Development and deployment of novel soft sensors in the process industry

The described impact has resulted from collaborative research between the Smart Technology Research Centre from Bournemouth University (BU) and the Evonik Industries, one of the world leaders in the specialty chemistry field. The efforts focused on development and deployment of a next generation of adaptive soft sensors for on-line prediction, monitoring and control in the process industry.

Details of the impact:

The fundamental research in the area of multi-component, multi-level, adaptive prediction systems was one of the focal points of Prof. Gabrys’ research between 1997 and 2006. Our research outputs from that period attracted interest from commercial organisations and resulted in financial support for a number of collaborative research and consultancy projects that followed.

One such collaboration which started in 2006 has been with Evonik Industries and the development and deployment of adaptive soft sensors on complex chemical plants is the subject of this impact case.Chemical_Plants

Due to increasing costs of labour, energy and resources in Europe, operational excellence is the key for competitiveness of Europe’s process industry. Operational excellence leads to higher cost efficiency, better exploitation of plant capacity, loss reduction as well as achieving the compliance with environmental and safety regulations.

Driven by the need for highly adaptive predictive systems and as part of looking for alternative solutions departing from the expensive and inefficient practices for development and maintenance of soft sensors, Evonik and Prof. Gabrys’ group initialised an European Task Force on Nature-Inspired self-healing Soft Sensors for Process Industry (NiSSPI) within the NiSIS project. This task force gathered an inter-disciplinary cross-section of control and telecommunication engineers, computer scientists, biologists and chemists from around Europe. The collaboration between BU and Evonik initiated within NiSIS was one of the driving forces behind a successful pilot project between BU and Evonik [2006-2009] which resulted in defining a novel, open, flexible architecture for development of highly adaptable predictive systems and several successful case studies.

Based on these results, a new substantially extended research collaboration with an additional partner – REC Global from Poland – has been initiated within the INFER project [2010-2014]. The invitation of a skilled software engineering company has been motivated by the need to fill in the gap between the research software, which mainly focuses on proofs of concept, and production-level software, which puts more focus on robustness, scalability and an efficient user interface.

In order to realise the impact by both the development of the software tools and deployment of the proposed adaptive soft sensors for a number of different chemical processes, there has been extensive staff exchange within INFER through a broad secondment scheme which saw 16 industrial researchers getting the opportunity to gather knowledge in academia, 10 academic researchers absorbing knowledge in industry and 2 new experienced researchers directly recruited to the project.

With the introduced adaptive soft sensors, and as illustrated in the proof of concept studies, product quality monitoring can be performed more easily and reliably in a dynamically optimal manner, so that product sampling and expensive laboratory based chemical analysis can be reduced while probing intervals for control purposes can be extended.Control_room

Previously developed adaptive soft sensors have also been directly applied to Evonik’s processes by Dr Kadlec who after working on a co-funded by Evonik project, receiving a PhD from BU in 2009 and a short period of employment as a Lecturer at BU was employed by Evonik Industries in Nov 2010 where he made a significant impact by direct deployment of data analytics tools leading to substantial annual savings per chemical plant. Dr Kadlec, and his colleagues at Evonik Industries, were also playing a very important role in INFER by being responsible for the application of the developed methods and tools to Evonik processes between 2010 and 2014.

Key projects:

  1. Coordination Action project on Nature-inspired Smart Information Systems (NiSIS), 2005-08, funded by EC FP6 IST Future and Emerging Technologies (€1,000,000), involving 56 European Centres (including BT and Evonik). The Nature-inspired Data Technologies (NiDT) Focus Group was co-led by Prof. Gabrys. He was also a member of the Nature-inspired Self-healing Soft Sensors for Process Industry (NiSSPI) Task Force led by Evonik. It focused on raising awareness, collecting ideas, architectures and concepts for constructing and implementing self-understanding and self-healing soft sensors for the process industry. The activities of NiSSPI culminated in Ruta (former Gabrys’ PhD and a senior visiting research fellow from BT at Computational Intelligence Research Group (CIRG) in BU at the time of the competition) winning a predictive modelling competition organized by Evonik. The winning solution stemmed from the previous joint research of Ruta and Gabrys, and was later extended and published by BU team in [3]. Joint participation in NiSIS, co-involvement in NiDT and NiSSPI as well as winning the competition which used the solutions promoted within CIRG [1, 2, 3], formed a basis of a long-term collaboration between BU and Evonik within further projects listed below.
  2. Self-adapting and Monitoring Soft Sensors for Process Industry, 2006-2009, PhD project co-funded by Evonik, proposed and supervised by Gabrys and completed by Kadlec. The project investigated the issues of predictive modelling approaches used in the process industry [4 , 5] and addressed the identified issues by proposing a conceptual hierarchical architecture for development of robust and adaptive soft sensors [6, 7]. A successful pilot study using real-world data provided by Evonik has validated the proposed approach [8,9]. Kadlec was subsequently employed by Evonik in 2010.
  3. Computational Intelligence Platform for Evolving and Robust Predictive Systems (INFER), 2010-14, funded by EU FP7 People Programme, Marie Curie Industry Academia Partnerships and Pathways (€1,555,380, BU: €890,421, 2010-2014, Ref: FP7-PEOPLE-2009-IAPP-251617). The project was coordinated by Gabrys and involved in total 38 academic and industrial researchers. The project focused on pervasively adaptive software systems for the development of a modular computational INtelligence software platform For Evolving and Robust predictive systems applicable in various commercial settings and industries, with a focus on a process industry case studies which followed from the successful pilot project. Further details on INFER can be found at the project website and in the final project summary. A small selection of papers particularly relevant to the described case study can be found below while the full list of publications resulting from the INFER project can be found here.

