure system reliability. However, existing demagnetization diagnosis techniques predominantly rely on single-sensor signals, which suffer from inadequate fault information and unstable sampling. Moreover, the lack of sufficient labeled samples in practical scenarios further undermines diagnostic accuracy and generalizability. To address these challenges, this study proposes a novel multi-sensor information fusion graph convolutional network (MSF-GCN) framework for robust demagnetization fault diagnosis. First, a hybrid CNN+LSTM module is employed to extract features from multi-sensor data (torque, three-phase currents), enabling initial fault identification and quantifying sensor-specific contribution factors. To enhance model generalization with limited samples, two graphs are generated based on features and contribution factors respectively. These graphs are then fused and processed through a graph convolutional network to achieve cross-domain demagnetization pattern recognition, with domain discrepancy via maximum mean discrepancy (MMD) regularization. Finally, extensive experiments on a custom-built demagnetization test platform validate the proposed method. Results demonstrate that multi sensor fusion outperforms single-sensor approaches by uncovering richer fault-related information. The proposed method achieves high accuracy and robustness in demagnetization fault diagnosis under diverse operating conditions and limited sample sizes.
This paper investigates leader-following consensus for heterogeneous double-integrator multi-agent systems subject to actuator loss of effectiveness and input saturation. A fixedgain distributed observer-based protocol is developed using local output measurements and neighboring information, without requiring global topology knowledge, online parameter estimation, or a fault-diagnosis and isolation module. For exact locally assigned effective-input gains, a Lyapunov analysis combining integral terms with LaSalle's invariance principle establishes global asymptotic leader-following tracking for every follower retaining nonzero control authority. For unknown but bounded constant mismatches in the inertia and actuator-effectiveness parameters, explicit sufficient robustness conditions and uniformly ultimately bounded tracking-error estimates are derived. Separate conditions are also obtained to preserve asymptotic tracking under bounded controller- and observer-gain tuning inaccuracies. Followers subject to complete actuator failure are excluded from the tracking guarantee because their control authority is lost. Numerical studies cover fault-free operation, partial and complete actuator failures, external disturbances, parameter mismatch, gain-tuning inaccuracies. Furthermore, a 20- follower switching network illustrates the predicted performance, robustness margins, scalability, and practical limitations of the proposed approach.