The invention discloses a clustering algorithm-based transformer fault diagnosis system. The system comprises a sample collection module, a data pre-processing module, a sample analysis clustering module, a neural network training module and a fault category judging module, wherein the sample collection module is used for collecting data of gas dissolved in oil of a transformer and taking the data of gas dissolved in oil as original samples; the data pre-processing module is used for carrying out cleaning and screening pre-processing on the data in the original samples; the sample analysis clustering module is used for carrying out clustering analysis on the pre-processed original samples by adoption of a K-means clustering method, so as to obtain training samples and test samples; the neural network training module is used for constructing a neural network, carrying out optimal training and determining the number, central positions, widths and output weights of hidden layers; and the fault category judging module is used for inputting the training samples to the trained neural network, solving a posterior probability and judging the fault categories. The system disclosed by the invention is easy to implement and can improve the fault diagnosis correct and the work efficiency.