Intelligent Prognostics and Health Management of Modular Systems

Description

To handle the increasing complexity of systems, modular design principles are usually followed to achieve the benefits of scalability, tractability, and hierarchical architectures, in industries such as manufacturing, aerospace, etc. On the other hand, the combination,interface and interactions of modules will unavoidably lead to additional system performance variability and data dependence in system testing, resulting in inconsistency and inaccuracy in system reliability assessment, and studies on theseimportant problems are rare and inadequate for modular systems. This has led to difficulties in effectively implementing Prognostics and Health Management (PHM) for such systems although it is important in industries for preventing system failures andrescuing the operating costs.This project focuses on the PHM of modular systems where various types of information, including structural and data at module or system level, are available. While there have been many research efforts on system reliability modelling and analysis, as well as datadriven methods for general purposes, our project is innovative in a way that it focuses on the unique features of modular system data and information and addresses the system PHM issues through an integrative data-driven and system modelling approach.In this project, we will first develop methods for modelling data and structural dependency in modular systems. To address the issues with different levels of dependent data, a model segmentation approach is proposed. By properly reconfiguring the standardgraph model, an arc reversal-based method will be developed to allow general causal inference. Following that, we will utilize the module-level data and related information about module dependence to perform both module-level and system-level prognostics byenhancing appropriate deep-learning methods. A novel adaptive residual convolutional neural network is designed to enhance the anti-noise performance of module-level prognostics models. Then, an improved deep domain adaptation network is designed toimprove the model's generality and efficiency. A new graph convolutional network will be investigated for system-level analysis to develop a practical prognostic approach for utilizing module-level data and module-dependence information simultaneously.Several industry partners have been consulted about the problems and methodologies to be investigated in this project. Our concepts are found widespread acceptance among our industrial partners, and they are enthusiastic about participating in the testing andevaluation phase by providing the research team with real data from their companies. Our methods shall also be refined and adjusted for domain-specific applications, while recommendations for general applications will be provided.