A Multi-Gradient-Descent-Integrated Physics-Informed Autoencoder for Sensor Fault Detection in Data Center Chillers

Area of Science:

• Data Center Operations
• Artificial Intelligence in Engineering
• Thermal Management Systems

Background:

• Chiller performance is crucial for data center temperature stability and energy efficiency.
• Existing sensor fault detection methods, like Physics-Informed Autoencoders (PIAE), struggle with complex faults and generalization.
• Challenges include gradient conflicts and limited adaptability to sensor drift, stuck faults, and precision attenuation.

Purpose of the Study:

• To develop an advanced sensor fault detection system for data center chillers.
• To overcome the limitations of existing PIAE models in handling complex fault scenarios.
• To improve the accuracy, robustness, and adaptability of fault detection for enhanced operational stability.

Main Methods:

• Proposed a novel Multi-Gradient Descent Algorithm-integrated Physics-Informed Autoencoder (MGDA-PIAE).
• Embedded the chiller thermal balance equation as a hard constraint within the model.
• Dynamically determined Pareto-optimal weights balancing data reconstruction and physical consistency.

Main Results:

• MGDA-PIAE demonstrated significantly improved performance over conventional Autoencoders (AE) and PIAE models.
• Average recall increased by ~20%, and F1-score by ~10% in validation tests.
• For flow sensor faults, the F1-score improved by over 80% compared to AE and 20% compared to PIAE.

Conclusions:

• The MGDA-PIAE model offers superior generalization and stable performance across diverse operational conditions.
• It achieves high-precision fault detection with low false negatives, adaptable to various sensor types and operating modes.
• The system provides a practical, reliable solution for maintaining efficient, stable, and safe data center refrigeration operations.