Fault Diagnosis of Motor Bearings Based on Multi-dimensional Vibration Feature Fusion

To address the issues of limited and unstable unidirectional vibration data features, as well as poor training and transferability in motor bearing fault diagnosis, this paper proposes a fault diagnosis method based on multi-dimensional vibration time-frequency feature fusion. The method synchronously collects vibration data along the x, y, and z axes of the bearing using three-axis vibration sensors under different working conditions. Continuous Wavelet Transform (CWT) is applied to extract time-frequency features, generating 2D grayscale images. Based on the channel characteristics of RGB images, the time-frequency grayscale images of the three directions are fused to generate an RGB image, where the feature fusion and enhancement are achieved through the combination of brightness and color. A deep learning model, ConvNeXt, is then used to learn the fused features for cross-condition diagnosis. The experimental results show that, compared with methods using unidirectional vibration data, the proposed method achieves improvements in training accuracy, training loss, and cross-condition recall rate. Moreover, it demonstrates significant advantages even in noisy data experiments. In comparison with weighted average fusion and PCA fusion methods, the proposed method performs better in feature fusion. Therefore, the proposed method effectively improves the accuracy and transferability of motor bearing fault diagnosis.