Data-driven modelling and monitoring of industrial processes with applications in the nuclear waste processing industry
Lead Research Organisation:
Newcastle University
Department Name: Sch of Engineering
Abstract
In the area of process monitoring and control, building accurate and reliable models is at its foundation. These models can be used in control schemes or in the estimation of key process variables. Monitoring process variables is of importance on industrial plants because of the need to ensure product quality and to monitor for unwanted process variations. For many companies, falling outside of certain process limits can not only lead to financial losses, but also breaching rules and regulations. In particular, the emphasis on the reduction of environmental pollution in more recent years has led to stricter laws that companies have to adhere to.
In many processes, monitoring a wide range of process variables using hardware sensors can be of great expense and in some cases, on-line measuring may not be feasible. Off-line sampling is an alternative, however, this typically produces a lower rate of samples and significant delays. In addition, faults can arise in hardware sensors and planned maintenance leaves sensors out of order for a time.
The alternative arises in "soft sensors", which predict difficult to measure quality variables from easy to measure process variables using computation. The advantages of soft sensors are that they can be much cheaper than hardware sensors, provide on-line predictions with a lower delay than hardware sensors, and can be used in conjunction with hardware sensors to aid in fault detection and take over during hardware maintenance.
The two primary ways to produce soft sensors are through using mechanistic models and data-driven models. Mechanistic models utilise first principle mathematical equations to describe the process. Many modern industrial processes are highly complex and the production of a mechanistic model can be very time consuming and require many assumptions that can lead to a sub-optimal model. Data-driven soft sensors make use of process data in their production. There are many techniques that can be used to create a data-driven soft sensor and they are typically based around multivariate statistics and computation intelligence methods.
The area of data-driven soft sensors is constantly evolving due to the development of new techniques and advances in computational hardware that enable more complex models to be built without excessive computational effort. The need to create more accurate and robust models is of great importance to the process industry to reduce operational costs through enhanced process control and monitoring.
The basis of this project is to innovate and develop non-linear multivariate data analysis and data-driven modelling techniques for industrial processes, in particular the nuclear waste vitrification process at Sellafield Ltd. The challenge of long term storage for high level nuclear waste is of great importance to ensure there is no radioactive leakage into the environment. The application of cutting edge, data-driven modelling techniques can lead to improvements in long term storage of the waste, as well as the vitrification process itself. Methods to be investigated will include non-linear multivariate statistics, computational intelligence and hybrid techniques. In addition, methods for improving the reliability and generalisation of these data-driven models will be investigated. Development of accurate and robust data-driven models will be at the heart of the project through the collaboration with Sellafield Ltd, since this will provide application to the nuclear waste processing industry.
The project is a studentship that is undertaken between Newcastle University and Sellafield Ltd under the EPSRC Industrial CASE award scheme. The studentship is funded by the EPSRC and Sellafield Ltd for a period of four years.
In many processes, monitoring a wide range of process variables using hardware sensors can be of great expense and in some cases, on-line measuring may not be feasible. Off-line sampling is an alternative, however, this typically produces a lower rate of samples and significant delays. In addition, faults can arise in hardware sensors and planned maintenance leaves sensors out of order for a time.
The alternative arises in "soft sensors", which predict difficult to measure quality variables from easy to measure process variables using computation. The advantages of soft sensors are that they can be much cheaper than hardware sensors, provide on-line predictions with a lower delay than hardware sensors, and can be used in conjunction with hardware sensors to aid in fault detection and take over during hardware maintenance.
The two primary ways to produce soft sensors are through using mechanistic models and data-driven models. Mechanistic models utilise first principle mathematical equations to describe the process. Many modern industrial processes are highly complex and the production of a mechanistic model can be very time consuming and require many assumptions that can lead to a sub-optimal model. Data-driven soft sensors make use of process data in their production. There are many techniques that can be used to create a data-driven soft sensor and they are typically based around multivariate statistics and computation intelligence methods.
The area of data-driven soft sensors is constantly evolving due to the development of new techniques and advances in computational hardware that enable more complex models to be built without excessive computational effort. The need to create more accurate and robust models is of great importance to the process industry to reduce operational costs through enhanced process control and monitoring.
The basis of this project is to innovate and develop non-linear multivariate data analysis and data-driven modelling techniques for industrial processes, in particular the nuclear waste vitrification process at Sellafield Ltd. The challenge of long term storage for high level nuclear waste is of great importance to ensure there is no radioactive leakage into the environment. The application of cutting edge, data-driven modelling techniques can lead to improvements in long term storage of the waste, as well as the vitrification process itself. Methods to be investigated will include non-linear multivariate statistics, computational intelligence and hybrid techniques. In addition, methods for improving the reliability and generalisation of these data-driven models will be investigated. Development of accurate and robust data-driven models will be at the heart of the project through the collaboration with Sellafield Ltd, since this will provide application to the nuclear waste processing industry.
The project is a studentship that is undertaken between Newcastle University and Sellafield Ltd under the EPSRC Industrial CASE award scheme. The studentship is funded by the EPSRC and Sellafield Ltd for a period of four years.
Studentship Projects
| Project Reference | Relationship | Related To | Start | End | Student Name |
|---|---|---|---|---|---|
| EP/N509528/1 | 01/10/2016 | 31/03/2022 | |||
| 1948800 | Studentship | EP/N509528/1 | 01/10/2017 | 30/09/2021 | Jeremiah Corrigan |
| Description | A Journal publication on the development of a data-driven modelling technique to enhance the performance of soft sensors was published. Soft sensors are predictive models used in the process industry where a model is used to predict a difficult to measure process variable from other easier to measure process variables. The technique developed incorporated dynamic slow feature analysis with neural networks. The incorporation of slow feature analysis as a statistical method prior to a neural network model lead to improved model performance, when applied to an industrial process benchmark and numerical example, when compared with other similar techniques. This was novel work in the data-driven modelling field. Additionally, this work was expanded to improve the technique and this has been submitted for publication. The slow feature analysis technique was enhanced for nonlinear applications by incorporating the kernel trick. This lead to improved model performance in nonlinear processes. |
| Exploitation Route | The use of kernel/linear slow feature analysis for process modelling can be further investigated to try and improve model performance. Additionally, the developed techniques can be applied to novel industrial applications. |
| Sectors | Chemicals,Digital/Communication/Information Technologies (including Software),Pharmaceuticals and Medical Biotechnology |