Abstract: A new method is proposed for extracting general rules from an artificial neural network, which is trained by destructive learning. The method consistes of two phases of preprocessing and rule extracting. The preprocessing phase contains three parts:dynamic modification, cluster and pruning. The dynamic modification generates automatically or constructs a fully connected or non fully connected preliminary topological network having one hidden layer from initial rule set. Redundant and unimportant hidden unites and links are deleted from a trained network respectively in the cluster and pruning phase, and then, the link weights remaining in the network are retrained to obtain the same MSE. Thus, we can obtain a simpler and smaller size network structure and extract a relatively small rule set by using pruned network. The method is applied to meteorologic cloud atlas data from AD reports of USA. Test results using test data set show the correctness and effectiveness of proposed method, and this method is a simple and feasible one.