CAREER: Model Free Fault Detection for Nonlinear Systems

Colorado School Of Mines, Golden CO

Investigators

Abstract

This proposal will deal with concerns of the development and application of methods for fault detection of nonlinear dynamical systems. Fault detection of linear systems is by now well established, but application of these techniques to nonlinear systems can lead to a loss of performance. The market for fault detection is growing, as more and more manufacturing and industrial centers are interested in improving their productivity by reducing unplanned maintenance and limiting losses to product due to equipment failure. In addition, many manufactures have installed sophisticated sensor and data collection networks in order to closely monitor their processes, which would enable the application of very sophisticated fault detection schemes. Fault detection can be as simple as placing limits on the allowable range of sensor measurements, but when the systems that are monitored are dynamic systems, more sophisticated methods can and should be applied. This is because dynamic systems, by the their very nature, have temporal relationships between the system variables that characterize both normal and faulted operating conditions. The challenge is to codify these temporal relationships. One method is to develop analytical models from first principles. In most model based fault detection schemes, output observers are designed which generate residual signals that measure the disagreement between the model and the measured data. However, these models are sometimes difficult to come by. The systems described above are nonlinear thermo-fluid systems that are difficult to model accurately and the resulting models would have many unknown parameters. In this case, system identification methods can be employed. This proposal is aimed at developing a theory of the interactions between the system identification and fault detection processes. The focus is on a general, practical method of nonlinear fault detection which is easily implemented on input/output models that are generated by modern nonlinear system identification methods. The basic question that is addressed is to determine how the choices made during the system identification process, such as the model structure. The objective function, the choice of regularization, and the validation procedure influence or interact with the fault detection process.