This article investigates the prescribed tracking problem for a class of stochastic switched nonlinear systems (SNSs) subject to generalized unmodeled dynamics (UDs), which are unmeasurable, depend on the full state vector, and are not assumed to satisfy the input-to-state practical stability condition. To handle such dynamics, the separation property of neural networks is applied to prevent the associated variable overflow and guarantee the solvability of the virtual controllers. A flexible performance function is then introduced to enable the achievement of multiple tracking performance objectives. Moreover, by reassigning the function variables, the coupling effect between the performance function and the UDs is eliminated. A switching law is codesigned with a controller such that all closed-loop signals remain bounded in probability and the tracking error is confined within a prescribed bound. Finally, simulation results demonstrate the advantages of the proposed approach.
Autonomous lunar landing relies heavily on vision-guided control and navigation (VGCN), in which visual landmarks, such as lunar craters, are used to estimate the lander’s position and attitude during descent. A major challenge in such systems is achieving robust crater detection and pose estimation under the stringent computational, power, and latency constraints associated with onboard spacecraft hardware. This article presents a real-time FPGA-based terrain-relative navigation (TRN) framework that combines learning-based offline threshold generation with deterministic online crater detection and pose estimation. The proposed approach employs a faster region-based convolutional neural network (Faster R-CNN) to derive statistically optimal reference threshold tuples (RTTs) offline through a bounded-suboptimality-based threshold-selection framework, which is subsequently integrated into a hardware-efficient multilevel Otsu segmentation architecture suitable for FPGA implementation. Unlike conventional deep-learning-based approaches that require computationally intensive onboard inference, the proposed framework enables deterministic real-time execution while maintaining robust crater-detection performance. The architecture is implemented on a Xilinx ZCU104 UltraScale+ MPSoC platform and validated through hardware-in-the-loop (HIL) simulations using an Opal-RT-based lunar landing environment. Experimental results demonstrate deterministic execution with image-processing rates of up to 60 frames per second (FPS), corresponding to a bounded processing latency of approximately 16.67 ms/frame, while maintaining acceptable navigation accuracy within a complete VGCN framework. The proposed methodology establishes a practical pathway toward low-power, deterministic, and real-time onboard perception architectures for future autonomous planetary landing missions.
In this article, we investigate the problem of asymptotic mean-square stabilization for Markovian cyber–physical systems (MCPSs) under stochastic denial-of-service (DoS) attacks. The increasing complexity and openness of networked environments make MCPSs particularly vulnerable to DoS attacks, while the presence of Markovian parameters further complicates the design of stabilizing control strategies. To address this challenge, we propose a generalized operator spectrum approach to avoid the traditional Lyapunov function construction and to derive necessary and sufficient conditions for stabilization. Furthermore, we design a nonfragile feedback controller based on linear matrix inequalities (LMIs) to enhance the system’s robustness. Additionally, a generalized stability notion, termed $D(0,\beta)$ -stabilization, is introduced to explicitly characterize the convergence rate through the location of the operator spectrum. A set of sufficient $D(0,\beta)$ -stabilization conditions is further derived based on LMIs, enabling systematic controller synthesis. Finally, numerical, power-system, and networked vehicle examples validate the effectiveness of the proposed approach, where comparisons with traditional Lyapunov-based controllers show faster convergence and smoother state trajectories.
Neural architecture search (NAS) aims to automate the design of network architectures and can be regarded as a systems engineering process. As one of the most vexing issues, enhancing its search efficiency has remained a persistent challenge. Moreover, architectural complexity control in NAS has also been a focus for deployment on resource-constrained platforms, with multiobjective optimization being recognized as a promising approach. However, the imbalance in difficulty between optimizing predictive accuracy and model complexity is rarely considered. In this work, we propose a reliable zero-cost metric, i.e., the double perturbation perception (DPP) score, to dramatically reduce the search cost of NAS in a training-free manner. In DPP, network performance is measured by counting the difference between the activation states of two noisy inputs. To handle the imbalanced difficulty in optimizing the objectives and improve architectural exploration of low and medium complexity regions, we develop a novel evolutionary multiobjective search algorithm. It combines Pareto dominance with a reference point-based technique to calculate appropriate fitness for promising architecture selection in different regions. The experimental results on various search spaces and datasets show the high efficiency of DPP and the high correlation between the score and test accuracy. DPP-NAS requires merely 0.009 GPU days on the CIFAR search space and 0.012-CIFAR search space and 0.012 GPU days for the MobileNet-ImageNet search space to explore multiple architectures, achieving superior performance in both parameter scale and classification accuracy.
Scene text reading is an essential task for scene understanding. As the fundamental work of text reading, scene text detection has received significant attention recently. Unlike general objects, scene texts exhibit geometric features (such as aspect ratio, scale, orientation, and shape) that vary significantly. Segmentation-based methods provide flexible pixel-level predictions for arbitrary-shaped text. However, pixel-level supervision alone may be insufficient to distinguish text from background patterns with similar local textures. To address this issue, we propose an instance sequence cognition module (ISCM) and a selective knowledge mutual module (SKMM). The former introduces character-count-guided instance supervision to learn complementary instance-level cues, which helps distinguish text from visually similar patterns. The latter divides the features into channel groups with different interaction roles and applies cross-attention to selected groups to integrate complementary instance-level and pixel-level cues. This interaction provides additional guidance for text-kernel segmentation. A series of ablation experiments demonstrates the effectiveness of the proposed modules. Extensive experiments demonstrate that the proposed method achieves competitive detection performance with a favorable accuracy–efficiency tradeoff across various datasets.
