Advanced hybrid LSTM-transformer architecture for real-time multi-task prediction in engineering systems

Area of Science:

• Engineering Systems
• Machine Learning
• Data Science

Background:

• Real-time predictions are crucial for operational performance, safety, and efficiency in engineering systems, especially in underground drilling and green stormwater management.
• Traditional predictive models often lack the adaptability and accuracy required for dynamic operational conditions.

Purpose of the Study:

• To introduce a novel LSTM-transformer hybrid architecture for multi-task real-time predictions.
• To enhance operational performance, safety, and efficiency in specialized engineering applications.
• To develop a predictive framework that dynamically adapts to changing conditions and new data.

Main Methods:

• Developed a hybrid LSTM-transformer architecture integrating attention mechanisms and sequence modeling.
• Incorporated online learning for dynamic adaptation to operational conditions.
• Utilized knowledge distillation to transfer insights from larger networks, optimizing computational resources.

Main Results:

• The proposed model demonstrates superior predictive accuracy compared to existing methods.
• Achieved high real-time adaptability to variable operational conditions.
• Showcased significant computational efficiency without compromising performance.

Conclusions:

• The novel LSTM-transformer hybrid architecture offers a robust and effective solution for real-time predictions in engineering systems.
• The model provides actionable insights and significant advantages in accuracy, adaptability, and efficiency.
• This work pioneers a predictive framework for targeted engineering applications, improving operational outcomes.