Developing a robust classifier for fault detection in production environment

Recently, machine learning algorithms are widely applied to production such as failures identification, finished products inspection, and process monitoring. Applying these algorithms to fault detection makes it possible to eliminate additional tests or experiments which usually involve high expense and highly risk. However, when applying machine learning methods to the real world data, the class imbalance problem usually has been ignored. This problem is caused by imbalanced data, in which almost all the examples are labeled as one class whilst far fewer objects are labeled as the other class. When deal with such imbalanced data, a classifier induced from an imbalanced data set has high classification accuracy for the majority class, but an unacceptable error rate for the minority class. To solve this problem, this work proposed a novel method, called SOM (Self-Organizing Maps) based methodology. A process monitoring data has been provided to demonstrate the effectiveness of the proposed method. Experimental results indicated the proposed method outperforms traditional techniques, under-sampling and cluster based sampling.