The fault identification of transformer based on the characteristics of dissolved gas contents in the oil is of great significance to its safe operation. The deep belief network (DBN) is selected as a fault identification model in this paper since it has unique for extracting features from sample data. Firstly, the original data set of dissolved gas in transformer oil is directly deployed as the inputs of the training model, and the DBN is processed through three intelligent search algorithms.Three important parameters of mid-batch processing, gradient descent learning rate, and number of neural units are intelligently optimized to solve the problem of a low fault recognition rate when input raw sample data are limited. It is shown that the performance of the proposed method is better than the particle swarm search (PSO) algorithm and genetic algorithm (GA) search optimization. The total recognition rate of CS-DBN is 4.2% higher than that of GA-DBN, 2.5% higher than PSO-DBN and 56.2% higher in evolution efficiency. It also has a good generalization performance.