Fault Detection and Diagnosis in Dynamic Multivariable Chemical Processes Using Speech Recognition Methods

Fault Detection and Diagnosis have become important topics in the process industries. The off-line diagnosis of past transient upsets can lead to important process or operation modifications that can improve the future behavior of the process. The rapid on-line diagnosis of faults is even more important since it can anticipate and minimize the impact ofotherwise costly effects. The first part of this thesis addressed the problem of fault diagnosis in multivariate, dynamic, continuous chemical processes. Two types of faults were considered: deterministic (whose root cause is a randomly occurring detenninistic event) and stochastic (caused by an underlying stochastic process). A realistic simulation of a chemical plant was used as a test bed for the proposed methods. Due to the lack of accurate dynamic models for this type of process, a Pattern Recognition approach was followed. Within this framework, several methods were designed for the on-line and offline diagnosis of both types of faults. All methods consisted of: 1) a feature extraction step, where magnitude invariant features are extracted from both the reference patterns and the pattern of the new unknown fault, and II) a similarity assessment step where the distance between the new pattern and each of the reference patterns is estimated using Dynamic Time Warping. Due to the use of magnitude invariant features and the ability of Dynamic Time Warping to synchronize similar patterns with distorted temporal correlations, the results were satisfactory in diagnosing detenninistic faults. In the case of stochastic faults, the results were inconclusive. The correlation pattern between the variables was used as the feature for the diagnosis of stochastic faults. However, the slow dynamics and the effect of the recycle in the simulated chemical plant meant that unrealistically long records of data are required for an accurate estimate of this feature.