RNN- and LSTM-Based Soft Sensors Transferability for an Industrial Process

Area of Science:

• Process Systems Engineering
• Artificial Intelligence in Industry
• Machine Learning Applications

Background:

• Soft Sensors (SSs) are crucial for the process industry, but their design faces challenges in model accuracy, data availability, and computational complexity.
• Black-box machine learning (ML) methods are commonly used for SS implementation, requiring extensive effort in variable selection, model class, order, and hyperparameter tuning.
• Knowledge transfer between similar industrial processes for SS design remains an underexplored area.

Purpose of the Study:

• To investigate the application of transfer learning for designing Soft Sensors (SSs) by transferring knowledge from a source to a similar target process.
• To evaluate the efficiency and effectiveness of transfer learning methods in reducing the computational time and effort required for SS design.
• To assess the performance of different recurrent neural structures, specifically recurrent neural networks and long short-term memory architectures, in the context of transfer learning for SSs.

Main Methods:

• Developed and proposed two transfer learning methods to facilitate knowledge transfer for SS design.
• Utilized recurrent neural networks (RNNs) and long short-term memory (LSTM) architectures to implement the Soft Sensors.
• Conducted an industrial case study to evaluate the performance of the proposed transfer learning procedures and compare them based on SS performance and computational cost.

Main Results:

• The proposed transfer learning methods demonstrated suitability for designing SSs based on nonlinear dynamical models for industrial systems.
• Recurrent neural structures, including RNNs and LSTMs, showed varying degrees of transferability, with specific architectures potentially offering better performance.
• The study addressed the challenge of labeled data scarcity in the target domain, showing that transfer learning can mitigate this issue.
• A favorable compromise between SS performance and computational effort was achieved through the implemented transfer learning strategies.

Conclusions:

• Transfer learning offers a viable and efficient approach to designing Soft Sensors for industrial processes, significantly reducing development time and computational resources.
• The effectiveness of transfer learning is confirmed for nonlinear dynamical systems, paving the way for more streamlined SS development.
• Recurrent neural network architectures, particularly LSTMs, show promise for knowledge transfer in SS applications, though careful selection is advised.
• The proposed methods provide a practical solution for leveraging existing knowledge in new, similar industrial contexts, even when labeled data is scarce.