Deep Neural Network-Embedded Stochastic Nonlinear State-Space Models and Their Applications to Process Monitoring.

Area of Science:

• Industrial Process Monitoring
• Machine Learning for Complex Systems
• Nonlinear Dynamic Modeling

Background:

• Complex industrial processes exhibit significant nonlinearities, dynamics, and uncertainties, challenging traditional monitoring techniques.
• Accurate modeling of nonlinear dynamic behavior is crucial for effective process monitoring.
• Existing methods like kernel mapping and stacked auto-encoders struggle with these complexities.

Purpose of the Study:

• To develop a robust process monitoring model capable of handling nonlinearities, dynamics, and uncertainties.
• To introduce a novel deep neural network (DNN) based stochastic nonlinear state-space model.
• To enable real-time anomaly detection in complex industrial processes.

Main Methods:

• Construction of a stochastic nonlinear state-space model using DNNs for state transition and observation generation.
• Development of a novel bidirectional recurrent neural network (RNN) for filtering and smoothing estimations.
• Offline learning of the smoothing estimator and process model, followed by fine-tuning of the filtering estimator for real-time monitoring.
• Design of two indices for process anomaly monitoring based on the learned model.

Main Results:

• The proposed DNN-based bidirectional RNN model effectively estimates process states for real-time monitoring.
• The model successfully addresses complex nonlinearities, dynamics, and uncertainties inherent in industrial processes.
• Case studies demonstrate that the proposed method outperforms comparative techniques by at least 10% in average fault detection rate.

Conclusions:

• The developed process monitoring model offers a superior approach for handling complex industrial process characteristics.
• The integration of DNNs and bidirectional RNNs provides a powerful framework for advanced process monitoring and anomaly detection.
• The method shows significant potential for improving safety and efficiency in industrial operations through enhanced fault detection.