Hierarchical Self-Attention Network for Industrial Data Series Modeling With Different Sampling Rates Between the...

Area of Science:

• Industrial Process Control
• Data Science
• Machine Learning

Background:

• Dynamic modeling of industrial data series is crucial for quality prediction.
• Traditional models struggle with varying sampling rates and discard unlabeled data.
• Existing methods often overlook variable and sample interactions in quality prediction.

Purpose of the Study:

• To develop an adaptive dynamic modeling approach for industrial quality prediction.
• To address challenges posed by different sampling rates and underutilization of unlabeled data.
• To improve the consideration of variable and sample interactions in prediction models.

Main Methods:

• A hierarchical self-attention network (HSAN) was designed for adaptive dynamic modeling.
• Dynamic data augmentation was employed to incorporate unlabeled input sequences.
• Variable-level and sample-level self-attention layers were utilized to capture interactions and temporal dependencies.
• A long short-term memory (LSTM) network was integrated for final sequence modeling.

Main Results:

• The HSAN effectively integrates unlabeled data, overcoming sampling rate discrepancies.
• The model successfully captures both short-interval (variable interactions) and long-interval (sample dependencies) temporal dynamics.
• Experiments on an industrial hydrocracking process demonstrated the HSAN's effectiveness in quality prediction.

Conclusions:

• The proposed HSAN offers a robust solution for adaptive dynamic modeling in industrial quality prediction.
• HSAN enhances prediction accuracy by fully leveraging available data and capturing complex interactions.
• This methodology provides a significant advancement for real-time quality monitoring and control in industrial settings.