Course summary

A funded PhD studentship is available within the Autonomous and CyberPhysical Systems Centre in the School of Aerospace, Transport and manufacturing (SATM), Cranfield University, Bedfordshire, UK. As aerospace platforms go through their service life, gradual performance degradations and unwarranted system failures can occur. There is certain physical information known a priori in such aerospace platform operations. The main research hypothesis to be tested in this research is that it should be possible to significantly improve the performance of extreme learning and assure safe and reliable maintenance operation by integrating this prior knowledge into the learning mechanism. The integrating should enable to guarantee certain properties of the learned functions, while keep leveraging the strength of the data-driven modelling. Most of, if not all, the traditional statistical methods are not suitable for big data due to their certain characteristics: heterogeneity, statistical biases, noise accumulations, spurious correlation, and incidental endogeneity. Therefore, big data demands new statistical thinking and methods. As data size increases, each feature and parameter also becomes highly correlated. Then, their relations get highly complicated too and hidden patterns of big data might not be possible to be captured by traditional modelling approaches. This implies that mathematical modelling of such data is infeasible. The data-driven modelling approach could resolve this issue and we could use obtain data-driven models using machine learning algorithms such as artificial neural networks, reinforcement learning, and deep learning. A typical caveat of data-driven modelling using learning algorithms as Extreme Learning Machine (ELM) is that training data should cover the entire domain of process parameters to achieve accurate generalization of the trained model to new process configurations. In practice, this might not be possible, that is the sample data could cover only some space, not entire space, of process parameters. Integrating prior knowledge into the learning could enable accurate generalization of the data-driven model even when the space of system parameters is only sampled sparsely. Consequently, it will improve the performance of the learning. Integration of the prior knowledge of the system into the learning procedure will be quite challenging since the key enabler of its very powers is the universal approximation capabilities. Sampled data are generally noisy, outliers occur, and there always exist a risk of overfitting corrupted data. Therefore, the learned function may violate a constraint that is present in the ideal function, from which the training data sampled. This PhD study will address this research challenge.