DyGAT-FTNet: A Dynamic Graph Attention Network for Multi-Sensor Fault Diagnosis and Time-Frequency Data Fusion.

Area of Science:

• Industrial Fault Diagnosis
• Machine Learning
• Sensor Networks

Background:

• Traditional fault diagnosis methods struggle with complex spatiotemporal dependencies in multi-sensor data.
• Dynamic features between sensors are often overlooked, limiting diagnostic accuracy.
• Ensuring equipment reliability and operational safety necessitates advanced fault detection techniques.

Purpose of the Study:

• To propose DyGAT-FTNet, a novel graph neural network model for enhanced multi-sensor fault detection.
• To address limitations in capturing spatiotemporal dependencies and dynamic features in fault diagnosis.
• To improve the accuracy and reliability of fault detection in industrial systems.

Main Methods:

• DyGAT-FTNet dynamically constructs association graphs using a learnable mechanism based on time-frequency features from Short-Time Fourier Transform (STFT).
• A Dynamic Graph Attention Network (DyGAT) is employed to extract spatiotemporal dependencies by adaptively weighting nodes.
• A time-frequency graph pooling layer is utilized for aggregating information and optimizing feature representation.

Main Results:

• DyGAT-FTNet achieved superior classification accuracy on benchmark datasets (XJTUSuprgear and SEU), reaching 1.0000 and 0.9995, respectively.
• The model significantly outperformed existing methods in multi-sensor fault detection tasks.
• Experimental results demonstrate the effectiveness of the dynamic graph construction and attention mechanisms.

Conclusions:

• DyGAT-FTNet offers a powerful and accurate solution for multi-sensor fault detection.
• The proposed model effectively captures complex spatiotemporal and dynamic features for improved fault diagnosis.
• The high accuracy achieved highlights the practical applicability of DyGAT-FTNet in industrial settings.