Abstract— This paper describes an approach to design self-developing and self-tuning inferential soft sensors applicable to process industries. The proposal is for a Takagi-Sugeno-fuzzy system framework that has evolving (open structure) architecture, and an on-line (possibly real-time) learning algorithm. The proposed methodology is novel and it addresses the problems of self-development and self-calibration caused by drift in the data patterns due to changes in the operating regimes, catalysts ageing, industrial equipment wearing, contamination etc. The proposed computational technique is data-driven and parameter-free (it only requires a couple of parameters with clear meaning and suggested values). In this paper a case study of four problems of estimation of chemical properties is considered, however, the methodology has a much wider validity. The optimal inputs to the proposed evolving inferential sensor are determined a priori and off-line using a multi-objective genetic-programming-based optimization. Different on-line input selection techniques are under development. The methodology is validated on real data provided by The Dow Chemical Company, USA. I. INTRODUCTION A. Background and state-of-the-art NFERENTIAL sensors also known as soft sensors [1], are applied nowadays extensively in a range of industries, such as processing, chemical, petro-chemical, manufacturing, etc. One of the typical application of soft sensors is for process quality monitoring [2],[3]. The black-box-model-based inferential sensors applied currently [1],[4],[5] has big advantages over the conventional solutions that rely on laboratory tests and manual intervention in terms of overall process automation and costs. The most widespread methods that are used to design inferential sensors are principle component analysis (PCA) for reducing the input dimensions and correlation between raw data readings and partial least squares (PLS) to train the models [1]. Alternative techniques that are used for soft sensors design are neural networks (NN) [6], support vector machines [7], genetic programming [8]. The main problem of inferential sensors based on these techniques is caused by the fact that the real industrial processes are highly non-linear, non-stationary, they have different operating regimes, the environment and the Dr. P Angelov and Mr. X Zhou are with the Intelligent Systems Research Lab, Dept of Communication Systems, InfoLab21, Lancaster University, Lancaster, LA1 4WA, UK; phone +44 (1524) 510391; e-mail: p.angelov@lancaster.ac.uk Dr. A Kordon is with the Data Mining and Modelling Group, Corporate Work Processes & Six Sigma Centre, The Dow Chemical Company, Freeport, TX, USA; phone +1 (979) 238 5149; e-mail: akkordon@dow.com industrial equipment, raw materials and catalysts are changing. This dynamically (often unpredictably) evolving environment leads to pre-trained and designed in off-line mode inferential sensors to have unacceptable drop in their performance and to require periodic and costly re-training, re-calibration and sometimes re-development. In this way, the life-cycle costs of these sensors become comparable or higher than that of laboratory tests. Another significant disadvantage of NN and other ‘black box’ techniques is their lack of interpretability and transparency, which is very important for human operators of these expensive industrial processes who sometimes rely on their experience and intuition. This reality calls for the development of new techniques that are adaptive and able to react to the complex changes in the process such as wearing out or contamination of the equipment, quality alteration of raw materials, etc. Ideally, an intelligent inferential sensor will have an online (possibly in real-time) structural learning ability in response to the fundamental shifts in the process. B. Modes of operation of inferential industrial sensors In most industrial process, the inferential sensors are expected to work in one of the following three modes of operation. For cases with high degree of stationarity and low level of variability of the raw materials, catalysts, environment and the equipment inferential sensors with a pre-trained fixed structure and parameters may be satisfactory. For processes with frequent non-fundamental changes, sensors are required to continuously adapt to these changes online. Such inferential sensors can have a fixed structure but require the ability to automatically re-tune their parameters in response to the changes. These sensors will be called adaptive or self-adaptive (if adaptation is automatic and on-line, but does not concern the structure of the inferential sensor) and self-tuning inferential sensors. Finally, there is also a type of inferential sensors that will be called evolving sensors, eSensors which evolve their structure as well as adapt their parameters. They are especially important and needed when there is a significant shift in the data pattern, which changes the way the original process works. Mathematical models with fixed structures are not able to react with parameters tuning to such situations. In order to automatically detect, learn and react to both fundamental and non-fundamental changes, specific novel techniques and methodologies are needed, which requires the soft sensor to detect shifts and drifts in the process and learn not only the parameters, but also to evolve its own structure in Evolving Fuzzy Inferential Sensors for Process Industry Plamen Angelov, Senior Member IEEE, Arthur Kordon, Member IEEE, and Xiaowei Zhou, Student Member IEEE I 3rd International Workshop on Genetic and Evolving Fuzzy Systems Witten-Bommerholz, Germany, March 2008 978-1-4244-1613-4/08/$25.00 ©2008IEEE 41 Authorized licensed use limited to: IEEE Xplore. Downloaded on October 29, 2008 at 10:17 from IEEE Xplore. Restrictions apply.