Interpretable Self-Attention Dynamic Inner Neural Network Model for Nonlinear Dynamic Process Monitoring

Area of Science:

• Industrial Process Monitoring
• Machine Learning
• Artificial Intelligence

Background:

• Nonlinearity and complex dynamics are key challenges in industrial process monitoring.
• Accurate fitting of these nonlinear dynamic relationships is crucial for effective monitoring performance.
• Existing methods often struggle with low computational efficiency and poor interpretability.

Purpose of the Study:

• To propose a novel process monitoring algorithm addressing nonlinearity, dynamics, computational complexity, and interpretability.
• To develop a method that accurately models dynamic relationships in the latent variable space.
• To enhance fault detectability and sensitivity analysis for neural network-based monitoring.

Main Methods:

• Development of the dynamic inner neural network based on a simplified self-attention mechanism (DiNN-SSAM).
• Mapping process variables into a latent variable space using neural networks to capture dynamic relationships.
• Fitting dynamic relationships using a simplified self-attention mechanism for efficient modeling.
• Introduction of fault detectability and sensitivity analyses for improved interpretability.

Main Results:

• DiNN-SSAM successfully establishes accurate dynamic models for industrial processes.
• The algorithm significantly reduces the complexity of online computations.
• The proposed fault detectability and sensitivity analyses enhance the interpretability of neural network models.
• Superior performance demonstrated through simulations on numerical examples, the Tennessee Eastman (TE) process, and catalytic cracking units.

Conclusions:

• DiNN-SSAM offers an effective solution for nonlinear dynamic process monitoring.
• The algorithm balances accuracy, computational efficiency, and interpretability.
• The novel interpretability analyses provide valuable insights into neural network-based monitoring systems.