Abstract
By estimation, the cost of process failure is about 20 billions $US in U.S., about 5% of the total GDP contributed by the petroleum refining and petrochemical industry. As the year of 2000, the total GDP contributed by petrochemical industry in Taiwan will reach 2 trillions $NT. Proportionally, the loss due to process faults will as much as 100 billion $NT. So, the research of how to improve the process safety is valuable for academic and economic.In modern chemical plants, computers record hundreds of thousands data in a short period. Some of these data are associated with "normal" operating states. Others belong to various "faulty" situations. The goal of fault detection and diagnosis is to recognize those faulty situations. Basically, fault detection and diagnosis is a software technology that uses the existing hardware to detect and isolate the incipient or latent faults. Then inform the operator to take some suitable actions before the system suffers from severe damage. Therefore, among the variety of process safety improving methods, fault detection and diagnosis is the most valuable research item and has potential of future developments.For a suitable on-line fault detection and diagnosis tool, it must satisfy a condition: when a new fault emerge, the tool should recognize it immediately and learn while not disturbing the old classification result. In neural network research, we called this problem as stability-plasticity dilemma. Since those supervised learning methods or PCA cannot satisfy this condition. Hence, unsupervised learning becomes the only way to find a suitable method for fault detection and diagnosis. Our research motive and goal is to choose and develop the most suitable method to detect and isolate fault.So far, only the neural network model, adaptive resonance theory (ART), developed by Grossberg in 70s tried to solve the stability-plasticity dilemma. ART is an unsupervised learning model and has much potential to apply to process monitoring. However, the initial design concept of ART is primarily for image recognition. We find such versions of ART will produce some problems if we apply them to fault detection and diagnosis directly. In order to preserve the advantages of ART and remove all the problems when applying to fault detection and diagnosis, we develop a new ART version called Digital ART that is an evolution from ART1. We found Digital ART can have better and more reasonable results than those old ART versions do both in static and dynamic systems.