Abstract: Fault diagnosis of rotating machinery is crucial for industrial safety, yet standard depthwise separable convolution in existing deep hybrid models is constrained by a fixed receptive field, making it difficult to capture multi-scale fault characteristics, ranging from local impulses and global trends. To address this, we propose a Multi-scale Folding Depthwise Separable Convolution network (MFDSC), which introduces folding into depthwise separable convolution to equivalently enlarge the receptive field without increasing the parameter count, and constructs a multi-branch parallel structure with different kernel sizes to extract multi-scale features simultaneously. Feature fusion is performed via 1×1 convolution, followed by Longformer and LSTM for deep temporal modeling, and fully connected layers produce the final diagnosis. Experimental results show that the proposed method achieves over 99% diagnostic accuracy on two datasets, and reaches an inference speed of 47 samples/s on edge devices. It also maintains over 98% accuracy under noise and cross-platform scenarios. These results demonstrate high diagnostic accuracy, lightweight deployment, and strong robustness.