In recent years, the semiconductor industry has developed rapidly, increasing competition among companies. The chip testing process is complex and costly. If a machine learning method is used to establish a model for identifying chip quality anomalies, the testing cost can be reduced. However, the chip data in the actual production environment drifts, which affects the performance of the model. Therefore, this paper proposes an incremental bagging Broad Learning System method. First, the offline data is balanced by resampling, and then the bagging sampling method is used to form multiple sub-datasets and train multiple BLS networks to make up for the problem of insufficient performance of a single BLS. At the same time, the model is incrementally updated using the data arriving online to adapt the model to the drifting data. The effectiveness of our method has been verified through experiments on real new chip test data.