(325f) Fault Detection and Isolation of Hybrid Process Systems Using a Combined Data-Driven and Observer-Design Methodology

The effective and timely monitoring of chemical processes is of paramount importance in process system engineering to ensure process safety and to achieve high-quality, consistent products. Significant progress has been made over the past few decades by introducing approaches that utilize causal models for model-based fault detection and isolation (FDI), as well as multivariate statistical process monitoring (MSPM), referred to as data-driven techniques. Despite a rich body of literature in process monitoring, the majority of existing methods have been developed for purely continuous processes with only one single operating condition. However, many chemical processes are operated under substantially different operating conditions and more appropriately modeled by hybrid systems [1]. As a result, the multimodality of a hybrid system is characterized by switching between a finite number of operating modes or subsystems. Research on this issue has mainly focused on developing monitoring schemes by using either model-based or data-driven approaches. As for model-based FDI, the main challenge lies in the development of causal models for modern industrial plants with increasing size and complexity. In contrast, the data-driven techniques can often handle large-scale processes since only historical process data is required. However, these approaches may not work well with respect to fault diagnosis and yield misleading results.

The most important task in the design of FDI schemes for hybrid systems with switching modes is the identification and description of specific operating modes. An effort to address these issues was initiated in [2] where a family of mode observers is designed in a way that facilitates the identification of the active mode without information from controllers. However, the mode observers are only solvable under certain constraints and it requires sufficient accurate information available about the system state-space model. To relax these limitations, a novel mode identification scheme needs to be developed and incorporated into the hybrid systems structure. Another key issue is the ability of the fault isolation scheme to diagnose the abnormal situations effectively. Although MSPM has no requirement for prior system knowledge, conventional fault diagnosis using contribution plots, variable reconstruction, or statistical discriminate analysis may yield ambiguous and misleading results given that the underlying process dynamics is insufficiently or inaccurately described. To avoid potential errors, a fault isolation scheme that can distinguish different faults accurately needs to be addressed.

To tackle these challenges, we propose a new framework for fault detection and isolation of hybrid processes with switching modes. This framework combines previous research on data-driven and observer-design methodology by integrating Gaussian mixture models (GMM), subspace model identification (SMI), and results from unknown input observer theory (UIO), and consists of the following steps: (1) A GMM is built to identify and describe the multimodality of hybrid systems by using the recorded input/output process data, (2) A state-space model is obtained for each specific operating mode based on SMI, (3) An UIO is designed to estimate the system states robustly, based on which the fault detection is laid out through a multivariate analysis of the residuals, (4) A set of unknown input matrices are designed for specific fault scenarios, and fault isolation is carried out through the disturbance-decoupling principle from UIO theory. As for multimodal process, it is reasonable to assume that the process data of each individual mode possesses a specific data characteristic. Therefore, GMM can be utilized for mode identification and description. For this purpose, only sampled data is required and the constraints of the mode observers are relaxed. Furthermore, practical experience has shown that industrial processes can be approximated with sufficient accuracy by linear-invariant systems of finite dimension. Thus, the identified state-space models that describe the dynamics of normal conditions are faithful representations for each operating mode if the accurate first-principle model is difficult to obtain. The UIO is adopted here for the purpose of residual generation, and then a multivariate statistic, T², is calculated as the fault indicator, which takes the residual correlation into consideration. Fault isolation is an issue that is not adequately investigated for the hybrid process systems. Motivated by the decoupling-principle from UIO theory, a set of unknown input matrices can be designed to compensate for specific fault scenarios. Given that the isolated fault is decoupled from residual generation, only the UIO with the corresponding unknown input matrix would generate a T² statistic that does not violate the control limit. A significant benefit of combining data-driven and observer-design methodology for hybrid system monitoring is that they overcome the limitations of individual approaches. Finally, the fault detection and isolation capability of the new framework is demonstrated through three application examples: a numerical simulation, a simulated continuous stirred tank heater process, and the Tennessee Eastman process. The results show a significant potential for this approach.

References:

[1] Christofides, P.D. and El-Farra, N.H. Control of Nonlinear and Hybrid Process Systems: Designs for Uncertainty, Constraints and Time-Delays, 446 pages, Springer-Verlag, Berlin, Germany, 2005.

[2] Hu, Y. and El-Farra, N.H. Robust Fault Detection and Monitoring of Hybrid Process Systems with Uncertain Mode Transitions. AIChE Journal, 57(10), 2783-2794, 2011.