A life prediction method based on MDFF and DITCN-ABiGRU mixed network model.

Area of Science:

• Mechanical Engineering
• Signal Processing
• Artificial Intelligence

Background:

• Single network models struggle with accurate life prediction for rotating machinery due to diverse fault types and unpredictable occurrences.
• Effective prognostics require advanced methods capable of handling complex signal features and temporal dependencies.

Purpose of the Study:

• To develop a more accurate and robust life prediction model for rotating machinery.
• To address the limitations of existing models in handling various fault types and uncertain fault occurrences.

Main Methods:

• Multi-domain feature fusion (MDFF) of vibration signals from time, frequency, and time-frequency domains.
• Dimensionality reduction optimization of extracted multi-domain features.
• A distributed TCN-Attention-BiGRU (DITCN-ABiGRU) network integrating Temporal Convolutional Networks (TCN) with an attention mechanism and Bidirectional Gated Recurrent Units (BiGRU).
• Construction of a health indicator (HI) curve for precise lifespan prediction.

Main Results:

• The proposed MDFF and DITCN-ABiGRU model demonstrated superior performance compared to Convolutional Neural Networks (CNN) and Gated Recurrent Units (GRU) models.
• Achieved a better score and lower prediction error in life prediction tasks.
• Validated on rolling bearings (IEEE PHM Challenge 2012 dataset) and ball screw pairs.

Conclusions:

• The MDFF and DITCN-ABiGRU model offers a significant advancement in rotating machinery prognostics.
• The integrated approach effectively captures critical fault information and temporal dependencies for accurate life prediction.
• This method provides a more reliable solution for predicting the remaining useful life of rotating machinery.