Key publications:

  1. Ruta, D. and Gabrys, B. 2005. Classifier selection for majority voting. Information Fusion, 6 (1), pp. 63-81.

    Novel multistage selection-fusion (MSF) model, an example of multilevel combination structures, introduced within our Computational Intelligence Research Group and offering improved generalisation performance was proposed for the first time. This concept was later further developed and applied in various application areas together with various industrial/business partners.

  2. Ruta, D. and Gabrys, B. 2007. Neural network ensembles for time series prediction. Proceedings of the International Joint Conference on Neural Networks (IJCNN 2007), pp. 1204-1209.
  3. Ruta, D., Gabrys, B. and Lemke, C., 2011. A Generic Multilevel Architecture for Time Series Prediction. IEEE Transactions on Knowledge and Data Engineering, 23(3), pp. 350-359. DOI: 10.1109/TKDE.2010.137.

    International forecasting competition winning approach following a class of promoted and developed in our Smart Technology Research Centre multi-level, multi-component prediction frameworks and approaches based on diversified ensembles of predictors. This work won the NiSIS 2006 competition, which was set up by Evonik Degussa and concerned the prediction of catalyst activity, and led to a successful co-funded, collaborative pilot project between Evonik and BU (£85k, 2006-2009) and later funding of a large EU project (INFER €1,555,380, BU: €890,421, 2010-2014, Ref: FP7-PEOPLE-2009-IAPP-251617).

  4. Kadlec, P., Gabrys, B. and Strandt, S., 2009. Data-driven soft sensors in the process industry. Computers & Chemical Engineering, 33 (4), pp. 795-814.
  5. Kadlec, P., R. Grbic and B. Gabrys, Review of Adaptation Mechanisms for Data-driven Soft Sensing. Computers and Chemical Engineering. 35 (1), pp.1-24, Jan 2011.

    Agenda setting review articles which since their publication in May 2009 and Jan 2011 respectively have attracted significant attention and are very often cited by other researchers in the fields of intelligent/ inferential/ soft sensors and predictive modelling in the chemical engineering area. According to the CaCE journal webpage (http://www.journals.elsevier.com/computers-and-chemical-engineering/most-cited-articles/) they have been the most cited articles in the CaCE journal.

    Apart from the comprehensive systematic reviews and description of industry practices based on original research within a process industry and carried out in collaboration with practitioners, the identified in the papers open issues (and proposed solutions) concerned with automating the process of soft sensors development and maintenance led to a number of further grant applications including the EU funded INFER project within the strategic Industry-Academia Partnerships and Pathways (IAPP) scheme. The research undertaken as part of the collecting of the material for this paper (and a series of papers on robust adaptive predictive systems that preceded and followed after it) has also been disseminated by invited keynote and plenary talks delivered by Gabrys at the HAIS 2008 (Spain), ICONIP 2008 (New Zealand), ICIAP 2009 (Italy), CISIS 2009 (Spain), INFER-APS 2011 (UK) and ICMMI’2013 (Poland) conferences.

    A selection of publications resulting from the pilot and INFER projects

  6. Gabrys, B., Self-adapting architecture for building powerful predictive models. Invited contribution to ICONIP2008 conference. Nov 2008.
  7. Kadlec, P. and Gabrys, B., 2009. Architecture for development of adaptive on-line prediction models. Memetic Computing, 1 (4), pp. 241-269.
  8. Kadlec, P. and B. Gabrys, 2010. Adaptive on-line prediction soft sensing without historical data. Proceedings of the 2010 International Joint Conference on Neural Networks (IJCNN), pp. 1-8 Barcelona, Spain.
  9. Kadlec, P. and Gabrys, B., 2011. Local learning‐based adaptive soft sensor for catalyst activation prediction. AIChE Journal, 57 (5), pp. 1288-1301.
  10. Gabrys, B, Robust Adaptive Predictive Modelling and Data Deluge, Proceedings of the International Conference on Man-Machine Interactions (ICMMI’2013), The Beskids, Poland, Oct 2013.
  11. Zliobaite, I. and B. Gabrys, Adaptive Pre-processing for Streaming Data, IEEE Transactions on Knowledge and Data Engineering, 26(2), pp. 309-321, Feb 2014.
  12. Budka, M., Eastwood, M., Gabrys, B., Kadlec, P., Salvador, M.M., Schwan, S., Tsakonas, A., and Zliobaite, I., From Sensor Readings to Predictions: on the Process of Developing Practical Soft Sensors, Proceedings of the 13th International Symposium on Intelligent Data Analysis (IDA’2014), pp 49-60, Lueven, Belgium, Oct. 2014.
  13. Salvador, M, B. Gabrys and I. Zliobaite, Online detection of shutdown periods in chemical plants: a case study, Procedia Computer Science, 35, pp. 580-588, Dec. 2014.