Research on Vibration Signal Processing of Rotating Components Based on Image Reconstruction Models

Aiming at the non-stationary characteristics of vibration signals in rotating components and the limitations of traditional time-frequency analysis methods, such as poor noise resistance and low energy concentration, this paper proposes a deep learning-based time-frequency image reconstruction model. By integrating skip-connections in a fully convolutional network with a channel attention mechanism (SE block), an encoderdecoder architecture with feature enhancement capability is constructed, along with a joint optimization function incorporating edge loss, pixel loss, and focal loss. Instance normalization and a regularized sparsity strategy are introduced to improve the model's ability to capture high-frequency components, while early stopping is employed to prevent overfitting. Experiments on multi-component simulated signals demonstrate that the proposed method (rTFAE) effectively enhances time-frequency energy concentration, as measured by kurtosis values, compared to STFT and SET methods within the 1-20 dB signal-to-noise ratio range. The reconstructed timefrequency images exhibit significantly smoother trajectories and superior noise robustness over conventional approaches.