The nature of failure data sets poses particular challenges to modellers. Failures, particularly of critical equipment, are rare. Lots of equipment are replaced in whole or in part before end of life. As a result failure data-sets are unbalanced and sparse. Failures are seldom labelled accurately, and there is often no ground truth for validation.

Condition monitoring data, when available, is often collected at different time intervals. Poor quality data results in greater model complexity that at best muddies inference, and at worst misleads inference and produces persistent prediction bias. These contextual issues, if not dealt with rigorously in model selection and validation practice, leads to poor model performance and a loss of trust by decision makers.  To manage these risks, this theme is explicitly cross-disciplinary combining the Bayesian statistics, engineering, nonlinear system identification, machine learning and deep learning.

Initial project areas for Theme 2 include:

Project 1: Dynamic modelling and nonlinear time series

Early signs of asset failure result in nonlinear changes in system dynamics of complex systems. That is, many assets consist of complex interacting part, and are a non-autonomous component in a larger system. Failure of each component will result in subtle changes in the dynamical behaviour of that system and these changes are best detected with a suite of new nonlinear signal processing tools. The ability to detect these would provide an earlier indication than traditional linear condition monitoring techniques. The aim of this project is to develop and to apply a range of nonlinear time series analysis methodologies – including state-space based dynamical reconstruction, frequency domain characterization, and novel machine learning paradigms – to provide a new and improved indicator of asset failure.  Condition monitoring data is collected across a range of assets across all our industry partners – this may include chemical reaction and mixing processes, pump vibration time series data, multi-modal condition monitoring of heavy equipment or process throughput control. In all cases this time varying condition data provides a proxy for the system health, which is imperfectly understood. The aim of this project is to develop the tools to better characterize and understand conditions of these systems across operations. The project will produce real-time diagnostic algorithms which can be deploy and embedded within current operations.

Project 2: Bayesian models for failure prediction and remaining useful life estimation

This project aims to deliver improved predictability of failure (with uncertainty estimates) for individual assets (rather than a population of assets) from longitudinal data. Remaining useful life (RUL) estimation provides a probabilistic maximum likelihood estimate of the expected time to failure. This is naturally a stochastic quantity. The aim of this project is to apply Bayesian methodologies in conjunction with other data driven modelling paradigms to optimally estimate expected failure. This will include an estimate not only of the RUL, but also the uncertainty of this estimate. Combined, these quantities can then be drawn upon for optimal maintenance scheduling and planning and for empirical expected-value based planning of asset replacement and retirement.

All three industrial partners are currently monitoring machine and process operations of relevance to these techniques. In all cases, whenever assets are deployed within a larger complex system, planning of RUL and estimates of prediction uncertainty will allow for improved operation. The outcome of this project will be algorithms to provide RUL estimates across elements of the processing chain. Integrating these will result in improved maintenance scheduling and planned obsolescence.

Project 3: Fault diagnosis and prediction through advanced spectral analysis techniques

In addition to time series data (the subject of the preceding two projects, and probably a good justification to seek progress on them first), asset health is often monitored with multi-modal data. By combining that data from different modalities (video, spectral, time series) the objective of this project is to improve fault diagnosis. The primary research question here is two levelled – first, how is that data best understood individually; and, second, how can one best integrate data from different modalities for optimal prediction?

Additional information concerning corrosion, contamination, degradation, congestion and failure can be obtained from video and image data (in addition to time series and audio). The combination of 2D (image) and 3D (video) data with techniques honed for 1D (time series/audio) data requires both novel mathematics and new computation algorithms. The objective of this project is to develop the algorithmic techniques to allow for integration of multi-channel multi-modal and multi-dimensional data from multiple sources for better predictions. Applications will include proves monitoring and overall system health across a range of industrial processes in mining, oil and gas and processing.

Project 4: Optimise maintenance management work flow

Equipment maintenance typically involves many different individuals with different duties within an organisation interacting with the equipment. Each transaction (work request, routine maintenance, replacement, service, etc.) is recorded in a Maintenance Work Management System (typically SAP or equivalent systems) and provides a rich quantitative data set of records of interaction and interdependence among individuals and equipment. Whereas the other projects in Theme 2 view the physical assets as a complex system, here we treat the maintenance workforce as a virtual complex system of human capital. The topology of this system, and how it interacts with the maintenance objects defines a set of complicated interdependencies which lead to redundancy (and hence reliability) or leanness (and hence fragility) in the maintenance system. The object of this project is to describe and optimize that system, identify bottlenecks and inefficiencies in across the entire maintenance process.