Stacking-Based Multi-Features Fusion Ensemble Learning for Product Quality Prediction

Product quality prediction, which predicts the quality of products with process parameters in industrial production lines, is a critical factor for quality control in the Industrial Internet of Things (IIoT). Current researches rely on a large amount of data to train to achieve accurate prediction results. Empirically, the more balanced data are used for training, the higher accuracy of product quality prediction can be achieved. However, in the modern real-world industry, multiple varieties and small batches become the characteristics of the production environment, resulting in insufficient data being available for training. It is thus raised that “how to predict product quality with small sample data in an IIoT environment”. To this end, in this paper, a stacking-based multi-features fusion ensemble learning for product quality prediction is proposed, where multiple base models generate stacking features while the final model learns from the fusion features and predicts product quality. Firstly, a 5-fold cross-validation is applied for multiple base models to generate 5 sets of predicted values. Then, the base models that have completed training predict product quality. Secondly, the predicted values from the training and prediction phases are both stacked and fused. At last, the fused features are learned by the final model for product quality prediction. Experimental results demonstrate the feasibility and effectiveness of our method. We also apply the proposed method to a real-world industrial project as a case study, where our model reaches an acceptable practical range. The source code of the proposed method is released on https://github.com/labcodell06/SMFFEL.