Learning in the model space for cognitive fault diagnosis

Area of Science:

• Engineering
• Data Science
• Artificial Intelligence

Background:

• Large sensor networks generate real-time data for complex systems.
• Data can be incomplete, inconsistent, or from time-varying environments.
• Existing methods struggle with these data challenges.

Purpose of the Study:

• To develop an innovative cognitive fault diagnosis framework.
• To address challenges of incomplete/inconsistent data and unformulated environments.
• To enable fault diagnosis in the model space.

Main Methods:

• Fitting models to signal segments using a sliding window.
• Employing one-class learning algorithms to discriminate faulty from healthy models.
• Theoretically investigating pairwise model distance and incorporating it into learning.

Main Results:

• Effectively discriminated faulty models from healthy ones.
• Constructed a fault library for unknown faults (cognitive fault isolation).
• Demonstrated framework effectiveness on benchmark applications and a water distribution network.

Conclusions:

• The cognitive fault diagnosis framework operates effectively in the model space.
• The approach handles incomplete data and unformulated environments.
• The method enhances fault diagnosis and isolation capabilities for complex systems.