Data-Driven Designs of Fault Detection Systems via Neural Network-Aided Learning

Area of Science:

• Control Engineering
• Artificial Intelligence
• System Dynamics

Background:

• Dynamic systems require robust fault detection (FD) methods for reliable operation.
• Traditional model-based FD can be limited by system uncertainties and complexity.
• Neural networks offer a data-driven alternative for enhancing FD capabilities.

Purpose of the Study:

• To develop novel data-driven fault detection designs for dynamic systems using neural networks.
• To generate residual signals using two distinct neural network architectures: finite impulse response (FIR) and recursive.
• To establish theoretical links between model-based and neural network-based FD methodologies.

Main Methods:

• Utilizing neural networks for data-driven residual signal generation.
• Implementing a finite impulse response (FIR) filter-based neural network design.
• Developing a recursive neural network design for residual signal generation.
• Employing self-organizing learning for optimal neural network architecture determination.
• Conducting theoretical analysis to bridge model-based and neural network-based FD.

Main Results:

• Two effective neural network-aided fault detection algorithms were developed.
• The proposed neural networks demonstrated the ability to find optimal architectures for FD.
• Theoretical connections were established between model-based and neural network-based FD methods.
• The effectiveness of the proposed algorithms was validated through experiments on a three-tank system.

Conclusions:

• The developed neural network designs provide effective data-driven solutions for fault detection in dynamic systems.
• The study successfully linked model-based and neural network-based fault detection paradigms.
• The proposed methods offer a promising approach for enhancing the reliability and performance of dynamic systems.