A cyber-physical system-enabled deep learning-based methodology for cognitive intelligent condition monitoring of manufacturing thing

Abstract

Manufacturing thing (e.g., machine tools and robots) is the basic manufacturing unit in a production system, of which condition monitoring plays an essential role in terms of keeping the smooth manufacturing process operation by accurately identifying the status of manufacturing thing. Recently, with the rapid development of information and communication technologies, such as the industrial Internet-of-Things, cloud computing, etc., the whole manufacturing industry is heading for smart manufacturing envisioned with a higher level of cognitive intelligence. However, current condition monitoring of manufacturing thing needs to be propelled to possess the self-X (i.e., self-awareness, self-learning, self-adaptation, and self-adjustment) capabilities featured cognitive intelligent development paradigm in the following aspects.

Firstly, existing studies seldom considered realizing condition-based self-awareness in a cost-effective manner. Secondly, most previous methods scarcely allowed the input of multi-domain spaces, feature weighting, and hybrid feature fusion simultaneously in feature-based self-learning for more sophisticated feature expression. Thirdly, major self-adaptation methods discussed the model adaptivity without thoroughly considering the different distribution of training data and testing data, the missing relationships between the input signals, and the insufficient data size, restricting the model adaptation under different operating settings. Lastly, most research on current knowledge graph (KG) embedding to advance effective knowledge management lacks a holistic consideration of the relation-specific, entity-differentiated, and across-layer aggregations, which is prone to limit model performance and insufficient to provide comprehensive knowledge reference for the self-adjustment.

Motivated by the above issues, this thesis proposes a cyber-physical system (CPS)-enabled deep learning (DL)-based methodology, in which CPS constructs a seamless link between the physical space and cyber space to provide manufacturing thing with connectivity, computation as well as control. Meanwhile, the developed DL methods serve as the core functions to deliver the envisaged self-X capabilities based on data-driven intelligence so as to achieve the cognitive intelligent condition monitoring of manufacturing thing. The main contents are briefly summarized as follows.

Firstly, a cost-effective approach for condition-based self-awareness is investigated based on deep transfer learning. The proposed approach overcomes some drawbacks of the conventional methods (i.e., machine learning (ML) and DL), for example, the dependence on manual handcrafted feature engineering as well as the massive training cost of data samples and training time. The proposed approach demonstrates its cost-effectiveness in terms of higher model performance, fewer training samples, and less training time. (Chapter 3)

Secondly, an adaptive parallel feature learning and fusion approach is proposed to deal with the input of multiple data domains, feature importance weight generation, and handcrafted statistical feature fusion. The proposed approach allows to derive more refined representation during the feature-based self-learning process and boosts the model performance. (Chapter 4)

Thirdly, a multi-hop branch ensemble-based graph adaptation framework for edge cloud-orchestrated self-adaptation is developed to enhance the model adaptivity under different operating settings based on the maximum mean discrepancy to evaluate the data difference, the graph-based format to incorporate the data correlation, and the pseudo labeled data from edge-cloud orchestration to supplement the data size. (Chapter 5)

Finally, a multi-hierarchical aggregation-based graph convolutional network embedding for KG-assisted self-adjustment is established to provide more accurate knowledge reference from the perspectives of achieving relation-specific transformations, differentiating the neighbor nodes, and exploiting the intermediate information generated during the KG embedding learning process. (Chapter 6)

Case studies were conducted to show the feasibility and effectiveness of the proposed methodology with extensive experiments including comparative studies with some state-of-the-art methods. Prototype systems were also constructed to illustrate the CPS context. The cognitive intelligence of the developed condition monitoring of manufacturing thing reflects in the achieved self-X capabilities, i.e., condition-based self-awareness, feature-based self-learning, edge cloud-orchestrated self-adaptation, and KG-assisted self-adjustment.

Most of the research work in this thesis has been reported in three journal articles. It is anticipated that the research outcomes could provide enterprises with useful guidance and practical examples for condition monitoring development in a cognitive intelligent paradigm featured with self-X capabilities.