Trend-Conditioned Residual Learning for Early Fault Warning in Nonstationary Multi-Sensor Oil Monitoring.

Area of Science:

• Mechanical Engineering
• Condition Monitoring
• Data Science

Background:

• Industrial gas turbine lubrication oil monitoring is crucial for early fault detection and maintenance.
• Nonstationary thermodynamic drifts in sensor data often mask subtle degradation signatures.
• Existing models struggle with complex residual distributions and trend separation.

Purpose of the Study:

• To develop a novel framework, ResAD-Net, for enhanced early fault warning in nonstationary multi-sensor oil monitoring.
• To address limitations in separating trends from residuals and modeling complex residual distributions.
• To improve the accuracy and reliability of incipient degradation detection in industrial gas turbines.

Main Methods:

• Implemented a signal trend-residual decoupling strategy to isolate degradation-sensitive fluctuations.
• Utilized a trend-conditioned diffusion model for characterizing state-dependent residual distributions.
• Employed a graph-based variational autoencoder for learning inter-sensor dependencies in the residual domain.

Main Results:

• ResAD-Net achieved a high average F1-score of 0.985 on real-world industrial oil monitoring data.
• The framework demonstrated no false positives within the predefined pre-alarm interval.
• Early residual distributional shifts were detected before macroscopic deviations, providing diagnostic cues.

Conclusions:

• ResAD-Net offers a robust solution for early fault warning in challenging nonstationary oil monitoring environments.
• The proposed methods effectively handle complex data characteristics, improving diagnostic accuracy.
• The framework provides valuable insights for interpreting oil monitoring changes and guiding maintenance decisions.