Neural networks and fault probability evaluation for diagnosis issues

Area of Science:

• Control Engineering
• Artificial Intelligence
• System Identification

Background:

• Fault detection and isolation (FDI) is critical for system reliability and safety.
• Unknown nonlinear systems pose significant challenges for traditional FDI methods.
• Existing FDI techniques often lack robustness and accurate fault localization capabilities.

Purpose of the Study:

• To develop a new FDI technique for fault detection and isolation in unknown nonlinear systems.
• To leverage artificial intelligence and probabilistic methods for residual generation and analysis.
• To quantitatively assess the performance and reliability of the proposed FDI approach.

Main Methods:

• Utilizing artificial neural networks for modeling fault-free and faulty system behaviors.
• Generating residuals based on learned system models.
• Applying probabilistic criteria for fault evaluation and candidate fault identification.
• Calculating a confidence factor to assess FDI decision reliability.

Main Results:

• The proposed FDI technique successfully detected and isolated 19 fault candidates in the DAMADICS benchmark.
• Quantitative indicators were established to assess the performance of the AI and probabilistic tools.
• The method demonstrated suitability for evaluating FDI decision reliability through confidence factor computation.
• Comparison with a standard thresholding method highlighted the proposed scheme's effectiveness.

Conclusions:

• The integration of artificial neural networks and probabilistic methods offers a robust solution for FDI in unknown nonlinear systems.
• The developed technique provides reliable fault detection and isolation with a quantifiable measure of confidence.
• The study establishes a benchmark for evaluating FDI performance using quantitative indicators and a challenging industrial application.