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

Abstract

In this paper, FaultFormer is proposed as a new hybrid intelligent mechanism for automatically diagnosing rolling bearing faults using vibration signals from an induction motor. It combines multi-domain feature fusion based on handcrafted statistical features, frequency-domain analysis, and deep features learned through a hybrid convolutional neural networks (CNN)-transformer-based approach. Whereas CNN layers capture local temporal vibration patterns, the transformer encoder leverages multi-head self-attention to capture long-range temporal dependencies that are important for identifying periodic mechanical faults. The framework was validated using the Case Western Reserve University bearing dataset, which covers 10 fault categories across different load modes and defect severities. Experimental results show that the model achieves 99.98% classification mean accuracy across 20 runs, with a precision of 99.68%, a recall of 99.68%, and an F1-score of 99.80%, outperforming some of the newly introduced state-of-the-art deep learning and hybrid models. This integrated approach to feature fusion successfully represents global signal patterns with localized transient features, providing a robust, repeatable approach to the analysis of industrial machine condition.