Multi-Level Modeling of Process Variability in Industrial Systems

Abstract

The implementation of Industry 4.0 technology in manufacturing operations creates a large quantity of heterogeneous and multidimensional datasets from the industry, thereby necessitating variability modeling to assure production quality and faults to achieve efficiency in manufacturing operations. However, traditional methods of building models often do not account for temporal information as well as structural relations between different machinery and procedures at different levels of operations in industries, hence resulting in poor performance predictions in such a dynamic setting. For this reason, an intelligent multi-level variability modeling approach based on the Long Short-Term Graph Neural Network (LST-GNN) is designed. The data produced from the industrial smart factory is pre-processed using Min-Max normalization method, and Linear Discriminant Analysis (LDA) is used for extracting features that are compact and informative but retain variability information. LSTM extracts time dependencies within the sequential data generated by the manufacturing process, while GNN learns structure dependencies in the network of interconnected machines, sensors, and production systems. It is shown that our LST-GNN model outperforms conventional approaches in terms of predictive accuracy, with Mean Absolute Error (MAE) and R² values being 0.030000 and 0.939500, respectively. Our proposed model provides higher accuracy, lower error, and higher robustness when used for process control and fault prediction within the domain of smart manufacturing. The LST-GNN is developed using Python programming language with support from TensorFlow, PyTorch Geometric, Scikit-learn, NumPy, Pandas, and Matplotlib libraries.