Transformers for Multi-Horizon Forecasting in an Industry 4.0 Use Case.

Area of Science:

• Industry 4.0
• Robotics
• Artificial Intelligence

Background:

• Automated guided vehicles (AGVs) in Industry 4.0 rely on 5G multi-access edge computing (MEC) for remote control.
• Communication disruptions in 5G networks pose safety risks due to potential AGV deviations.
• Existing deep learning models for trajectory prediction lack flexibility and robustness against network instability.

Purpose of the Study:

• To propose a novel multi-horizon forecasting approach for predicting remotely controlled AGV deviations.
• To introduce two new transformer-based architectures optimized for multi-horizon prediction.
• To evaluate the performance of these novel models against traditional deep learning methods.

Main Methods:

• Development of two novel transformer architectures for multi-horizon forecasting.
• Comparative analysis with Long Short-Term Memory (LSTM) neural networks.
• Evaluation of prediction accuracy and real-time inference capabilities.

Main Results:

• Transformer-based models demonstrated superior performance over LSTM in both multi-horizon and fixed-horizon scenarios.
• The best multi-horizon model achieved prediction accuracy comparable to the best fixed-horizon model.
• Models incorporating time-sequence structures in inputs showed enhanced performance in multi-horizon predictions.
• Proposed models met real-time inference time constraints for decision-making.

Conclusions:

• The novel transformer-based multi-horizon forecasting approach effectively predicts AGV deviations, enhancing operational safety.
• These models offer a robust and flexible solution for managing AGV navigation in dynamic industrial environments.
• The developed models are suitable for real-time applications, ensuring timely corrective actions and mitigating risks associated with network instability.