Conventional eigenvalue alarm methods for fault detection in low-speed, heavy-load, ultra-large rolling bearings often result in high rates of misdiagnosis and missed diagnosis. By contrast, deep learning techniques typically require large datasets, but the signal characteristics of bearings change slowly and are challenging to capture. To address these challenges, an intelligent fault diagnosis method based on a one-dimensional convolutional neural network (1DCNN) that incorporates a channel attention mechanism (CAM), is proposed. First, a signal acquisition scheme that combines low-frequency long sampling with charge amplifiers is designed to collect fault characteristic signals from low-speed, heavy-load rolling bearings. The collected dataset is then processed through a combination of fixed-length random segmentation and overlapping sampling to ensure sample diversity while minimizing feature loss from fixed segmentation. This approach effectively expands the training data volume under small-sample conditions, allowing for the construction of training, validation, and test sets. Next, the CAM module is integrated into the 1DCNN to enhance the model’s ability to learn essential features. The resulting CAM-1DCNN model is trained via the training data, with backpropagation employed to optimize network parameters at each layer, ultimately yielding a robust CAM-based 1DCNN fault diagnosis model. Finally, parameter tuning is conducted on the validation set to finalize fault classification. Experimental results from tests on a shield main bearing test bench demonstrate that the proposed method achieves a diagnostic accuracy of 97.13%, significantly outperforming traditional 1DCNN models in both performance and stability under small-sample conditions. The method effectively reduces deep learning’s reliance on large datasets, offering a practical solution for fault diagnosis with this type of bearing.