Let the Model Choose Its Own Frequency: An Adaptive Frequency-Aware Inverted Transformer for Noise-Robust Gearbox...

Area of Science:

• Mechanical Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Industrial condition monitoring faces challenges in gearbox fault diagnosis due to noise.
• Standard deep learning models struggle with frequency relevance and adaptive noise suppression.
• Existing transformer methods lack cross-sensor dependency analysis, and traditional denoising is inflexible.

Purpose of the Study:

• To develop a lightweight transformer framework for adaptive gearbox fault diagnosis.
• To enable models to dynamically learn and focus on fault-relevant frequencies.
• To improve diagnostic accuracy under noisy operating conditions.

Main Methods:

• Proposed the Adaptive Frequency-Aware Inverted Transformer (AF-iTransformer) framework.
• Introduced a learnable spectral filter using FFT and a soft frequency mask for adaptive noise suppression.
• Implemented channel-level tokenization and cross-channel attention for heterogeneous sensor data.
• Utilized feature-wise linear modulation and residual attention propagation for enhanced training and stability.

Main Results:

• Achieved 99.63% accuracy on clean data and 95.1% at 0 dB SNR on the UConn Gearbox dataset.
• Demonstrated superior performance over baselines in noisy conditions.
• Attained 99.74% accuracy on the SEU gearbox dataset.
• Showcased efficient real-time deployment capabilities on edge GPUs and CPUs with low latency and memory usage.

Conclusions:

• AF-iTransformer effectively addresses the challenge of gearbox fault diagnosis in noisy environments.
• The adaptive frequency-aware approach significantly enhances diagnostic accuracy and robustness.
• The framework's efficiency confirms its suitability for real-time industrial applications.