Dual Attention-Based Encoder-Decoder: A Customized Sequence-to-Sequence Learning for Soft Sensor Development.

Area of Science:

• Industrial Process Monitoring
• Machine Learning Applications
• Soft Sensor Development

Background:

• Soft sensors predict difficult-to-measure quality variables using easily measured process data.
• Sequential dependencies among process and quality variables are crucial for accurate modeling.
• Existing methods may not fully exploit temporal relationships in industrial data.

Purpose of the Study:

• To develop a customized sequence-to-sequence learning model for soft sensors.
• To explicitly utilize the sequential nature of both process and quality variables.
• To improve the accuracy of predicting hard-to-measure quality variables.

Main Methods:

• A dual attention-based encoder-decoder architecture was developed.
• Long short-term memory (LSTM) networks were used for encoder and decoder modules.
• A dual attention mechanism was integrated to identify key process variables and time points.
• The model was evaluated on cigarette production and multiphase flow processes.

Main Results:

• The proposed dual attention encoder-decoder model demonstrated effectiveness in soft sensor applications.
• The model successfully captured sequential dependencies in process and quality variables.
• The dual attention mechanism improved the identification of influential process variables and time instances.
• Accurate prediction of hard-to-measure quality variables was achieved.

Conclusions:

• The developed dual attention encoder-decoder model offers a powerful approach for soft sensor development.
• Exploiting sequential information and temporal dependencies enhances prediction accuracy.
• This method provides a fine-grained quality prediction for industrial processes.