Yang Yu

The natural heterogeneity of raw materials, such as fluctuations in hydroxysafflor yellow A concentration, poses significant engineering challenges in standardizing safflower ( Carthamus tinctorius L .) processing. This study proposes an integrated data‐driven framework combining near‐infrared hyperspectral imaging and reinforcement learning to achieve dynamic blending control. First, a rapid, non‐destructive monitoring system was developed using a SNV‐CARS‐SVR model, achieving excellent predictive accuracy for HSYA content (R _c ² = 0.9808, R _p ² = 0.9218). Subsequently, the industrial blending process was modeled as a Markov Decision Process. A dynamic batching strategy utilizing the Proximal Policy Optimization algorithm was implemented to optimize the feedstock sequence. Experimental validation demonstrated that the proposed RL‐driven blending system effectively suppressed feed concentration fluctuations, reducing the standard deviation of HSYA concentration from 0.4730 to 0.086 (an 81.8% reduction). This framework provides a scalable, intelligent control solution for continuous quality assurance in the processing of complex natural products.

Video anomaly detection (VAD) aims to identify rare and unpredictable abnormal events in long video streams under open world conditions. While existing surveys predominantly organize content by model architectures or supervision settings, they often obscure the cross-modal alignment challenges, capability boundaries, and practical limitations of diverse sensory cues. To address this gap, this paper presents a unified review from the perspective of multi-modal perception and modeling. We group the existing literature along the dimensions of information sources and semantic granularity into three main categories. These categories are the video modality, model inferred semantic modalities such as skeletons objects and text, and other physical sensing modalities including audio, event streams, and thermal imaging. We systematically analyze the common modeling approaches perceptual information capabilities and environmental limitations of each modality. Furthermore, we summarize the common paradigms of cross-modal fusion such as early, mid, and late fusion, and compare their interaction mechanisms and dependencies. Crucially, we discuss how practical challenges like alignment errors modality degradation and missingness affect system reliability across different fusion stages and outline potential future directions to address these issues. Overall this review provides a structured reference for building robust multi-modal VAD systems and advancing reliable anomaly understanding in real world scenarios.

This study presents an integrated framework combining near-infrared hyperspectral imaging (NIR-HSI) and laser-induced breakdown spectroscopy (LIBS) with deep learning for the origin traceability and component prediction of Salvia miltiorrhiza(SM). Initially, a 1D-CNN model based on LIBS data achieved a 95.95% accuracy in origin identification. To enhance discriminative performance, a multimodal fusion approach utilizing a GWO-SE-CNN framework was proposed combining LIBS and HSI data to achieve superior accuracy, recall and F1-scores. For quantitative quality assessment, a ResNet-based model demonstrated robust predictive capabilities yielding R2p values of 0.9311 for Tanshinone IIA, 0.9099 for Salvianolic acid B and 0.9324 for moisture content. Furthermore, model performance was optimized through a dual strategy involving feature selection via the IWCARS algorithm and data augmentation using a WGAN-GP network to generate high-fidelity synthetic spectra. This research demonstrates the potential of multispectral data fusion and advanced deep learning as a non-destructive tool for the comprehensive quality control of SM.

This study presents an integrated framework combining near-infrared hyperspectral imaging (NIR-HSI) and laser-induced breakdown spectroscopy (LIBS) with deep learning for the origin traceability and component prediction of Salvia miltiorrhiza(SM). Initially, a 1D-CNN model based on LIBS data achieved a 95.95% accuracy in origin identification. To enhance discriminative performance, a multimodal fusion approach utilizing a GWO-SE-CNN framework was proposed combining LIBS and HSI data to achieve superior accuracy, recall and F1-scores. For quantitative quality assessment, a ResNet-based model demonstrated robust predictive capabilities yielding R2p values of 0.9311 for Tanshinone IIA, 0.9099 for Salvianolic acid B and 0.9324 for moisture content. Furthermore, model performance was optimized through a dual strategy involving feature selection via the IWCARS algorithm and data augmentation using a WGAN-GP network to generate high-fidelity synthetic spectra. This research demonstrates the potential of multispectral data fusion and advanced deep learning as a non-destructive tool for the comprehensive quality control of SM.

To address the challenges of real-time observation of gas-solid two-phase flow and the high computational cost/lack of real-time capability of traditional Computational Fluid Dynamics (CFD) simulations in fluidized bed drying processes, this paper proposes a rapid flow field reconstruction method based on Physics-Informed Neural Networks (PINNs). This method integrates sparse observation data with governing flow equations to achieve simultaneous prediction of key physical quantities including gas-phase velocity, solid-phase velocity, pressure, and solid volume fraction. A systematic evaluation was conducted under various conditions: spatial sparsity (20%, 5%, 1%), temporal sparsity (50%, 20%, 10%), noise interference (1%–20%), and cross-temporal extrapolation. Results show that with a spatial sampling rate as low as 5%, PINNs can still accurately reconstruct typical fluidization features such as bubble motion, solid-phase stratification, and pressure gradients. Even under an extreme sparsity of 1%, the overall flow pattern remains correct. Under dual spatio-temporal sparsity, the model demonstrates good information compensation capability, generating continuous and plausible transient evolution. With noise levels up to 20%, reconstruction errors exhibit only minor fluctuations, indicating strong noise resistance. Cross-temporal extrapolation experiments reveal that error growth over unseen time periods is stable and controllable, without non-physical oscillations. This study validates the stability, physical consistency, and robustness of PINNs under complex industrial scenarios involving sparsity, noise, and extrapolation, providing a viable technical pathway for real-time flow field reconstruction and intelligent monitoring in fluidized bed drying processes.