Nuclear power plants have abnormal operating procedures to prepare abnormal events occurring. An operator should choose and follow the appropriate procedure according to alarms and plant parameters which indicate the plant state. However, with enormous information, it is sometimes hard for the operators to judge the plant state in a short period of time. In the field, the skilled operators are well trained in the entry conditions of the abnormal operating procedures, so that they can quickly select a procedure that is appropriate to the current situation. Nevertheless, this task has a potential risk for less skilled operators to make mistakes of the judgement, which would result in response time delayed. Therefore, this paper suggests nuclear power plants abnormality diagnosis algorithm to support the judgement. This paper covers two of three steps to develop the diagnosis system; setting the training data production environment by analyzing the abnormal operating procedures and comparison between deep learning algorithms using the convolutional and recurrent neural networks. The abnormal operating data were generated from the nuclear power plant simulator. In addition, to reduce the dimensionality of the data, principal component analysis was used as data preprocessing. The algorithm is expected to reduce work load of the operators by providing selection of the proper procedure in a short time with high accuracy.