Explainable Fault Classification and Severity Diagnosis in Rotating Machinery using Kolmogorov-Arnold Networks

In this open-access paper, available here, we introduce an explainable, efficient and adaptive framework based on Kolmogorov-Arnold Networks (KANs) for fault detection, classification and severity diagnosis in rotating machinery. The framework utilizes automatic feature selection and symbolic regression techniques to create interpretable models suitable for real-time condition monitoring.

What Are Kolmogorov-Arnold Networks (KANs)? Link to heading

KANs are neural network architectures inspired by the Kolmogorov-Arnold representation theorem, distinguished by their use of trainable symbolic activation functions. This design makes KANs highly interpretable, overcoming the common “black-box” limitations of traditional neural networks.

Main Contributions Link to heading

• Automatic and Explainable Feature Selection: We employ sparsity-inducing regularization coupled with an explicit feature attribution mechanism, enabling our framework to automatically select the most informative features from extensive sensor data.

• Unified Fault Diagnosis Framework: Our approach handles multiple tasks seamlessly; fault detection, detailed fault classification (e.g., bearing faults, imbalance, misalignment), and severity estimation, all within a consistent, unified methodology.

• Interpretable Symbolic Models: Activation functions trained in KANs are approximated by symbolic mathematical expressions, significantly enhancing the interpretability of our models.

• High Performance and Real-world Applicability: Tested extensively on the widely-recognized CWRU and MaFaulDa datasets, our models consistently achieved perfect or near-perfect F1-scores, demonstrating adaptability across diverse fault scenarios and highlighting potential applications in industrial monitoring systems.

Framework Overview Link to heading

Our methodology comprises distinct stages, including data augmentation, automated feature extraction and selection, adaptive hyper-parameter tuning and symbolic regression, ensuring both interpretability and predictive performance.

Below is a visualization of the complete KAN-based fault diagnosis framework, illustrating how the different phases integrate into an interpretable and efficient pipeline:

@article{e27040403,
      author = {Rigas, Spyros and Papachristou, Michalis and Sotiropoulos, Ioannis and Alexandridis, Georgios},
      journal = {Entropy}, 
      title = {Explainable Fault Classification and Severity Diagnosis in Rotating Machinery Using Kolmogorov–Arnold Networks}, 
      year = {2025},
      volume = {27},
      number = {4},
	  article-number = {403},
      doi = {10.3390/e27040403}
}