An Identification Tool for Uncertainty Estimation of Process Measurements Riku-Pekka Nikula, Esko Juuso, Kauko Leiviskä University of Oulu, Control Engineering Laboratory, forename.surname@oulu.fi KEYWORDS identification, modelling, sensor validation, soft sensor ABSTRACT This paper presents a tool, which is used in the identification of models for the estimation of the uncertainty of measurements in an on-line process monitoring application. The tool has three separate parts. The first part is used to identify simple linear regression models and limits for the measurements. The second part is used in the generation of inequality equations between the variables. The third part is used to identify soft sensors which have the basis in physical relationships between the variables. To estimate the overall uncertainty of a measurement, the models from the tool need to be used in conjunction with each other and together with other methods. The identification tool is built in the Matlab platform and the necessary historical data, which is needed for identification, is fetched from a database using SQL. The performance of the tool is demonstrated with an industrial data set from a combined heat and power plant. The main benefit of the tool is the fast implementation of the identified models into the database and for the monitoring purposes. The use of the tool presumably simplifies and speeds up the implementation of the models into the on-line application compared with manual implementations without the tool. 1 INTRODUCTION Industrial plants have numerous measurements, which provide data that is used for operational, reporting and analysis purposes. Critical parts such as control loops and performance monitoring continuously use real time information from the process. To maintain a high level of economic or environmental efficiency and safety, it is essential that this information is correct, reliable and up-to-date. The measurement uncertainty is caused by internal and external reasons. Internal uncertainty originates from the measurement method, parts and materials, cabling and mounting of the sensing device, and I/O-card. Therefore, a single measurement has several possible sources of errors. The increased wear of parts, the corrosion of equipment, the fouling of sensors, calibration difficulties and other challenges that originate from the harsh industrial environments are examples of external reasons for uncertainty. The uncertainty can be decreased by the introduction of high quality sensing systems with regular maintenance and through advanced monitoring of the measurements themselves using computational methods. This study concentrates on the latter one. Sensor validation, consisting of the tasks of sensor failure detection, isolation, and accommodation, has its basis in the redundancy of several sensors or in a sensor’s own health information /3/. The redundancy approaches can be coarsely divided into physical and analytical approaches. Physical redundancy is often used in safety critical areas such as aviation and it is based on the utilization of redundant sensors measuring the same parameter of the system /3/. This kind of redundancy is best supplied as a part of the original system installation, because the retrofit work could require system downtime. Analytical redundancy approaches are model-based validation methods. The models are equations derived from the first principles /2/, /20/ or empirically derived mathematical relations estimated using data-driven system identification approaches /12/, /7/, /17/. The analytical redundancy approaches are commonly regarded as soft computing approaches or soft sensors /9/. Redundancy can also be approached with knowledge-based methods /13/, where problems are solved with qualitative rule-based models. Temporal redundancy of a sensor value is obtained through repetitive measurements of the sensor values at regular intervals /11/. The information originating from the redundant sensors or the derived models is compared with the original sensor values and the congruence is then analyzed to clarify the differences. The actual sensor faults can be classified into bias, drifting, precision degradation and complete failure /16/. The sensor faults have to be distinguished from process changes which can also be abnormal and influence the measurements. Several approaches to process monitoring and process fault detection have been proposed in the application area of soft sensors /9/. In recent years, many research projects have developed software applications for the analysis and monitoring of process data and fault detection also in power plant environments. The general goal of such approaches is to improve the overall reliability of the process, and therefore, they themselves must be very reliable. The computational effort must be reasonable and the used methods should be understandable and tunable. An unsupervised neural network algorithm, called Self-Organizing Map (SOM), has been a popular choice in