The working environment of vibrating screens is complex and variable, and the distribution of vibration data is easily disturbed and changes, leading to a decline in the generalization ability of the model. Additionally, the occurrence frequencies of different fault types vary, resulting in differences in the number of samples between fault categories. To address above issues, a cross-domain fault diagnosis method for vibrating screens with class imbalance, combining cloud oversampling and domain adaptation geometric constructive network, is proposed. Firstly, by integrating feature extraction methods from signal processing and statistical analysis, a joint analysis of time domain, frequency domain, and time-frequency domain is conducted to extract key statistical features that comprehensively reflect the dynamic characteristics of vibration signals. Secondly, to balance the number of samples of each class in the dataset, the uncertainty transformation mechanism of the cloud model is introduced to enhance minority class samples in the cloud feature space. Finally, the maximum mean and covariance difference is introduced to reduce the global feature distribution difference, constructing a domain adaptation geometric constructive network fault diagnosis model, thereby enhancing the model’s generalization ability. Taking the actual industrial cross-domain data of vibrating screen as an example, the results show that, compared with cross-domain and class-imbalanced fault diagnosis methods such as convolutional neural networks （CNN）, random vector functional link neural network （RVFLNN）, and active broad transfer for class-imbalanced learning （ABTCI）, the average fault diagnosis accuracy of this method is above 0.93. Furthermore, its classification accuracy in various diagnostic tasks is also higher than these methods, further demonstrating the effectiveness and superior generalization ability of this method.