Detecting slow-moving objects in complex environments remains a core challenge for radar perception, particularly for low-altitude uncrewed aerial vehicles (UAVs) and vulnerable road users whose weak motion signatures are easily masked by strong clutter and noise. Here, we propose a multidirectional radar state space model that jointly exploits time-domain raw ADC signals and frequency-domain representations, enabling complementary cross-domain feature learning. A subspace-based clutter suppression strategy is further introduced, reformulating conventional static clutter suppression as a dynamic metalearning-driven perception routing problem, thereby effectively decoupling weak objects from dominant background interference. In addition, an eight-directional scanning mechanism is incorporated to capture subtle long-term trajectory evolution of slow-moving objects within a CPI. Extensive experiments conducted on the public CRUW dataset and the newly constructed radarUAV-24K dataset demonstrate that the proposed approach consistently achieves competitive performance under low signal-to-noise ratio and slow-moving conditions, establishing a promising framework for robust radar perception of high-value objects in low-altitude and near-ground scenarios.
Process monitoring is an indispensable safeguard for ensuring operational safety in modern industrial systems, where multivariate statistical analysis has emerged as a cornerstone of data-driven monitoring techniques due to its high efficiency. Recently, stationary subspace analysis (SSA), a classical multivariate method, has proven effective in extracting stationary sources ($\mathcal {SS}$s) for reliable process monitoring. However, most SSA-based approaches assume linear relationships among variables and are unable to mine deep $\mathcal {SS}$s from nonlinear, nonstationary time series prevalent in industrial scenarios. To overcome these constraints, we propose an $x$-deep learning-aided SSA ($x$-DLSSA) framework. This work embodies three key innovations: first, a novel objective function reformulated via nonlinear network mapping, based on which adaptive mining of complex, time-invariant information is presented; second, a theoretically rigorous alternating optimization algorithm with proven convergence properties is developed for stable parameter estimation; and third, a modular design incorporating the specialized neural network components, enhancing flexibility to accommodate diverse industrial data characteristics. Rigorous theoretical analysis establishes explicit bounds for the optimization gap and convergence rate, providing a solid mathematical foundation. Validation studies using real-world blast furnace ironmaking datasets demonstrate superior performance in $\mathcal {SS}$ estimation and process monitoring, verifying its practical applicability.
This study investigates the problems of critical-bus-based (CBB) attack strategy design and the corresponding resilient $H_\infty$ estimator design for the smart grid. First, a hierarchical neighbor PageRank framework is generated to identify the critical buses in the smart grid by integrating global structural hierarchy and electrical functional influence after quantifying the importance degree of the buses. A CBB attack strategy is proposed from the attackers’ perspective, which can focus on targeting these selected critical buses to maximize the attack effect. Second, in order to mitigate the negative effect of the proposed CBB attack, an $H_\infty$ state estimator is designed to establish sufficient conditions for the mean-square asymptotic stability of the smart grid with disturbance attenuation. Simulations on the IEEE 39-Bus system demonstrate that: first, the proposed CBB strategy significantly outperforms traditional random and degree-based attacks in degrading the estimation performance; second, the proposed resilient $H_\infty$ estimator effectively alleviates the negative effect caused by the CBB attacks.
Non-Gaussian disturbances pose significant challenges to key performance indicator (KPI)-related process monitoring. To address this issue, this article proposes a streamlined adversarial autoencoder-concurrent projection to latent structures (SAAE-CPLS) framework centered on interpretable high-dimensional Gaussianized data transformation. Unlike a conventional autoencoder, the proposed SAAE removes the decoder and reconstruction objective, and functions as an encoder-based adversarial transformation network that maps non-Gaussian process and KPI variables into Gaussianized feature spaces. A Gaussian radial-based gated linear unit and an adaptive dimension-selection strategy are developed to improve the interpretability of the transformation and reduce empirical parameter tuning. The transformed features are subsequently decomposed by CPLS into orthogonal KPI-related and KPI-unrelated latent structures for monitoring-statistic construction. Experiments on the Skoltech Anomaly Benchmark and Tennessee–Eastman process demonstrate that the proposed framework achieves effective Gaussianized data transformation and KPI-related fault monitoring while reducing computational cost compared with decoder-based adversarial autoencoder models.
Industrial visual anomaly detection is essential for intelligent manufacturing, where inspection must operate with scarce defective samples and diverse appearance variations. Under the few-shot protocol, only a few normal samples are used for training, while defective samples are scarce and diverse. Although recent few-shot methods perform well on single-view benchmarks, modern inspection systems commonly acquire synchronized observations from multiple cameras. This setting introduces weak cross-view correspondence, viewpoint-dependent appearance changes, and view-specific noise such as illumination variation or specular reflection. To address these issues, this article proposes CVAD, a cross-view learning framework for few-shot multiview anomaly detection based on a frozen vision foundation model. CVAD contains three key designs: cross-view feature learning (CVFL), multiview consensus gating (MVCG), and prototype-consistency-guided (PCG) decoding. MVCG enhances features through sparse message passing and consensus-aware gating, CVFL retrieves compact candidate regions and aggregates them into multiview normal prototypes (MVP), and PCG constrains reconstruction toward normal modes. Experiments on Real-IAD under 1/2/4-shot protocols show that CVAD achieves a strong balance across detection and localization, with particularly clear gains in pixel-level anomaly localization.
