Near-infrared (NIR) spectroscopy plays a critical role in online analysis of difficult-to-measure properties of petrochemicals. In industrial applications, a calibration model among NIR spectra and properties must be established. However, it is a challenge to obtain a precision NIR model in the majority of petrochemical processes since industrial data present strong non-Gaussian and uncertainty characteristics. To deal with these problems, in the present work a probabilistic regression modeling method based on a biased minimax probability machine (BMPM) is proposed, without assuming any specific distributions for the data. In addition, a multiple locally weighted updating approach with a new supervised similarity distance is introduced to cope with process changes. The greatest advantage of the proposed approach is that it has superior capability in dealing with uncertainties and variations. The effectiveness of the method is illustrated through its application in an actual gasoline blending process and a simulated fermentation process.
30 April 2024 in IEEE Transactions on Instrumentation and Measurement
In this article, a novel shared-database decentralized federated learning (SDeceFL) framework is developed for wafer defect pattern recognition (DPR). Specifically, a differential privacy shared-database strategy is proposed to overcome the interclass heterogeneity problem of different clients and enhance data privacy. A deformable convolutional autoencoder (DCAE) is designed for data augmentation for handling class imbalance. The vision transformer (ViT) is employed for wafer DPR. The proposed DCAE-ViT-SDeceFL framework is validated on three public datasets (e.g., WM-811K, NEU-CLS-64, and CIFAR-100). The experimental results show the superiority of the SDeceFL framework over Ratio Loss-FedAvg, MOON, FedNH, BalanceFL, federated averaging (FedAvg), DeceFL, and swarm learning (SL). Compared with some deep learning methods, experimental results exhibit the effectiveness of the proposed DCAE-ViT-SDeceFL method for wafer DPR on WM-811K.
22 October 2025 in IEEE Transactions on Industrial Informatics
In many real-world industrial scenarios, acquiring accurately labeled data are often challenging due to limited resources or unexpected errors. Learning with noisy labels (LNL) has emerged as a significant research topic, aiming to develop reliable deep learning models using noisy-labeled training data. In this article, a novel Transformer-embedded LNL framework with fuzzy-clustering-assisted contrastive learning is developed for industrial time series outlier detection under noisy labels. Specifically, a fuzzy-clustering-assisted contrastive learning strategy is proposed to enhance the robustness of the Transformer encoder against noisy labels by leveraging the intrinsic characteristics of raw data. Furthermore, a dynamic two-stage training scheme is introduced to train the outlier detector. In the first training stage, the Transformer encoder is pretrained through data reconstruction to improve feature extraction capabilities for industrial time series. In the second stage, the outlier detector is jointly trained with the Transformer encoder, incorporating a joint learning strategy. Furthermore, a label-consistency regularization term is designed to enhance the robustness of the outlier detector against noisy labels by minimizing the discrepancy between the outputs of the outlier detector and the clustering algorithm. The proposed framework is applied to industrial time series data collected from a real-world wire arc additive manufacturing (WAAM) process. Experimental results demonstrate that the developed framework outperforms selected representative LNL approaches in WAAM outlier detection under both low and high noise ratios.
11 October 2024 in IEEE Transactions on Instrumentation and Measurement
This paper investigates the problem of dynamic event-triggered fault detection (FD) for a class of nonlinear systems. Under the dynamic event-triggered strategy, an observer-based nonlinear fault detection filter (FDF) is developed as a residual generator through the processes of time-update and event-update. Different from the existing schemes of nonlinear event-triggered FD, the generated residual signal is fully decoupled from the event-triggered transmission error, which allows an independent development of the residual generator and the dynamic event detector. Moreover, the design of event-triggered FDF is formulated as an Hi/H ∞ optimization problem so that an optimal tradeoff can be achieved between the robustness of residual against unknown input and the sensitivity to fault. Based on Taylor series expansion and first-order linearization, the existing conditions of the dynamic event-triggered FDF and a feasible solution of the Hi/H ∞ optimization problem are given in terms of Riccati recursions. Finally, a three-tank system is considered to illustrate the validity of our proposed approach.