(476a) Active Fault Isolation of Nonlinear Process Systems

Reliability and safety are of particular importance to process industries. Automatic control techniques have significantly improved the efficiency, reliability, and safety of chemical process operations. In a chemical plant, however, various faults are ubiquitously present, which can take place in control or processing equipment, due to reasons such as power failure, long-term use, or human-made mistakes. These abnormalities can jeopardize the profit of process operations or even lead to safety hazards to equipment and personnel. This realization has motivated significant research efforts to develop automated techniques that can detect the occurrence of faults, identify the failed equipment, and achieve fault tolerance. These techniques can be used to enhance safety and avoid potential economic losses in an intelligent way. The existing results on fault detect and isolation (FDI) can be categorized into data-based and model-based approaches.

In the data-based approach, faults are often detected through statistical process monitoring. From normal plant operating data, empirical correlation models can be built by using multivariate latent variable methods (e.g., [1]), such as principal component analysis (PCA) and partial least squares (PLS). These models are low dimensional and can capture the key information in normal process data. The current process data are compared with the normal variation contained in these low dimensional models, and the abnormal behavior is detected through statistical tests. Faults can then be isolated by using contribution plots, which are usually powerful to isolate simple faults, the effects of which are not propagated into other variables. The isolation of complex faults, the effects of which are propagated into many other variables (e.g., through feedback control), is improved by using additional data on past faults in [2]. The major benefits of this approach are that it can handle the case where there are a large number of measured variables and first principles models are unavailable. Due to the non-causal nature of the model built from normal plant operating data, however, the problem of fault isolation remains difficult when data on past faults are unavailable, which motivates the development of the model-based approach.

In the model-based approach, FDI is often achieved by generating residuals through the system model and input/output data (see [3] for a review). Under fault-free conditions, the magnitudes of these residuals are small. A fault is reported when a residual breaches the user-specified threshold. This approach has been studied extensively for linear (e.g., [4, 5]) and nonlinear (e.g., [6, 7]) systems. Due to the presence of plant-model mismatch, residuals that are sensitive to faults but insensitive to uncertainty and disturbances are desired. Unknown input observers are developed in [4] to decouple the effect of unknown inputs, such as disturbances, from that of the faults for linear systems. For nonlinear systems, the problem has been studied by using uniform thresholds in [6] (and adaptive thresholds in [7]), where fault isolation relies on the existence of a state variable which is directly and uniquely affected by the potential fault. Most of the existing FDI methods, however, are designed by using the measurement data, which are obtained from the process under the nominal control law. Therefore, these methods may not remain effective if the structure of the close-loop system inherently does not allow fault isolation under the nominal control law.

In comparison, there exist limited results on active fault isolation, namely utilizing the control action to facilitate fault isolation. Along this line, a method of decoupling the dependency between certain process state variables through feedback control (exploiting the system structure) has recently been proposed to enhance fault isolation under full state feedback control (e.g., [8]), where it is assumed that the state measurements are normally distributed. More recently, this approach has been extended to handle the case where only output measurements are available and studied with the use of model predictive control to optimize the input cost in [9]. Furthermore, the problem of distinguishing between faults that may directly affect the dynamics of the same process state (e.g., process disturbances and actuator faults) remains difficult. This problem is partly addressed in [10] for actuator faults by monitoring the outputs of the actuators, where it is assumed that the outputs of the (healthy or failed) actuators are constant between two consecutive discrete times and there exists a subsystem of the plant that satisfies a full rank condition.

Motivated by the above considerations, we in this work consider the problem of designing an active fault isolation scheme for nonlinear process systems. In particular, we consider the case where multiple faults may directly affect the dynamics of the same process state. To this end, we first develop a fault-tolerant control law by treating faults as disturbances, which can be used to achieve the desired control objective in FDI. We also develop a fault detection method that accounts for plant-model mismatch to provide timely and accurate information on the occurrence of faults. To utilize the control action to facilitate fault isolation, we design a mechanism to operate the process, upon fault detection, in a way such that for a given fault, the effect of other potential faults on the process dynamics can be reduced to an insignificant level. This allows the isolation of a particular fault by treating other potential faults as process disturbances. A simulation example of a chemical reactor is used to illustrate the effectiveness of the proposed method.

References

[1] J. V. Kresta, J. F. MacGregor, and T. E. Marlin. Multivariate statistical monitoring of process operating performance. Can. J. Chem. Eng., 69:35–47, 1991.

[2] S. Yoon and J. F. MacGregor. Fault diagnosis with multivariate statistical models part I: Using steady state fault signatures. J. Proc. Contr., 11:387–400, 2001.

[3] P. M. Frank. Fault diagnosis in dynamic systems using analytical and knowledge-based redundancy: A survey and some new results. Automatica, 26:459–474, 1990.

[4] J. Chen, R. J. Patton, and H.-Y. Zhang. Design of unknown input observers and robust fault detection filters. Int. J. Contr., 63:85–105, 1996.

[5] F. Hamelin and D. Sauter. Robust fault detection in uncertain dynamic systems. Automatica, 36:1747–1754, 2000.

[6] P. Mhaskar, C. McFall, A. Gani, P. D. Christofides, and J. F. Davis. Isolation and handling of actuator faults in nonlinear systems. Automatica, 44:53–62, 2008.

[7] X. Zhang, M. M. Polycarpou, and T. Parisini. Fault diagnosis of a class of nonlinear uncertain systems with Lipschitz nonlinearities using adaptive estimation. Automatica, 46:290–299, 2010.

[8] B. J. Ohran, D. Munoz de la Pena, J. F. Davis, and P. D. Christofides. Enhancing data-based fault isolation through nonlinear control. AIChE J., 54:223–241, 2008.

[9] B. J. Ohran, J. Liu, D. Munoz de la Pena, P. D. Christofides, and J. F. Davis. Data-based fault detection and isolation using feedback control: Output feedback and optimality. Chem. Eng. Sci., 64:2370–2383, 2009.

[10] M. Du, J. Nease, and P. Mhaskar. An integrated fault diagnosis and safe-parking framework for fault-tolerant control of nonlinear systems. Int. J. Rob. & Non. Contr., Submitted.