A Hybrid Neural–Symbolic Framework for Interpretable Fault Diagnosis in Rotating Machinery

Abstract

The advent of Industry 4.0 has precipitated a paradigm shift in the maintenance strategies of complex industrial systems, particularly rotating machinery, moving from reactive to predictive maintenance regimes. While deep learning models have achieved state-of-the-art performance in fault diagnosis due to their powerful feature extraction capabilities, they suffer from a critical lack of transparency, often described as the black-box problem. This opacity hinders their adoption in safety-critical environments where understanding the etiology of a fault is as significant as its detection. This paper proposes a novel Hybrid Neural–Symbolic Framework (HNSF) that integrates the perceptual capabilities of deep neural networks with the inferential transparency of symbolic logic. We utilize a One-Dimensional Convolutional Neural Network (1D-CNN) to extract high-level latent features from raw vibration signals, which are subsequently mapped to semantic concepts within a predefined knowledge graph. A differentiable reasoning layer then applies First-Order Logic rules to these concepts to deduce fault classes, ensuring that the model's predictions are consistent with domain knowledge. Experimental validation on the Case Western Reserve University (CWRU) bearing dataset demonstrates that our framework not only achieves classification accuracy comparable to pure deep learning models but also provides human-readable explanations for its decisions. The results suggest that bridging the subsymbolic and symbolic gap offers a robust pathway toward trustworthy industrial artificial intelligence.