This paper proposes an association rules reasoning model based on ameliorated trapezoidal cloud transformation. It is aimed primarily at complexity and randomness geosciences data bears. The traditional trapezoidal cloud transformation is improved in order to avoid lack of data mutation information and to finish reasonable and sensitive exchange from qualification to quantification. A set of attributes for simulating faults extraction algorithm is designed, which breaks through limitations of traditional visual interpretation and ensures an effectiveness and completeness of test data. Multi-Level Association Rules (MLAR) model [1] is also adopted to reason and predict unknown faults and fault properties in Chengdu Office zone. The result shows that the MLAR algorithm enhanced an association mining between fault types with their classified attributes.