Bearings are widely used in equipment such as robots. Therefore, timely diagnosis of their faults is the necessary basis for maintaining production safety. With the promotion of intelligence technology, deep learning has become an important tool for the intelligent diagnosis of bearing faults and has received much research attention. However, existing research models usually require many failure samples to achieve preset performance. In actual production, because of the rugged operating environment of bearings and the short operating time under abnormal conditions, failure data is missing. To address this problem, an intelligent fault diagnosis model called the multi-scale convolutional kernels attention network (MSCKAN) is proposed in this context. The model is grounded in a two-dimensional convolutional neural network (2D-CNN), which obtains input through the reconstruction of time-domain samples. Then, it combines multi-scale convolution kernels and the efficient channel attention (ECA) mechanism to increase its feature extraction and recognition performance for small samples, thereby reducing the need for fault data. The performance of the MSCKAN was tested through the constructed small-sample experimental datasets and compared with that of other models. The testing results strongly prove that the MSCKAN has superior performance in small-sample diagnosis for bearing faults.