Yan Qin

Robust small object detection in adverse environments remains challenging due to physical degradations and unstable feature representations. Existing detectors often struggle to maintain semantic consistency in haze, rain, and low-light conditions, leading to degraded accuracy and poor generalization. To address these issues, we propose a unified neural information bottleneck with physics-guided diffusion (NIB-PGD) framework. The proposed framework couples physical priors with information-theoretic regularization in a dual-diffusion paradigm, enabling robust feature encoding and noise-resilient object localization. A multiscale enhance–fuse–context attention (EFC-A) encoder captures hierarchical semantics and contextual dependencies, while a physics-guided feature-level diffusion injects realistic degradation kernels into the noise schedule to enhance robustness. In parallel, a box-level diffusion reformulates detection as iterative denoising, progressively refining bounding boxes without anchors or queries. A second-order information bottleneck (2O-IB) constraint dynamically regulates mutual information to suppress redundancy and preserve task-relevant semantics. On our composite benchmark (constructed from BDD100K, SODA-D, and an in-house small object set), NIB-PGD achieves 52.3% AP and 35.7% AP ${}_{s}$ , outperforming the strongest AP baseline by + 1.2 AP and the strongest AP ${}_{s}$ baseline by +1.1 AP ${}_{s}$ . Comprehensive experiments on multiple benchmarks demonstrate that NIB-PGD achieves state-of-the-art performance in detection accuracy, robustness, and generalization.

Cyber-physical systems (CPSs) are increasingly vulnerable to sophisticated cyber-attacks due to their growing complexity and connectivity. Current detection approaches face significant challenges in distinguishing stealthy attacks from normal operational variations, as adversaries can craft attacks that preserve statistical correlations while disrupting physical processes. To address this critical gap, we propose a novel causal representation learning (CRL) framework that leverages intrinsic physical invariants for robust attack detection in industrial CPS. We develop two complementary methodologies, unsupervised secure causal feature extraction (USCFE) and supervised secure causal feature extraction (SSCFE), which extract low-dimensional causal features resilient to adversarial manipulation. Furthermore, theoretical guarantees on detection probability bounds and false alarm control are provided. Experimental validation shows that proposed methods outperform both traditional statistical methods and contemporary deep learning approaches, thus significantly advancing CPS security.

Modern industrial systems require real-time prediction for critical variables that are expensive or impossible to measure online. Existing soft sensing approaches suffer from catastrophic performance degradation under distribution shifts and sensor perturbations. This study develops a novel framework that bridges causal inference with robust optimization for reliable industrial soft sensing. We propose a causal robust soft sensor (CRSS) framework that employs spatiotemporal filtering to extract Granger-causal features that capture predictive temporal relationships reflecting fundamental process mechanisms. The framework incorporates explicit robustness constraints via regularization that penalizes sensitivity to bounded perturbations. Extensive experimental verification on real industrial refinery control system data demonstrates that CRSS achieves the best predictive accuracy (RMSE $=0.42$, $R^{2}=0.90$). It maintains stable performance under 20% sensor noise with small degradation, and remains robust under complete sensor outages, nonlinear distortions, and heavy-tailed perturbations beyond the Gaussian model. CRSS provides theoretically grounded prediction error bounds, offering a practical solution for industrial soft sensing applications with interpretable Granger-causal insights.

Developing trustworthy models for industrial processes is challenging due to changing operating conditions and limited data. In this work, we present an innovative and effective framework for establishing trustworthy process models from limited data. Our approach introduces the uniform manifold approximation and projection (UMAP) algorithm to uncover the essential low-dimensional structure in real industrial data, which allows high-quality virtual sample generation (VSG) that captures the underlying process dynamics. Furthermore, we propose a new algorithm for stable learning that uses sample reweighting to effectively mitigate spurious correlations that can undermine the stability and reliability of the model. Two case studies, the Tennessee Eastman process (TEP) and a commercial fluid catalytic cracking (FCC) unit, demonstrate the effectiveness of the proposed framework. In the Tennessee Eastman benchmark, the proposed method reduces the average error by 12.5% compared with VSG alone and by 47.8% compared with the OLS baseline. In the FCC case study, it further achieves a 41.5% RMSE reduction compared with the strongest competing UMAP-augmented model. Furthermore, the proposed framework maintains robust performance under changing operating conditions. Our framework enables the development of trustworthy industrial process models from limited data, which offers a powerful method to improve operational safety and efficiency in real-world applications.

Industrial processes are governed by causal relationships whose understanding is essential for root cause analysis and predictive maintenance. However, existing causal discovery methods operate in batch mode and assume static causal structures, conflicting with the inherent nonstationarity caused by equipment degradation and regime changes. Current approaches cannot track how causal effects vary with operating parameters in real-time, nor detect structural change points. This article introduces Online-PCT-CD, a framework that extends parametric causal discovery to streaming data settings. The proposed method integrates incremental Polynomial Chaos Expansion estimation via recursive least squares, dual-layer cumulative sum change point detection, and an adaptive sliding window strategy. Experiments on an industrial chemical process benchmark demonstrate that Online-PCT-CD achieves 93.7% average F1-score, outperforming 17 baseline methods while maintaining 450 samples per second throughput suitable for real-time deployment.