Model-based Data Fusion in Industrial Process Instrumentation

Abstract

Process sensors form an essential component of modern industrial production processes. They usually have to be operated under stringent environmental, safety, and cost constraints. Some of the key requirements on process instrumentation are: operation under harsh and varying environmental conditions, high reliability, fault tolerance, and low cost. The failure of process sensors may cause high losses due to plant breakdowns or out-ofspecification products. Therefore it is important to know as much as possible about the momentary states of a production process. In addition, the estimates need to have low uncertainty and high reliability. Sensor and data fusion can be key techniques to economically reach these goals and achieve higher performance than with isolated single point sensors alone. These methods enable the quantification of otherwise inaccessible quantities that cannot be deduced from a single sensor or management principle. Examples are the concentration measurement of ternary solutions or the tomographic estimation of spatially distributed material parameters from arrays of single point sensors. Industrial processes are usually operated within a defined environment, although there may be very harsh conditions like temperature variations, aggressive fluids, and high humidity. There are certain limits of operation. The nominal parameters, like desired product specifications as well as normal or acceptable fluctuations are known in advance while unknown encounters like in classical sensor fusion for target tracking, autonomous guidance, and battlefield surveillance are not within the scope of operation. This fact potentially establishes a precise and specific model of the industrial process at hand. The knowledge then contained in the process model can be fruitfully exploited in model-based data fusion. Generally, model-based approaches reach beyond straightforward methods like physical redundancy with majority votings or heuristic filtering operations. The achievable measurement precision as well as decision making reliability is usually higher in model-based approaches due to the additional regularization of the state or hypotheses space that is achieved with an appropriate model. However, at the same time special care needs to be taken to choose a model that allows for sufficient and representative variation. This is the only way the inherent variability of a process can be adequately represented. O pe n A cc es s D at ab as e w w w .in te ch w eb .o rg