Integrating Causal Models and Trend Analysis for Process Fault Diagnosis

Mano Ram Maurya - 1

Thesis Summary

Integrating Causal Models and Trend Analysis for Process Fault Diagnosis

Mano Ram Maurya

Venkat Venkatasubramanian

Quick fault detection and diagnosis (FDD) are key issues for safe and optimal process operation and hence they form the central part of any Abnormal Event Management (AEM) system. Quick corrective action can help in minimizing the quality and productivity offsets and can assist in averting hazardous consequences in abnormal situations. Further, quick detection and diagnosis allows more time for the operator to perform counter-measure planning. Hence incipient fault diagnosis (IFD) is an important area of research.

The focus of my thesis is two-fold: (1) development of techniques to utilize various forms of information, such as sensor data, qualitative models etc., for fault diagnosis, and (2) seamless integration of various techniques to design a robust diagnostic system with high resolution. The model representations used in the diagnostic system are: signed directed graphs (SDG), trend-based fault-signatures and principal components. Unlike the parallel interactions between various components of blackboard architecture-based frameworks proposed in the literature, various techniques in the proposed diagnostic system interact in a series-parallel manner, thus resulting in improved isolation of the faults. Among various models, SDG-based models have been widely used for IFD due to their ability to deal with uncertainty, incomplete information and noise. Reasoning is also easy with SDG-based models. Despite considerable work on SDG, large-scale applications are seldom seen in the literature. Also, till recently, very little attention has been given to automatic development of SDG and their analysis.

My thesis deals with in-depth analysis of SDG. It involves thorough analysis of the algorithms for digraph generation, design of methodologies for digraph analysis, and their applications for qualitative simulation and fault diagnosis. A framework to model and analyze control loops by using SDG also has been presented. We have shown that the framework can handle systems with compensatory response. Due to qualitative nature of SDG models, prediction of spurious results is inherent in SDG-based fault diagnosis. Hence, to improve diagnostic resolution (discrimination between different faults), we propose to choose sensors by using SDG-based optimal sensor location.

We have also developed an SDG and QTA (qualitative trend analysis) based framework to enhance resolution. QTA is a data-driven approach that works by representing sensor trends as a sequence of certain basic shapes (called primitives) and matching the current sensor-trends against the sensor-trends for various faults stored in a database. The qualitative trends are extracted by using an interval-halving methodology. The extent of match between two trends is calculated using a fuzzy-inference approach. QTA, being a data-driven approach, can be easily applied in real industrial processes for fault diagnosis. Though the SDG-QTA framework is very efficient in detecting and diagnosing faults at an early stage, sometime false negatives (the actual fault is missed) are generated due to the inability to measure the initial response correctly. We have proposed wavelet-based de-nosing to minimize incorrect measurement of initial response. The framework is tested on the Tennessee Eastman (TE) case study. The combination of SDG and QTA approach with robust initial change detection based on wavelet-based de-noising ensures completeness (due to the SDG) and at the same time results in a diagnostic system with high resolution. Further, the computational complexity is reduced as compared to a diagnostic system that uses QTA alone.

My thesis also deals with the use of principal component analysis (PCA) for fault diagnosis. Though, PCA has been widely used for process monitoring and fault detection, for fault diagnosis their application is limited to the diagnosis of sensor and actuator faults. PCA works by projecting the data onto a lower-dimensional space while ensuring that most of the variations in the data are captured. To perform process monitoring and fault detection, T2 (cumulative sum of the squared normalized principal components) and SPE (squared prediction error) plots are monitored. Thus the operator does not need to monitor individual sensors. In my thesis, an attempt has been made to perform fault diagnosis by analyzing the qualitative trends of the principal components. The application of the proposed approach on TE process has shown that the computation time is reduced by a factor of two or even more while preserving sufficient fault discrimination capability.