My PhD
Title:
Developing Novel Machine Learning Techniques with Human-in-the-Loop Approach to Enable Better Decision Making on Operations Maintenance
External Stakeholder:
Tata Steel
The Research:
Asset monitoring in large scale industries is challenging. Users (human-supervisors or operators) monitor the assets and any failure or emergency can cause them to stress and impact their mental health. Further, this can impact their efficiency. In addition to mental health, it can have catastroph-ic failures or unplanned downtime and reduce quality losses. To address this challenge, the Fourth Industrial Revolution, also called Industry 4.0. is set to change the operation and maintenance of manufacturing processes through better integration of physical and digital systems. Industry 4.0 relies on the pervasive and ubiquitous use of ICT, sensors and data to deliver the next generation of intelligent, co-operating and interconnected manufacturing systems. These systems can collect, process and store a large amount of data related to physical processes, including equipment availability, condition monitoring, events and alarms. The real-time asset monitoring and data analysis can provide valuable information and knowledge about the health status of assets, leading to improved decision-making capabilities that can significantly reduce the cost of maintenance, avoid catastrophic failure or unplanned downtime, reduce quality losses and will further reduce supervisor’s stress and positively impact their health and well-being.
Predictive Maintenance (PdM) is an emerging maintenance strategy that uses predictive tools to determine the optimal strategy as opposed to scheduling maintenance at regular intervals, therefore optimising the utilisation of assets, their availability, and operations. Traditional PdM approaches include methods based on statistics and physical based models. These approaches are typically used for single component systems without considering how several assets may interoperate and interact and they rely on calculation of thresholds to determine the health status of equipment, thorough monitoring signals such as vibration, temperature, and torque. More recently, Machine Learning (ML) has emerged as a key technology in Predictive Maintenance applications due to the ability to discover complex pattern related to wear conditions for single or multiple component systems, using large volume and high dimensional datasets. However, compared to physics-based methods, ML methods do not take into consideration specific domain knowledge or ‘human expertise’ about maintenance operations that has accumulated over the years; hence they lack the ability to generalise when new failure modes occurs or to transfer knowledge/experience from one system component to another. Furthermore, due to the ‘black box’ nature of ML systems it is difficult to interpret or use the results of the predictions to effectively support decision making about the optimal maintenance strategy.
In large scale and complex manufacturing operations such as steel making the sheer volume of streaming data collected by monitoring multiple assets pose further challenges in data processing and subsequent analysis for predictive maintenance. In particular, the deployment of ML approaches for PdM in steel production facilities is hindered by the lack of scalable and robust Machine Learning methods as well as the difficulty in acquiring labelled datasets for training. ML models typically take a lot of time and computational resources for training and performance optimisation, hence making it unpractical for real world deployments.