- Opens: Wednesday 6 April 2022
- Number of places: 1
- Duration: 3 years
- Funding: Home fee, Stipend
OverviewThis PhD seeks to make a contribution to fundamental research in the field of cognitive intelligence, and specifically in addressing the challenges of representing human decision making in a repeatable and reusable manner in a digital form to support co-generation of human-driven, artificial intelligence supported decisions within industrial contexts.
Minimum 2:1 degree in Engineering, Computer Science or Physics. Strong programming skills desirable.
There is substantial research in the field of developing and applying machine learning approaches to specific challenges, often with the requirement for large, well-curated sets of data. However, these approaches often require significant computational skills to develop, and the results generated are considered “black box” where there is not a clear and auditable set of reasoning steps to explain the conclusions. In many industrial applications, the end users while tending to be technical experts in their field, don’t have the associated computer-related backgrounds to fully understand or accept the outcomes. Furthermore, in heavily regulated environments, there is the additional need for regulatory acceptance of such techniques before they can be deployed, which introduces an additional role which is required to understand the outcomes from automated, or semi-automated decision making. This is particularly true of the nuclear industry, which is keen to benefit from recent advances in data science in this field, but requires to address this barrier of explicability for acceptance in this rightly conservative domain.
Knowledge graphs are an approach which has generated significant interest in many applications, including search engines and image recognition systems, which aim to capture, derive and represent the underlying relationships between concepts from large volumes of data, such as text or images, particularly where these relationships are not necessarily fully understood or are ill-defined. However, in an engineering context, data which needs to be understood and reasoned about is often a digital mirror of an underlying physical system or process, while complex, has a structured form which can be reasoned about, by a domain expert. However, there exists a gap whereby this domain expertise can be readily captured, represented and integrated with knowledge and evidence derived from data driven techniques.
The aim of this PhD is to demonstrate the unification of both a bottom up, data-driven knowledge graph representation of assessment of condition monitoring data with a top down, domain expert driven representation of the same task, using a formal knowledge modelling methodology, such as commonKADS. The project will also demonstrate the use of the combined knowledge to provide a solution to the engineering task, along with a transparent explanation of the decisions made, the supporting evidence derived from the raw data and associated confidence values for dealing with uncertainty. The project will be contextualized through industry-provided case study data and access to the time of engineers and domain specialists to derive and validated the domain specialist driven component.