Imbalanced data classification for defective product prediction based on industrial wireless sensor network
Zhou Hong, Kun-Ming Yu · 2017
In the Industry 4.0 era, manufacturers can establish smart factories based on the industry wireless sensor network (IWSN). And data analysis plays a vital role to realize smart manufacturing. However, data collected in the real production through IWSN generally represent features as incomplete and imbalanced resulting in incorrect or biased analysis results. Therefore, a solution is proposed to resolve this problem, in which K Nearest Neighbor (KNN) algorithm is applied to do missing value imputation and Adaptive Synthetic Sampling algorithm is utilized to generate a balanced dataset. Furthermore, a 2-layer feedforward neural network (FNN) is designed as a classifier to predict defective products. The classification performance in testing using the resolution proposed is far superior to that of 2-layer FNN using the original dataset directly or employing the KNN algorithm for preprocessing first whose recall value is 94%, the precision value is 87.9%, and the F1-measure value is 90.8%. To sum up, the solution proposed can improve the classification performance dramatically, especially for the minority class, when encountering the incomplete and imbalanced data.