1 August 2021 in Analytica Chimica Acta
19 July 2022 in IEEE Transactions on Instrumentation and Measurement
To solve the problem of insufficient labeled samples in online total nitrogen (TN) detection, a novel semisupervised learning (SSL) method called double regularized structure graph learning (DRSGL) was proposed, which can effectively extract useful features and support the TN detection equipment to establish accurate detection models with few labeled samples. Based on temporal and spectral informativity, a high-quality graph structure method was designed first, which utilizes the spectrum-temporal prior knowledge hidden in the spectrum and enhances the efficiency of the graph model. Then, considering the double sparsity of TN spectrum samples, a double structural sparse feature selection method (DS 2 FS) was invented accordingly, which can excavate useful information from both spectral and temporal dimensions. Finally, to address the insufficient problem of labeled samples, an adaptive SSL method was constructed by combining graph learning strategy and DS 2 FS and applied to the TN rapid detection prototype, which can iteratively select important features and update graph structure of samples correspondingly. According to the experimental results based on practical application, DRSGL can effectively solve the problem brought by the insufficiency of labeled samples, which can utilize only 20% labeled samples to establish an accurate detection model satisfied national detection standard with a relative error lower than 10%.
30 May 2023 in IEEE Transactions on Industrial Electronics
The outstanding spectral resolution of full-spectrum detection (FSD) ensures excellent performance for complex scenarios in which target signals are affected by particle interference. However, FSD is limited by the cost of its measurement devices and demanding install requirements. Therefore, it is not always practical to utilize FSD for online water quality monitoring in wastewater treatment plants. In this article, we develop a novel detection approach entitled “compressive spectrum detection (CSD)” to overcome the hardware restrictions posed by FSD. An intelligent source-encoding laser is first developed to perform spectrum encoding, which avoids the utilization of sophisticated equipment and reduces the detection cost. Subsequently, a dual-stage compressive detection framework is proposed based on incremental learning to accurately reconstruct spectral signals with full-spectrum resolution, which consisted of three innovative algorithms and could solve difficulties brought about by sample insufficiency and changing environments in practical situations. Based on our constructed online detection prototype, the superiority of CSD is evaluated through practical application in which the training samples are reduced to a quarter of their original size with an overall improvement in detection precision greater than 50%.
31 October 2025 in IEEE Transactions on Computational Imaging
With the widespread adoption of utilizing coded aperture snapshot spectral imaging (CASSI) to acquire hyperspectral images (HSIs), reconstruction algorithms become increasingly important as one of the most critical parts of CASSI. By formulating the iterative optimization problem as a multistage end-to-end neural network, the deep unfolding framework has made gratifying progress in improving the reconstruction reliability and fidelity. Nevertheless, there still exist three knotty issues limiting the performance of current reconstruction algorithms: 1) predetermined sensing matrix ignoring the gap between real and ideal degradation procedure. 2) ill-structured feature extraction and fusion network making it difficult to effectively recover the spatial texture and ensure spectral consistency. 3) inadequate stage interaction module, which hinders the transmission of useful information. To address these issues, in this paper, we propose a hyperspectral explicable degradation-compensatory unfolding neural network (HEDUNet), which attempts to dynamically compensate for the differences between ideal and real degradation procedures using a residual network under the deep unfolding strategy. Moreover, a spatial-spectral interaction transformer is designed via plugging the Swin transformer under the framework of the MixS2 transformer to independently excavate spatial or spectral features and fuse them thoroughly. Finally, to mitigate the information loss between adjacent stages, stage interaction is also enhanced by proposing a specific feature aggregation module based on the multi-scale attention mechanism. Experiments on both simulation and practical datasets show the state-of-the-art (SOTA) performance and wide application prospects of our method.
18 March 2026 in SSRN Electronic Journal
Pollutant leaks can cause significant environmental damage. Achieving rapid and accurate prediction of pollutant dispersion concentration is key to miti- gating the impact of such accidents. Bridging the gap between the environ- mental rigidity of mechanistic models and the physical opacity of data-driven methods, this paper presents RDiGSr-GKAN, a novel framework that fuses a Gaussian prior with Kolmogorov-Arnold Networks (KAN) for enhanced pollutant dispersion prediction. This model constructs Gaussian feature through the Gaussian plume model to embed physical relationships, em- ploys the RDiGSr algorithm to optimize sample selection, and subsequently utilizes KAN network to learn complex mappings between variables, thereby enhancing both prediction performance and interpretability. Experiments on seven datasets, including Project Prairie Grass, demonstrate that the proposed model outperforms traditional machine learning and deep learning methods across multiple evaluation metrics. For instance, on the Hanford- 64 and Cabauw datasets, which have vastly diferent concentration ranges (0–14.9 mg/m3 and 0–2838 mg/m3 , respectively), the RMSE values achieved are 0.049 and 11.468, significantly better than those of comparative mod- els. Ablation experiments confirm the crucial role of Gaussian feature in model performance. This research provides a physics-guided, data-driven, high-precision modeling method for pollutant dispersion prediction, demon- strating excellent generalization capability and application potential.