ABSTRACT

Various activities required for acceptable/optimal and safe process operations, such as process control, and fault detection and diagnosis, depend critically on the sensors measuring various process variables. In the last few years the problem of sensor network design has gained increasing attention in the process systems research community. Sensor network design problem is to select the location, type and number of sensors measuring various process variables.

In this talk, a multiobjective optimization approach is proposed for locating sensors in a process. The objectives considered are reliability maximization, and cost minimization. The reliability of a variable is the probability of estimating the value of the variable, either by direct measurement or based on its relationship with other variables (through model equations) in presence of sensor failures. The sensor network reliability is defined as the lowest reliability amongst all variables. For a given set of measurements, the reliability of unmeasured variables is estimated by using equation-variable matching approach, wherein each unmeasured variable is assigned to a process model equation, and is expressed only in terms of measured variables. The only information used for this is the equation-variable occurrence matrix (whether a variable occurs in a particular equation or not), and the exact form (or the numerical values) of the equations is not required. Apart from maximizing reliability, cost minimization (total cost of locating sensors) is also considered.

A genetic algorithm is used to solve the resulting multiobjective optimization problem and generate the pareto front which shows the trade off between reliability and cost. Further, ideas similar to robust product design in experimental design literature are used to systematically select various tuning parameters associated with the Genetic Algorithm. The parameters so chosen would ensure robustness in solving the type of problems considered here, and hence the Genetic Algorithm need not be retuned if the problem size changes.

The application of the proposed approach is demonstrated via illustrative examples.