Fastformer: An Efficient Attention-Based Framework for Rapid Multi-Class Fault Diagnosis in High-End Equipment...

Area of Science:

• Mechanical Engineering
• Signal Processing
• Machine Learning

Background:

• Accurate multi-class fault diagnosis is critical for high-end equipment maintenance.
• Non-stationary vibration signals present challenges for reliable fault identification due to noise and complex dynamics.
• Existing methods struggle with redundant computation and distinguishing between fault types.

Purpose of the Study:

• To develop a lightweight and discriminative diagnostic framework for multi-class fault identification from vibration signals.
• To address the challenges of non-stationary signals, noise, and computational redundancy in fault diagnosis.
• To improve the separation of informative fault modes and enhance inter-class distinctions.

Main Methods:

• Fastformer framework integrating Empirical Mode Decomposition (EMD) for signal decomposition.
• Encoder-oriented Q/K/V dot-product scoring for compact spatiotemporal embeddings, avoiding full Transformer architecture.
• Validation-guided pruning of attention responses and a Margin-Enhanced Fault Softmax classifier for improved category separation.

Main Results:

• Fastformer achieved perfect scores (1.000) for precision, recall, F1-score, and AUC on the XJTU-SpurGear dataset.
• Demonstrated superior overall performance on the HUST bearing dataset with an AUC of 0.9596.
• Showcased faster and more stable convergence compared to other methods.

Conclusions:

• Fastformer offers an effective solution for lightweight and accurate multi-class fault diagnosis from vibration signals.
• The framework successfully handles non-stationary signals and reduces computational load.
• Combines stable decomposition, efficient attention mechanisms, and margin-based classification for robust fault identification.