Data-driven modelling and monitoring of industrial processes with applications in nuclear waste vitrification

Abstract

Process models are critical for process monitoring, control, and optimisation. With the increasing amount of process data and advancements in computational hardware, data-driven models are a good alternative to mechanistic models, which often have inaccuracies or are too costly to develop. One problem with data-driven models is the difficulty in ensuring that the models perform well on new data and produce accurate predictions in complex situations, which are frequently encountered in the process industry. Within this context, part of this thesis explores developing better data-driven models through using a latent variable technique, known as slow feature analysis, as a pre-processing step to regression. Slow feature analysis extracts slow varying features that contain underlying trends in the data, which can improve model performance through providing more meaningful information to regression, reducing noise, and reducing dimensionality. Firstly, the effectiveness of combining linear slow feature analysis with a neural network is demonstrated on two industrial case studies of soft sensor development and is compared with conventional techniques, such as neural networks and integration of principal component analysis with a neural network. It is shown that integration of slow feature analysis with neural networks can significantly improve model performance. However, linear slow feature analysis can fail to extract the driving forces behind data in nonlinear situations such as batch processes. Therefore, using kernel slow feature analysis with a neural network is proposed to further enhance process model performance. A numerical example was used to demonstrate the effective extraction of driving forces in a nonlinear case where linear slow feature analysis cannot. Model generalisation performance was improved using the proposed method on both this numerical example, and an industrial penicillin process case study. Dealing with radioactive nuclear waste is an important obstacle in nuclear energy. Sellafield Ltd have a nuclear waste vitrification plant which converts high-level nuclear waste into a more stable, lower volume glass form, which is more appropriate for long term storage in sealed containers. This thesis presents three applications of data-driven modelling to this nuclear waste vitrification process. A predictive model of the pour rate of processed nuclear waste into containers, an early detection system for blockages in the dust scrubber, and a model of the long-term chemical durability of the stored glass waste. These applications use the previously developed slow feature analysis methods, as well as other data-driven techniques such as extreme learning machine and bootstrap aggregation, for enhancing the model performance.