As a critical transmission component in transmission systems, inter-shaft bearings face three major challenges in fault diagnosis: the incompleteness of fault information from single sensors, the scarcity of fault samples, and strong noise interference in complex environments. To address these challenges, a multi-sensor multi-scale convolutional attention network (MSCA-Net) was proposed based on multi-sensor data fusion, multi-scale feature extraction, and an embedded spatial feature attention mechanism. Gramian angular difference field (GADF), Gramian angular summation field (GASF), and Markov transition field (MTF) methods were employed to transform original one-dimensional vibration signals collected by multiple sensors into two-dimensional grayscale images. These three types of images were mapped to RGB channels respectively to construct composite color images, effectively preserving the spatiotemporal correlations among inter-shaft bearing vibration signals. A multi-scale channel attention feature fusion module (MSCAFM) was designed to simultaneously capture different fault feature information. A spatial feature attention mechanism was embedded to learn spatial attention for features by integrating information from both channel and spatial dimensions to highlight important regions of fault characteristics. Fault diagnosis of inter-shaft bearings was accomplished through feature extraction and classification layers. Experimental results show that on the dual-rotor inter-shaft bearing test bench dataset, with only 4 training samples per fault condition, the diagnostic accuracy reaches above 91.3%. Under -2 dB noise environment, the method maintains 78.5% accuracy, which demonstrates the robustness and practical value for inter-shaft bearing health monitoring under small-sample and high-noise conditions.