Contrastive learning and dynamics embedding neural network for label-free interpretable machine fault diagnosis

Area of Science:

• Mechanical Engineering
• Data Science
• Machine Learning

Background:

• Industrial machinery generates vibration signals indicative of faults.
• Accurate fault diagnosis is challenged by the need for labeled vibration data.
• Existing supervised learning methods struggle with distinguishing vibration-related faults.

Purpose of the Study:

• To develop an innovative, interpretable approach for automatic fault diagnosis.
• To address the challenge of unlabelled vibrational signals in industrial machinery.
• To introduce the Bootstrap Your Own Latent and Dynamical Systems Model Discovery (BYOLDIS) algorithm.

Main Methods:

• Deriving differential equations to model faulty bearing dynamics.
• Utilizing contrastive learning and time-delay embedding for system reconstruction.
• Constructing a library of fault machine dynamic polynomial equations with physical constraints.

Main Results:

• BYOLDIS accurately diagnoses bearing faults in simulations and experiments.
• The method provides dynamic explanations for diagnostic outcomes.
• Demonstrated effectiveness in processing unlabelled vibrational data.

Conclusions:

• BYOLDIS offers a robust solution for interpretable, automatic fault diagnosis.
• The algorithm successfully handles unlabelled vibration data from industrial machinery.
• BYOLDIS shows significant promise for advancing machine health monitoring and diagnostics.