First-hand Experience And Second-hand Knowledge In Prediction Of Silicon Content In Hot Metal Tapped From Blast Furnace

It is essential to predict silicon content in hot metal accurately for the purpose of controlling blast furnace under good operation condition. In recent years, great progress has been achieved on applying rule-based artificial intelligence methodology to prediction of silicon content in hot metal. However, expert systems has disadvantage in knowledge acquisition. Self-organization experience evolution prediction model of prediction of silicon content in hot metal consists of extracting dynamic feature from data, classifying dynamic pattern and treating and improving prediction experience, which is composed of first hand experience and second hand experience. In this thesis, the mean, gradient and fluctuation of process variables of blast furnace have been used to extract the features of process variables. ART2 neural network has been used for classifier of dynamic patterns of process variables. Prediction experience has been represented based on production rule and frame description. Meanwhile, a data-mining algorithm used to discover relevant rules based on attribute reduce with maximal amount of information has been introduced.With process data of blast furnace No.1 in Tianjin Iron Plant, tests have been conducted to predict the silicon content in hot metal. It has been shown that the feature abstract by using the mean, gradient and fluctuation of process variables of blast furnace valid for predicting silicon content in hot metal. The data-mining algorithm could be used to extract effectively quantitative rules. The hit ratio of prediction of silicon content in hot metal will increase with higher support level of rules. In order to achieve better prediction, support level and confidence level of rule have to be 3 and 0.6 respectively. Most of the exctracted rules are subject to operation experience from human expert. However, few cases,in which exctracted rules are contradictionary with present operation experience from human expert, may imply that new knowledge about blast furnace operation would be discovered.It is also shown that the hit ratio of prediction is 75% with accuracy ±0.15 of silicon content in hot metal. When accuracy is ±0.10 of silicon content in hot metal, better prediction is acchived with random walk model rather than built up intelligent model, which brings better prediction when accuracy is ±0.15. The hit ratio dependson strongly the stability of blast furnace. Therefore it implies that lower hit ratio of prediction for an instable furnace condition could be more useful than higher hit ratio of prediction for stable furnace condition.Finally, it is pointed out that hit ratio should not be regarded as only criteria for prediction of silicon content in hot metal. Approach from coding space of patterns to concept based on pattern classification and knowledge discovery has been discussed.