An Integrated Framework of Fourier Transform and Transformer for Rotating Machinery Fault Diagnosis

CFF-Trans leverages the power of frequency-domain analysis, extracting salient features from vibration signals that are often obscured in the time domain. This unique approach enables the model to successfully differentiate between noise and fault-related components, thereby facilitating accurate fault identification even amidst high-noise environments. This is of particular importance in real-world industrial settings where machinery often operates under noisy conditions.The architecture of CFF-Trans is thoughtfully designed to optimize feature extraction and signal interpretation. It employs convolutional layers for initial feature extraction, capturing preliminary patterns in the input data. Fourier transforms are then utilized for detailed frequency-domain encoding, effectively translating the time-based signal patterns into a spectral representation. This translation unveils the hidden characteristics of the signal, providing a new perspective for fault diagnosis.The final component of the CFF-Trans framework is the integration of Transformer encoders for global dependency mapping. This feature enables the model to understand the interdependencies between different parts of the signal, enhancing its ability to decipher complex signal patterns and interactions.Diagnosing faults in rotating machinery is crucial for maintaining operational efficiency and safety. However, existing diagnostic methods, while sometimes effective, often struggle with the complexity of data and noise interference, creating a need for more advanced solutions. In light of this, this paper proposes CFF-Trans, a novel deep learning framework that combines Fourier and Transformer-based techniques for enhanced fault diagnosis in rotating machinery. CFF-Trans excels in extracting frequency domain features from vibration signals, enabling it to effectively differentiate between noise and fault-related components and accurately identify faults even in high-noise environments. Specifically, CFF-Trans integrates convolutional layers for initial feature extraction, Fourier transforms for detailed frequency domain encoding, and Transformer encoders for global dependency mapping. This approach not only improves the noise robustness of the diagnostic model but also enhances its capability to extract relevant features from complex vibration signals. Validated on a dataset simulating various faults in rotating machinery under different load conditions. The experimental results clearly demonstrate the superiority of CFF-Trans over existing models in terms of accuracy and F1 scores across different noise levels, demonstrating its promising practical value in this field.