The indicator diagram of the pumping unit visually shows the working conditions of the pumping unit. However， the actual working conditions show typical long-tailed distribution characteristics， and the categories are seriously unbalanced. Traditional methods cannot accurately identify small categories of working conditions， and cannot obtain accurate identification of underground working conditions. Aiming at this problem， a cost-sensitive active learning algorithm based on distribution-driven multi-class long-tail data （CALA） is proposed. First， considering the characteristics of data distribution， the optimal number of clusters for the data is determined by minimizing the cost as the optimization objective. Second， the optimal number of clusters obtained before is updated by adding the pre-classification error cost. Then， a classifier is constructed by integrating the classification models. Finally， balance the data distribution iteratively. Using the real indicator diagram data of an oil field to test， the significant experimental analysis proves that CALA has better performance in the diagnosis of small categories of working conditions.