The invention relates to the technical field of data mining, in particular to a machine error data classification method and system. A label propagation algorithm is introduced as a machine error data preprocessing step, and labels of uncalibrated data are rapidly estimated by the aid of a small number of calibrated machine data to form a classification training set. Based on the machine data and the labels in the classification training set, dictionary learning with consistent labels is performed, reconstruction errors are minimized, sparse coding errors are judged, the errors are classified, a reconstruction dictionary, sparse codes and a multi-class linear classifier are obtained, and the relationship among items in the dictionary and the data labels is kept. Furthermore, machine data characteristics are represented by the obtained sparse codes and inputted to the classifier for prediction, classes of test samples are determined, and the machine data errors are classified. By introducing efficient semi-supervised data preprocessing, the number of calibrated machine samples is increased, prior information is enriched, and machine data classification accuracy is effectively improved.