<p>With the current advancement of fault detection and diagnosis technology, a real-time fault diagnosis system is essential for immediate response to the machinery’s health control. However, the current deep learning models face challenges in the implementation for real-time monitoring conditions. The unreliable evaluation of deep learning models in real-time conditions leads to difficulties in the intelligent real-time fault diagnosis application, as it is hard to define the responsiveness of model based on the training time. Extensively, the Convolutional Neural Network (CNN), widely applied in the mechanical fault diagnosis system, has a limitation when applied with single convolutional kernels. Hence, the dual-path three-folded residual network (DPTTRCNN) is proposed. The parallel path with a different scale enables the different dimensional features. Secondly, the introduced three-folded residual block (TfRb) with the application of layer normalisation can effectively tackle the gradient degradation problem by preserving the information across the deep architecture of models. The feasibility of the proposed model is evaluated based on the online dataset of the CWRU bearing dataset, the NCRA bearing dataset, and the experiment dataset by comparing it to existing models. The proposed model achieves an average accuracy of 99% or above, indicating its superior performance in fault diagnosis. The real-time performance of deep learning models is also discussed intensively based on the prediction time per second of data segment in simulation conditions, with the current limitations and future implementation.</p>