The invention relates to a deep learning-based balanced sampling and modeling method for unbalanced data. The method comprises the steps of extracting majority class and minority class sample sets ina data set, and performing counting respectively; and carrying out multi-time K-Means clustering on the majority class sample set to obtain R clustering results; performing clustering fusion on the Rclustering results by adopting a clustering fusion algorithm based on an incidence matrix, thereby obtaining a new majority class sample set; performing over-sampling on the minority class sample set,thereby obtaining a new minority class sample set; combining the obtained new majority class sample set and new minority class sample set, thereby forming a new data set with class balance; extracting an abstract feature of the new data set with the class balance, and adding the abstract feature as a new feature into a feature set of the new data set with the class balance, thereby forming a newfeature set; and training a DBN model by adopting the obtained new feature set, thereby obtaining an optimal DBN model. The method overcomes the shortcomings in a single processing method, has a better processing capability, and has relatively high accuracy.