The invention provides a gear sample database construction method which comprises the following steps: firstly, collecting actual gear parameters, and determining a label; constructing a generator anda discriminator through a neural network, wherein the initial value of each layer of weight coefficient obeys uniform distribution; inputting a noise sequence signal and a set tag into the generatorto generate data; inputting the generated data and the original sampling data into the discriminator at the same time for classification and judgment; achieving dynamic balance by updating the weightof each neuron in the generator and the discriminator; performing sorting in an ascending order according to the distances; and sequentially selecting the different quantities of generated data from the generated data queues of each category, and fusing the generated data into the original sample data to serve as a final training database for a classifier to use. The method has the advantages thatthe requirement for the number of training samples of each category in a database can be met by generating new data under the condition that the sampling data is small, and the influence on the training effect of the classifier due to insufficient data volume is avoided.