A Multi-Gradient-Descent-Integrated Physics-Informed Autoencoder for Sensor Fault Detection in Data Center Chillers
Open Access
Abstract
The operating status of chillers in data centers critically affects temperature stability and energy efficiency, imposing stringent requirements on the accuracy, robustness, and adaptability of sensor fault detection systems. Existing Physics-Informed Autoencoders (PIAE) for chiller fault detection face challenges such as gradient conflicts and limited generalization, especially under complex fault scenarios like sensor drift, stuck faults, and precision attenuation. To address these limitations, we propose a Multi-Gradient Descent Algorithm-integrated PIAE (MGDA-PIAE), which embeds the chiller thermal balance equation as a hard constraint and dynamically determines Pareto-optimal weights between data reconstruction and physical consistency. By unifying feature learning with physical law constraints, the model achieves improved generalization and stable performance across diverse operational conditions. Validation using operational data from a data center in Guangzhou demonstrates that MGDA-PIAE significantly outperforms conventional AE and PIAE models: the average recall increases by approximately 20% and the F1-score by 10%, while for flow sensor faults, the F1-score improves by over 80% compared with AE and 20% compared with PIAE. Further tests on multiple sensor types confirm that the model maintains high-precision detection with low false negatives under varying operating modes. By adjusting a single physical threshold, MGDA-PIAE can flexibly meet different fault detection requirements, providing a practical and reliable solution for maintaining the efficient, stable, and safe operation of data center refrigeration systems.Keywords
Funding Information
- National Natural Science Foundation of China (52508137)
This publication has 37 references indexed in Scilit:
- Contrastive self-supervised learning for lightweight and automated fault detection and diagnosis in HVAC systemsApplied Energy, 2026
- A physics-constrained deep learning framework for dynamic modeling of vapor compression systemsApplied Thermal Engineering, 2024
- A novel evaluation method of measurement sensitivities on common faults in VAV HVAC systemsBuilding and Environment, 2024
- A two-stage fault diagnosis strategy for air handling units via a backpropagation multidimensional Taylor network fitter and a novel statistical process controlApplied Thermal Engineering, 2024
- An adaptive fault detection model based on variational auto-encoders and unsupervised transfer learningApplied Soft Computing, 2024
- Sensor fault diagnosis and correction for data center cooling system using hybrid multi-label random Forest and Bayesian InferenceBuilding and Environment, 2023
- A thermodynamic-law-integrated deep learning method for high-dimensional sensor fault detection in diverse complex HVAC systemsApplied Energy, 2023
- An enhanced principal component analysis method with Savitzky–Golay filter and clustering algorithm for sensor fault detection and diagnosisApplied Energy, 2023
- Unsupervised outlier detection using neural network-based mixtures of probabilistic principal component analyzers for building chiller fault diagnosisBuilding and Environment, 2022
- Diagnosis for multiple faults of chiller using ELM-KNN model enhanced by multi-label learning and specific feature combinationsBuilding and Environment, 2022