Proactive Fault Diagnosis of a Radiator: A Combination of Gaussian Mixture Model and LSTM Autoencoder

Area of Science:

• Mechanical Engineering
• Reliability Engineering
• Data Science

Background:

• Radiator reliability is critical in high-temperature and friction environments, necessitating early fault detection to prevent failures.
• Existing diagnostic methods may struggle with complex operational states and previously unencountered faults.

Purpose of the Study:

• To develop a proactive fault diagnosis system for radiators using advanced machine learning techniques.
• To enhance radiator reliability and enable effective life assessment in accelerated testing.

Main Methods:

• Collected radiator vibration signals from randomized durability vibration bench tests across normal, unknown, and faulty states.
• Extracted time-domain statistical features and applied Principal Component Analysis for data reduction.
• Integrated Gaussian Mixture Models for initial classification and Long-Short Term Memory autoencoders for advanced anomaly detection.

Main Results:

• The combined Gaussian Mixture Model and Long-Short Term Memory approach demonstrated effectiveness in radiator fault diagnosis.
• The methodology successfully captured time-dependent sequences for anomaly detection, including novel faults.
• Dynamic threshold adjustments were proposed for reliability life assessment using model distributions.

Conclusions:

• The integrated GMM-LSTM autoencoder model provides a robust and adaptable solution for radiator fault diagnosis.
• This proactive approach improves diagnostic capabilities, especially in demanding operational environments.
• The study lays groundwork for enhanced reliability life assessment through dynamic thresholding in accelerated life testing.