Design and evaluation of data driven prognostic models for component health management

Abstract

Throughout industry and academia, big data analytics are being used to discover features within components and systems, in order to improve their design, monitoring and management. The industrialisation of data analytics for health management of components is highly complex, due to challenges associated with data availability, data quality, variances in components and applications, as well as the operational demands e.g. computational eciency. This thesis identifies the limitations of vibrational data analysis associated with bearings for health management. There is a need for a comprehensive multi-sensor vibration dataset for training fault classification models in order to improve the levels of intelligence in embedded systems for Prognostics & Health Management (PHM) and thereby, automation in health management. Through an extensive and quantitative review of the literature it was found that a large quantity of the research makes use of a small number of open datasets or self-generated data containing few instances of faults. Therefore, a lab-based custom designed experiment using Accelerated Lifecycle Testing (ALT) was used to create a highly extensive dataset using multiple accelerometers to augment the data quantity for 172 instances with 15 separate fault classes covering 240hrs of testing. Fourier and Wavelet transforms were compared alongside raw-data based methods based on windowing and also a novel multi-channel analogy method for multi-sensor data. Their performance on this dataset was compared through the use of comparative multi-class performance metrics. The findings reveal the best performing pre-processing steps in the performance of Fourier transformed data and the novel multi-channel data which provide a similarly high level of performance. Though these networks provide di↵erent characteristics which a↵ects their suitability for use within embedded systems and on a cost/benefit ratio the Fourier transformed data provides best performance per unit data by reducing stored data volume by a factor of 2. A data pipeline was created for vibrational analysis, evaluating the performance of Convolutional Neural Networks (CNNs) on data inputs with di↵erent sample rates through the use of downsampling which discovered a non-linear relationship between performance and sensor sample rate. The findings within this ALT data base and data pipeline demonstrated that acceptable performances can still be achieved with reduced sample rates and the relationship is mapped as a design tool for the optimisation of embedded systems where computational requirements are limited and benefit from minimising the collection of redundant data. The findings inform the design and optimisation aspects of applying fault prediction in embedded systems for the purpose of providing accurate decision support to broader PHM methodologies aimed at understanding and improving the reliability of systems.