In practical industrial processes, the limitation of fault data collection leads to a decline in feature extraction capability and causes issues of incomplete data interpretation when using a single-channel approach. To address this problem, this paper proposes a diagnostic framework based on image fusion and a dual-channel convolutional neural network. First, the MTF-GADF-GASF image conversion method is employed to transform time-series signals into three types of images, which are then encoded into the red, green, and blue (RGB) channels of an image for fusion. Subsequently, one-dimensional signals and two-dimensional images are simultaneously fed into a parallel dual-channel model, where the first channel utilizes a Convolutional Neural Network (CNN) to extract spatial information, while the second channel employs Gated Recurrent Units (GRU) to capture temporal features from vibration signals. Finally, the spatiotemporal features extracted from both channels are fused and trained using a multi-head self-attention mechanism. Experimental results based on public and real-measured datasets indicate that the proposed diagnostic method achieves an average accuracy of 98.75% in fault type diagnosis under various operating conditions. Comparisons with other methods further demonstrate that this approach offers higher accuracy and robustness.