29 July 2022 in IEEE/ASME Transactions on Mechatronics
Deep learning-based methods have been developed and widely used for fault diagnosis, which rely on the sufficient data. However, fault data are extremely limited in some real-case scenarios. In this article, a meta-learning with adaptive learning rates (MLALR) method is proposed for few-shot fault diagnosis. MLALR learns from auxiliary tasks to find initialization parameters of the model that can adapt to target tasks with a few data. The keys of MLALR are the proposed adaptive learning rates for meta-training and fine-tuning, whose values are adjusted according to the distributions of extracted features to tackle the two common problems of few-shot learning, i.e., overfitting and underfitting. The loss functions are further improved to promote the model generalization capability and training stability. The effectiveness of the proposed method is validated using two bearing datasets. MLALR obtains higher accuracies and stabilities than the baseline methods and three other state-of-the-art methods.
5 January 2025 in Engineering Applications of Artificial Intelligence
24 June 2021 in IEEE Transactions on Industrial Electronics
Data-driven methods, especially those based on deep learning, have been successfully applied to various tasks. However, training and testing data may have different distributions in many real cases, which lead to unreliable results. Transfer learning can address this issue by learning domain-invariant features. Online transfer learning (OTL), as one type of transfer learning, deals with the situation where the data of the target domain arrive in a sequential manner. The limitations of the existing OTL methods are that they cannot deal with the online data generated by time-varying distribution because of varying operational conditions, for example, besides, they mainly focus on classification tasks. In this article, we propose a novel OTL framework for a regression task, which includes an offline stage and an online stage. In the former, a basic model is built to learn domain-invariant features. In the latter, an updating strategy is proposed to adapt the prediction model to the new data. Besides, an ensemble approach is further developed to avoid over- or underfitting and also fully utilize the knowledge from the source domain for the target domain. The effectiveness of the proposed framework is verified by a real dataset regarding the diesel hydrofining process collected from a petrochemical workshop.
4 July 2023 in IEEE Transactions on Instrumentation and Measurement
Deep learning (DL) methods can be used to construct health indicators (HIs) for remaining useful life (RUL) prediction. Existing DL methods consider previous and current sensor signals and utilize labeled data, which are limited in practice. To leverage unlabeled data for extracting HIs, semi-supervised methods, especially hybrid methods, can be employed. In this paper, a semi-supervised deep hybrid multi-task model (DHMTM) for RUL prediction is developed. DHMTM contains two temporal models for unlabeled and labeled multivariate time series data, respectively. In the model training process, adding an extra task of prediction of future sensor signal values, DHMTM can obtain HIs which improve the RUL prediction accuracy. Besides, temporal dependency of sensor signals is captured in the proposed DHMTM. The effectiveness of the proposed model is validated using the C-MAPSS and the lithium-ion batteries datasets. The results show that using the proposed method, the prediction errors for the two datasets have been reduced by 2.5% and 23.5% on average, respectively, compared to the fully supervised regression model, and 17% and 44%, respectively, on average compared to three other widely-used semi-supervised methods.
28 September 2021 in Quality and Reliability Engineering International