On robust and adaptive soft sensors.

Abstract

In process industries, there is a great demand for additional process information such as the product quality level or the exact process state estimation. At the same time, there is a large amount of process data like temperatures, pressures, etc. measured and stored every moment. This data is mainly measured for process control and monitoring purposes but its potential reaches far beyond these applications. The task of soft sensors is the maximal exploitation of this potential by extracting and transforming the latent information from the data into more useful process knowledge. Theoretically, achieving this goal should be straightforward since the process data as well as the tools for soft sensor development in the form of computational learning methods, are both readily available. However, contrary to this evidence, there are still several obstacles which prevent soft sensors from broader application in the process industry. The identification of the sources of these obstacles and proposing a concept for dealing with them is the general purpose of this work. The proposed solution addressing the issues of current soft sensors is a conceptual architecture for the development of robust and adaptive soft sensing algorithms. The architecture reflects the results of two review studies that were conducted during this project. The first one focuses on the process industry aspects of soft sensor development and application. The main conclusions of this study are that soft sensor development is currently being done in a non-systematic, ad-hoc way which results in a large amount of manual work needed for their development and maintenance. It is also found that a large part of the issues can be related to the process data upon which the soft sensors are built. The second review study dealt with the same topic but this time it was biased towards the machine learning viewpoint. The review focused on the identification of machine learning tools, which support the goals of this work. The machine learning concepts which are considered are: (i) general regression techniques for building of soft sensors; (ii) ensemble methods; (iii) local learning; (iv) meta-learning; and (v) concept drift detection and handling. The proposed architecture arranges the above techniques into a three-level hierarchy, where the actual prediction-making models operate at the bottom level. Their predictions are flexibly merged by applying ensemble methods at the next higher level. Finally from the top level, the underlying algorithm is managed by means of metalearning methods. The architecture has a modular structure that allows new pre-processing, predictive or adaptation methods to be plugged in. Another important property of the architecture is that each of the levels can be equipped with adaptation mechanisms, which aim at prolonging the lifetime of the resulting soft sensors. The relevance of the architecture is demonstrated by means of a complex soft sensing algorithm, which can be seen as its instance. This algorithm provides mechanisms for autonomous selection of data preprocessing and predictive methods and their parameters. It also includes five different adaptation mechanisms, some of which can be applied on a sample-by-sample basis without any requirement to store the on-line data. Other, more complex ones are started only on-demand if the performance of the soft sensor drops below a defined level. The actual soft sensors are built by applying the soft sensing algorithm to three industrial data sets. The different application scenarios aim at the analysis of the fulfilment of the defined goals. It is shown that the soft sensors are able to follow changes in dynamic environment and keep a stable performance level by exploiting the implemented adaptation mechanisms. It is also demonstrated that, although the algorithm is rather complex, it can be applied to develop simple and transparent soft sensors. In another experiment, the soft sensors are built without any manual model selection or parameter tuning, which demonstrates the ability of the algorithm to reduce the effort required for soft sensor development. However, if desirable, the algorithm is at the same time very flexible and provides a number of parameters that can be manually optimised. Evidence of the ability of the algorithm to deploy soft sensors with minimal training data and as such to provide the possibility to save the time consuming and costly training data collection is also given in this work.