The operating condition data of pumping unit can be described from the aspects of displacement,load and electric current.If only one feature vector is used to describe the operating condition of the pumping unit,the information will be too simplified,and it will lost some effective information.In view of the above problems and polysemy which is the essential characteristics of operating condition data,the fault diagnosis of pumping unit based on multi-instance and multi-label is presented.In this study,the displacement,load and current data of the pumping unit are used as multiple instances of pumping unit working condition data bags.Using k-medoids clustering algorithm cluster the bags and convert bags into several instances.Each dimension of the new instance indicates the distance from the bags to each cluster center,and then the MLSVM algorithm is used to solve the multi label problem.Experimental results show that multi-instance and multi-label learning can diagnose the trouble of oil pumping machine timely and accurately.