Guoqing Mu

Proton exchange membrane fuel cells (PEMFCs) offer broad prospects due to their high efficiency and environmental friendliness. Stack power, a key performance metric, is highly influenced by operating conditions such as temperature and pressure. Thus, building a data‐driven model that captures the nonlinear dynamics between process variables and stack power is essential for real‐time prediction and optimization. However, the high cost of testing hydrogen fuel cells has resulted in a paucity of experimental data, which has hindered the development of high‐precision data‐driven models. To address these challenges, this paper presents an enhanced prediction method for PEMFC stack power that integrates the multihead attention mechanism into the TimeGAN‐GRU framework. Specifically, the multihead attention‐driven TimeGAN enhances the model's capacity to capture the global dynamic characteristics of time series, thereby mitigating the issue of limited data. The GRU model integrated with the multihead attention mechanism enhances the accuracy of modeling and predicting the complex nonlinear relationship between external process variables and stack power. The results show that, compared with traditional deep learning models, the proposed method improves stack power prediction accuracy across test sets and offers better generalization and data generation ability.

Hemodialysis (HD) is the primary life-sustaining treatment for patients with end-stage renal disease (ESRD). However, current real-time monitoring methods during dialysis are costly, complex, and not widely adopted. Therefore, this study aims to propose a novel Kt/V prediction method that is accurate, low-cost, and compatible with various dialysis devices, in order to enhance the intelligence and efficiency of clinical hemodialysis treatment. We propose an improved Domain Adversarial Neural Network (DANN) framework for Kt/V prediction. The feature extractor integrates a Long Short-Term Memory (LSTM) network with a Kolmogorov-Arnold Network (KAN), where the B-spline activation functions in KAN enhance nonlinear modeling capabilities. Additionally, a multi-head attention mechanism is incorporated into the label predictor to capture long-range temporal dependencies. This architecture effectively mitigates data distribution shifts caused by inter-patient and inter-device variability. Experiments conducted on clinical and simulated data from 17 ESRD patients demonstrate that the proposed model achieves high prediction accuracy and significantly outperforms conventional deep learning models, including RNN, LSTM, and the original DANN. The proposed model enables precise, non-invasive estimation of dialysis adequacy (Kt/V), showing excellent generalization and clinical application potential. It provides a practical solution for real-time adequacy assessment, especially suitable for low- and mid-range dialysis devices.

The accurate measurement and monitoring of quality variables are crucial for optimizing productivity and efficiency in industrial processes. However, many existing data-driven methods fail to account for the dynamic nature of industrial data and the nonuniform distributions that arise due to fluctuating operational conditions. To address these challenges, attention-based dynamic latent variable models are proposed. This approach comprehensively considers time-variant relationships among latent variables using attention mechanisms for both samples and variables. An interactive iterative algorithm is introduced to compute model parameters efficiently. Experimental validation, conducted through both numerical simulations and real-world applications in an industrial ethylene oxychlorination process, demonstrates the effectiveness of the proposed approach in enhancing process modeling and prediction performance. Comparison results show that the fault detection rate (FDR) of the proposed attention-based methods is significantly higher than that of traditional approaches. The proposed bi-attention partial least squares (PLSs) achieves the highest FDR rate of 0.94 based on the squared prediction error (SPE) statistic in the application of industrial application.

To measure the moisture content of the product in the fluidized bed drying (FBD) process in real time, a hybrid integrated soft measurement method based on near-infrared (NIR) spectroscopy is proposed in this paper.

For using the infrared spectroscopy to conduct in-situ measurement of slurry component concentrations during batch crystallization or fermentation process in engineering practice, a novel just-in-time learning (JITL) based functional modelling method is proposed for spectral data analysis in this paper to improve on-line measurement accuracy, by addressing engineering issues of very limited assay samples, spectral absorption nonlinearity and high-dimensional spectral variables. The traditional univariate or multivariate partial least-squares (PLS) model is subtly extended to the corresponding functional form for describing the relationship between the measured spectra and slurry component concentrations, such that the salient problem of only a small number of slurry component concentration samples available in practice for spectral data modeling could be effectively solved. The wavelet functions are adopted to properly approximate the measured spectral curves owing to their multi-scale and orthogonal properties. Accordingly, a functional spectral data model is established by wavelet functional PLS for on-line measurement of single and multiple component concentrations, respectively. Meanwhile, a JITL strategy is introduced to select the most representative samples for such modeling, which could efficiently handle the disparity issue involved with real-time samples and guarantee the model validity. Experimental results show that root mean square error of model prediction on the solution concentration during L-glutamic acid cooling crystallization process and multiple componts of ethanol fermentation process, could be evidently reduced over 40% and 50%, respectively, compared with the traditional PLS and support vector regression methods.

To measure the moisture content of granules during the industrial fluidized bed drying (FBD) process, a semi-supervised calibration model is proposed for using the near-infrared (NIR) spectroscopy to conduct in situ measurement. To solve the dilemma of lacking sufficiently labeled samples as often encountered in various batch FBD processes, a semi-supervised variational inference partial least-squares (PLS) method is proposed to use up all the labeled and unlabeled spectra measured for calibration model building. Moreover, an adaptive Gamma distribution-based sparsing algorithm is established to select the spectral variables for modeling, to overcome the high-dimensional input collinearity. Owing to the use of a variational inference learning approach, the constructed model can ensure not only prediction accuracy but also credibility. A numerical example and experiments on batch FBD processes of silica gel granules are shown to demonstrate the effectiveness and merit of the proposed method.