Extreme learning machine has a series of achievements on batch processing due to high-activity processing， superior performance， less manual parameter settings and so on， which has been successfully applied in multi-label classification. However， data streams emerging in the real-world applications present the characteristics of high-volume， high-speed， multi-label and concept drift， which poses the challenges in accuracy， time and space consumptions for traditional multi-label classification algorithms. Therefore， this paper proposes a multi-label classification data stream ensemble approach based on kernel extreme learning machine （KELM）. Firstly， to adapt to the environment of data streams， the sliding window mechanism is used to partition data chunks， and an ensemble model consisted of k KELM models is built on k data chunks. Meanwhile， considering the label correlation， the Apriori algorithm is used to achieve the association rules of labels， and the confidence of label occurrence is introduced in the prediction using the generated model. Secondly， the MUENLForest model is introduced to detect whether a concept drift occurs in the new arriving data chunk， correspondingly the loss function is specified to update the ensemble model for adapting to concept drifts. Finally，massive experiments on the real multi label data sets demonstrate that the proposed approach outperforms the traditional multi label classification methods in accuracy and can adapt data drifts in multi label data streams quickly.