Can This Sensor be Removed? Investigating ML Models for Virtual Sensors in Injection Molding

Abstract

Injection molding relies on physical sensors to monitor critical variables such as pressure, temperature, and actuator positions. However, these sensors can be costly to install and maintain, especially in industrial environments with restricted access or calibration needs. This thesis investigates whether selected physical sensors can be replaced by data-driven virtual sensors and evaluates how different machine learning models compare in achieving this.

Using real-world multivariate time series data from Tetra Pak’s injection molding process, six machine learning algorithms of varying complexity were benchmarked: Linear Regression, XGBoost, Feedforward Neural Networks (FFNN), Gated Recurrent Units (GRU), Long Short-Term Memory networks (LSTM), and Transformers. Each model was trained to reconstruct the full signal of a target sensor using a fixed-length time window from the remaining sensors. Performance was measured using range-normalized root mean squared error (RA-RMSE) to enable cross-feature comparison.

Results show that GRU and Transformer architectures consistently achieved the lowest RA-RMSE values, particularly for pressure and actuator-related sensors. Temperature signals were harder to model accurately, likely due to long-term dependencies beyond the available input window. Additionally, feature importance analysis revealed trade-offs in sensor removal some sensors are highly predictable but also crucial for estimating others.

The findings highlight both the feasibility of using virtual sensors and the importance of model selection. Comparative analysis shows that model architecture significantly affects prediction accuracy, guiding future decisions in sensor design and digital transformation in manufacturing.