For the research of rough set theory and genetic algorithm, an improved adaptive genetic reduction algorithm
based on knowledge dependability is presented. There are four characters of the algorithm, in which the initial population of binary code is restricted by attribute core of decision table, the fitness function in the algorithm is defined through the dependability that decision attribute for condition attribute, the adaptive crossover probability and adaptive mutation probability are improved, and correction operator is added into the new generation individuals. By using the algorithm, radar faults is diagnosed and the simple rules can be obtained automatically. The diagnosis rules have the characteristic that fault premonitions and fault reasons parallelism can be correlated one by one. The defects of worse system veracity and lower efficiency can be avoided about expert fault diagnosis based on traditional fault tree.