Faultformer: A CNN–Transformer Multi-Domain Feature Fusion Network for Intelligent Bearing Fault Diagnosis

Why It Matters

Such near‑perfect diagnostic performance enables predictive maintenance, reducing unplanned downtime and maintenance costs for manufacturers.

Key Takeaways

• Hybrid CNN‑Transformer captures both local transients and long‑range dependencies
• Multi‑domain fusion integrates statistical, spectral and deep features for richer representation
• Achieves 99.98 % accuracy on CWRU dataset, surpassing state‑of‑the‑art models
• Enables reliable, repeatable bearing fault diagnosis for predictive maintenance programs

Pulse Analysis

Rolling‑bearing failures remain a leading cause of unexpected equipment shutdowns in manufacturing, prompting a shift from manual vibration analysis toward data‑driven diagnostics. Traditional approaches rely on handcrafted statistical or spectral features, which can miss subtle, time‑varying patterns. Recent advances in deep learning, particularly convolutional neural networks, improve local pattern recognition but often struggle with long‑range temporal dependencies that characterize periodic mechanical faults. The industry therefore seeks hybrid models that combine the strengths of both paradigms to achieve higher reliability.

FaultFormer addresses this gap by integrating multi‑domain feature fusion with a CNN‑Transformer backbone. Handcrafted statistical descriptors and frequency‑domain metrics are concatenated with deep features extracted by convolutional layers that capture transient spikes. A Transformer encoder then applies multi‑head self‑attention to model long‑range temporal relationships across the vibration signal, enabling the network to distinguish between similar fault signatures under varying loads. Validated on the widely used Case Western Reserve University bearing dataset, the system achieved a mean accuracy of 99.98 % across ten fault categories, outperforming contemporary hybrid and pure deep‑learning models while maintaining repeatable performance over multiple runs.

The implications for industrial stakeholders are significant. Near‑perfect fault classification supports condition‑based maintenance strategies, allowing plants to schedule interventions before catastrophic failures occur, thereby cutting downtime and maintenance expenditures. Moreover, the modular nature of FaultFormer’s feature fusion makes it adaptable to other rotating‑machinery domains, such as gearboxes or turbines. As manufacturers increasingly digitize operations, deploying such high‑precision diagnostic tools can accelerate the transition to smart factories, while ongoing research may further enhance scalability and real‑time inference capabilities.