Adaptive Partially Observed Sequential Change Detection and Isolation.

Area of Science:

• Industrial applications
• Data science
• Machine learning

Background:

• High-dimensional data is prevalent in industrial settings due to sensor accessibility.
• Challenges include incomplete measurements from resource constraints and identifying distinct system failure patterns.
• Online adaptive monitoring is crucial for systems with multiple potential failure modes.

Purpose of the Study:

• To develop an online adaptive monitoring strategy for high-dimensional data in resource-constrained environments.
• To effectively identify true failure patterns amidst multiple potential modes.
• To enhance change point detection and failure mode isolation accuracy.

Main Methods:

• Application of the Shiryaev-Roberts procedure at the failure mode level.
• Utilization of the multi-arm bandit algorithm to balance exploration and exploitation.
• Theoretical analysis of the algorithm's properties for failure mode isolation.

Main Results:

• The proposed algorithm demonstrates the capability to correctly isolate failure modes.
• Significant improvements in change point detection performance were observed.
• Enhanced accuracy in identifying the true failure mode in complex systems.

Conclusions:

• The developed method offers an effective solution for online adaptive monitoring of high-dimensional data under resource constraints.
• The integration of Shiryaev-Roberts procedure and multi-arm bandits provides robust failure mode identification.
• The approach significantly advances the state-of-the-art in industrial system monitoring and fault diagnosis.