Novel deep learning based soft sensor feature extraction for part weight prediction in injection molding processes

Abstract

In the current injection molding (IM) industry, it remains challenging to monitor and estimate production quality promptly. It is costly and time-consuming to measure part quality manually after each production cycle ends, which results in quality defects difficult to be captured in time. In this case, a soft sensor is essential to model the IM process and predict the final quality in real time with multi-source industrial production data. However, traditional data-driven modeling methods fail to take advantage of the information in complex high-frequency data from in-mold sensors, resulting in an inaccurate IM model and unsatisfactory quality prediction performance. To solve this problem, this paper proposes a novel soft sensor framework based on a teacher-student structure. After specialized preprocessing of multiple sensor time series data, a GRU-based autoencoder with an attention mechanism (GRU-A-AE) is trained as a teacher model, extracting deep implicit features involving valuable time sequential information. Then, a cascaded relationship among shallow feature points from sensor signals, deep features, and final part weights is established using back propagation neural networks (BPNNs). To demonstrate its effectiveness and superiority, the proposed soft sensor is trained and tested with practical IM data under normal and fluctuating production conditions, respectively. Compared with conventional methods, our method has higher prediction accuracy with testing RMSE of 0.1049 and $R^2$ of 0.9950 under normal conditions, which proves more valuable information in high-frequency sensor signals are explored from the teacher model and IM production dynamics are captured precisely. In addition, its better prediction performance in the case of production condition fluctuation verifies its strong robustness.