One of the key research areas for the asset management group has been the exploitation of machine learning and artificial intelligence in improving the operational effectiveness and reliability of industrial systems. Here are highlighted some of our major developments.

1. Distributed Collaborative Prognosis. This brings together advances in Internet of Things (IoT), distributed intelligence, and machine learning to create an ecosystem similar to social networks where machines communicate and learn from each other to improve their ability to predict failures. Our work in this area has shown that collaborative prognosis results in a reduction in the number of failure cycles that each machine takes to achieve good failure prediction. Furthermore, it allows machines to automatically adapt to changing environment (and hence changing failure patterns) by restricting their learning from machines that behave in the most similar manner. We have also developed several types of architectures for implementing this concept and performed a cost comparison to identify the situations where each architecture has potential value.
2. Federated Learning in Prognosis. One of the major drawbacks of collaborative prognosis is that it requires machines to share their sensor/performance data with each other. Realising that such an approach might have commercial and privacy sensitivities associated with it, we are exploring the use of federated learning approaches where machines share models instead of datasets to improve their prognostics ability. Initial results show promise, and we are hoping that this will provide a realistic pathway to implementation for prognosis and predictive maintenance.
3. Generative learning for Prognosis. Most complex industrial systems are highly reliable and failures are rare. This presents a fundamental challenge to data-centric approaches for failure prediction because amount of data pertaining to failures is often very low compared to that of normal operation. In order to improve the performance of failure prediction in cases where such a ‘data imbalance’ exists, we are using conditional generative adversarial networks to generate realistic failure data based on information/knowledge regarding the nature of failures. Our work so far has shown that this technique greatly improves prognostics performance – a testament to this is the best paper award we received for this work presented at the IEEE International Conference on Prognostics and Health Management.