Developing Novel Machine Learning Techniques with Human-in-the-Loop Approach to Enable Better Decision Making on Operations Maintenance

点击翻译按钮获取中文摘要

The project will support the development of new human-centric approaches to Predictive Maintenance in the steel industry. By enabling synergy between operators (and their knowledge) and data models for assets maintenance, it will be possible to optimize the maintenance schedule, resulting in cost savings and increased safety of operations. Workers will also benefit because it will enable them to plan maintenance at the right time avoiding the need to deal with unforeseen circumstances, hence improving the wellbeing of workers who areexposed to a very challenging and complex environment.The board aims and objectives are:1) Work with onsite engineers who have the domain knowledge to understand the origin of data and how it relates to the physical realities of the process.2) Use novel data analysis techniques with artificial intelligence and machine learning to create a digital twin of the asset.3) Create models which are scalable across assets within the Azure environment.4) Create event-based outputs from models into existing dashboards for use by maintenance teams.5) Create a guidance on the standard requirements for input data formats and code language(s) (Python, C++, C# etc) to be used.In this research project we will address the above challenges and study novel ML tools and workflows with 'humans (supervisors) in the loop' that integrate data driven approaches with knowledge modelling to develop robust, transferrable, adaptable and usable ML models for in-line and real time predictive maintenance. The research will follow three important strands:An inspection data analysis environment that will present a realistic display of inspection data and will be used as a human (supervisors)-in-the-loop approach to learn about the domain knowledge in terms of predicting failures. Supervisors' actions to predict failure will help with labelling the unlabelled data which will help develop ML models. Consequently, supervisors' feedback will help generate a transferrable, adaptable and usable ML model. Investigate and apply Transfer Learning approaches to improve scalability and adaptability of ML models across a manufacturing site (reducing time and complexity during the training process).Investigate, develop and study hybrid human-centric PdM approaches that integrate semantically enriched data with data-driven models to learn appropriate corrective actions associated to failures and drive optimal decision-making strategies. The research project will use datasets from the centralised asset management platform (AMDC) at Tata Steel, focusing on specific use cases. The findings of the research will create direct benefit for the industrial sponsor as it will enable Tata Steel. to reduce the overall cost of maintenance and reduce occurrences of failures, leading to increased sustainability and improved worker safety and wellbeing in the steel works. The research methodology will employ human-centric approaches and user involvement to drive the development of novel ML workflows in PdM to enhance human decision making in complex industrial environment. We expect that some of the research findings and methodologieswill be of general application to PdM and hence will bring societal and economic benefits through improved, safer and more sustainable industrial processes.The student will work closely with industrial users (onsite engineers and managers) who have the domain knowledge to understand the origin of data and how it relates to the physical processes. Stakeholders (shop floor workers, managers and maintenance operators) will be involved in the research design and development of solutions through formal workshops and day to day interactions at the plant. The users will be involved in evaluation of solutions in an iterative way to gain continuous feedback that will lead to further improvements.