Study on a Fault Diagnosis Method for Heterogeneous Chiller Units Based on Transfer Learning

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Data Science

Background:

• Chiller units are critical for cooling systems, but acquiring sufficient labeled fault data is challenging.
• Data heterogeneity across devices and operating conditions hinders traditional data-driven fault diagnosis.
• Existing methods struggle with accurate fault identification due to data limitations.

Purpose of the Study:

• To develop a deep transfer learning method for accurate fault diagnosis in chiller units.
• To address challenges of limited labeled data and data heterogeneity in practical applications.
• To enable precise small-sample knowledge transfer from source to target domains.

Main Methods:

• A heterogeneous transfer learning approach integrating a dual-channel autoencoder, domain adversarial training, and pseudo-label self-training.
• Utilizing a Gradient Reversal Layer (GRL) and domain discriminator for domain-invariant feature extraction.
• Employing high-confidence pseudo-labeled samples from the target domain for joint training.

Main Results:

• The proposed method achieves high fault diagnosis accuracy in industrial scenarios.
• Effective identification of common faults in various chiller unit types under conventional operating conditions.
• Outperforms traditional approaches and existing transfer learning methods in multi-class fault diagnosis tasks.

Conclusions:

• The developed method enables precise small-sample knowledge transfer for chiller unit fault diagnosis.
• Offers a novel perspective for intelligent operation and maintenance of cooling systems.
• Demonstrates significant improvements in accuracy and F1-scores compared to existing techniques.