Abstract: The field measurement of varied monitoring levels about the shaft freezing pressure in Dingji mine has shown that the freezing pressure changes with time and circumstance. It is easily influ- enced by many factors, such as strata depth, geotechnical moisture content, average freezing wall thick- ness and mean temperature, and that it has obvious uncertainty. The degree of uncertainty has been characterized by the variation coefficient, laying a foundation for optimizing the traditional RBF neural network and introducing variation coefficient of freezing pressure to fuzzy central value and weight value learning strategy to establish prediction model of shaft freezing pressure in deep alluvium. With variables of strata depth, geotechnical moisture content, average freezing wall thickness and mean tem- perature as the input information, the model has been used to distinguish clay strata and calcareous clay strata in training and learning with sample data from thirty three monitoring levels in seven shafts of Huainan and Huaibei areas, and finally has made an engineering prediction for the shaft freezing pres- sure in Kouzidong mine. The results have shown that the field measurement has well fitted with the pre- diction; the efficient and precise model algorithm has provided reliable basis for analysis and prediction of vertical shaft freezing pressure in deep alluvium of Huainan and Huaibei areas.

Key words: freezing pressure, variation coefficient, RBF fuzzy neural network, learning strategy optimization, engineering prediction