Slow-Varying Dynamics-Assisted Temporal Capsule Network for Machinery Remaining Useful Life Estimation.

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Convolutional Neural Networks (CNNs) dominate Remaining Useful Life (RUL) estimation, but struggle with long-term temporal correlations.
• Capsule Networks (CapsNets) capture hierarchical relationships but also lack long-term temporal correlation analysis.
• Existing RUL models overlook slow-varying dynamics, limiting their ability to analyze low-frequency mechanical behavior.

Purpose of the Study:

• To propose a novel model, SD-TemCapsNet, for accurate RUL estimation.
• To simultaneously learn slow-varying and temporal dynamics from equipment measurements.
• To address limitations in current RUL estimation methods, particularly CapsNets.

Main Methods:

• Decomposition of slow-varying features from normal data to capture low-frequency system dynamics.
• Integration of Long Short-Term Memory (LSTM) mechanism within CapsNet to analyze time series temporal correlations.
• Development of a hybrid approach combining feature decomposition and LSTM-enhanced CapsNet.

Main Results:

• SD-TemCapsNet demonstrated superior performance on aircraft engine and milling machine datasets.
• Aircraft engine RUL estimation accuracy improved by up to 24.97% (RMSE) compared to CapsNet.
• Milling machine RUL estimation accuracy improved by 23.57% (vs. LSTM) and 19.54% (vs. CapsNet).

Conclusions:

• The proposed SD-TemCapsNet effectively estimates RUL by integrating slow-varying and temporal dynamics.
• The model overcomes limitations of traditional CapsNets and CNNs in analyzing complex equipment degradation.
• SD-TemCapsNet offers a significant advancement for predictive maintenance and machinery health monitoring.