Decentralized adaptively weighted stacked autoencoder-based incipient fault detection for nonlinear industrial...

Area of Science:

• Industrial Process Monitoring
• Machine Learning for Fault Detection

Background:

• Modern industrial processes are large-scale and nonlinear.
• Detecting incipient faults is challenging due to faint signatures.

Purpose of the Study:

• To propose a decentralized adaptively weighted stacked autoencoder (DAWSAE) method for improved incipient fault detection.
• To enhance fault detection performance in large-scale nonlinear industrial processes.

Main Methods:

• Dividing the industrial process into sub-blocks for local analysis.
• Establishing local and global adaptively weighted stacked autoencoders (AWSAE).
• Constructing local and global statistics from feature and residual vectors for detection.

Main Results:

• The DAWSAE method effectively mines local and global information.
• Local and global statistics enable detection of sub-blocks and the whole process.
• Validation on a numerical example and the Tennessee Eastman Process (TEP).

Conclusions:

• The proposed DAWSAE method offers a robust solution for incipient fault detection.
• The decentralized approach effectively handles large-scale nonlinear industrial systems.
• The method demonstrates superior performance in identifying subtle fault signatures.