Deep learning plays an increasingly important role in industrial applications, such as the remaining useful life (RUL) prediction of machines. However, when dealing with multi-feature data, most deep learning approaches do not have effective mechanisms to weight the input features adaptively. In this paper, a novel feature-attention based end-to-end approach is proposed for RUL prediction. Firstly, the proposed feature-attention mechanism is directly applied to the input data, which gives greater attention weights to more important features dynamically in the training process. This helps the model focus more on those critical inputs, and the prediction performance is therefore improved. Next, Bi-directional Gated Recurrent Units (BGRU) are used to extract long-term dependencies from the weighted input data, and Convolutional Neural Networks (CNN) are employed to capture local features from the output sequences of BGRU. Finally, fully connected networks are used to learn the above abstract representations to predict the RUL. The proposed approach is validated in a case study of turbofan engines. The experimental results demonstrate that the proposed approach outperforms other latest existing approaches.

Digital twin (DT) is an emerging technology that enables sophisticated interaction between physical objects and their virtual replicas. Although DT has recently gained significant attraction in both industry and academia, there is no systematic understanding of DT from its development history to its different concepts and applications in disparate disciplines. The majority of DT literature focuses on the conceptual development of DT frameworks for a specific implementation area. Hence, this paper provides a state-of-the-art review of DT history, different definitions and models, and six types of key enabling technologies. The review also provides a comprehensive survey of DT applications from two perspectives: (1) applications in four product-lifecycle phases, i.e. product design, manufacturing, operation and maintenance, and recycling and (2) applications in four categorized engineering fields, including aerospace engineering, tunneling and underground engineering, wind engineering and Internet of things (IoT) applications. DT frameworks, characteristic components, key technologies and specific applications are extracted for each DT category in this paper. A comprehensive survey of the DT references reveals the following findings: (1) The majority of existing DT models only involve one-way data transfer from physical entities to virtual models and (2) There is a lack of consideration of the environmental coupling, which results in the inaccurate representation of the virtual components in existing DT models. Thus, this paper highlights the role of environmental factor in DT enabling technologies and in categorized engineering applications. In addition, the review discusses the key challenges and provides future work for constructing DTs of complex engineering systems.

This paper proposes a novel Transformer-based neural network model for accurate tool wear estimation to improve production quality and efficiency in intelligent manufacturing. The proposed method can realize indirect measurement of tool wear. Initially, the raw multi-sensor signals are processed into three kinds of temporal feature data. Next, three identical submodels are utilized to deal with the above feature data, respectively. Finally, the outputs of these three submodels are concatenated together as the input of a multi-layer fully connected network for the final estimation of tool wear. Concretely, the submodels used in this work are based on the Transformer model and self-attention mechanism to capture long-term dependency. This is the first attempt to adopt Transformer and self-attention for tool wear estimation. Besides, some improvements are made in this work. For example, a long short-term memory network is employed to enhance the ability of capturing position information. In addition, a submodel framework is applied to process the temporal feature data in parallel, which helps improve the model performance. The proposed method is demonstrated through a real-world milling dataset, including more than 900 experiments. Also, the superiority of the proposed method is verified by the comparison with other advanced methods.