As the demand for flexibility and efficiency in modern industrial equipment increases, motors often operate under variable speed conditions in real-world industrial applications. This poses challenges for traditional time-domain and frequency-domain fault diagnosis methods. These challenges arise primarily due to the non-linear and non-stationary characteristics of signals under variable speed conditions, which can affect fault feature extraction. Single deep learning models generally require training and test data to follow the same distribution, and domain adaptation or multi-source domain generalization methods are difficult to apply in the absence of target domain and multi-source domain data, limiting their ability to enhance the generalization of single-source domain models. To address these challenges, this study proposes a motor rolling bearing fault transfer diagnosis method that integrates angular domain resampling and feature enhancement.

First, to mitigate the issue of time-frequency characteristic offsets in vibration signals under different rotational speeds, angular domain resampling is employed. This technique processes vibration signals at varying speeds, obtaining angular domain vibration signals to minimize the offsets caused by speed changes. Second, to address the generalization limitations of deep learning models, fault data from constant speed conditions are used as the source domain for training the neural network. Covariance loss is introduced to amplify the feature differences among various classes in the source domain data. This allows the network to focus on more informative features for the classification task, thereby improving its generalization capability. Finally, the angular domain vibration signals under variable speed conditions are input into the trained model for fault classification.

The effectiveness of the proposed method is validated through several experiments. Initially, the time-frequency characteristics of vibration signals from an actual bearing inner ring fault are examined before and after angular domain resampling. Before resampling, the vibration signal intervals under variable speed conditions show significant variability. However, after resampling, the variability in the vibration intervals is significantly reduced. Furthermore, using t-SNE visualization, the study observes that networks without feature enhancement show slow gradient updates and minimal changes in feature distribution. In contrast, networks with feature enhancement exhibit continuous changes in feature distribution, even as the classification loss decreases, with increasing feature distances. The study also conducts four cross-working condition fault diagnosis experiments, comparing the proposed method with other methods. The results demonstrate that the proposed method improves fault identification accuracy by 35.04% compared to methods without angular domain resampling, especially in rolling element fault identification. When compared to methods without feature enhancement, the proposed method improves accuracy by 7.45%. Additionally, in transfer diagnosis tasks under different load conditions, the proposed method demonstrates high accuracy, recall, and F1 scores.

In conclusion, the study finds that: (1) Angular domain resampling effectively reduces time-frequency distribution differences caused by speed variations, proving its applicability and rationality in data preprocessing at different speeds. (2) The feature enhancement strategy, by increasing covariance loss between different class features, amplifies feature differences between various health status signals, enabling the network to capture more distinctive features and significantly improving generalization capability. (3) The proposed method, without requiring target domain data, achieves fault identification accuracy of up to 97.29% under variable speed conditions, demonstrating good robustness under variable load conditions.