<p>Rolling ball bearings are the essential components of rotating machines that can be affected by various faults, increasing the maintenance costs and downtime of the machine. The paper introduces a novel CNN-GRU prediction model, aimed to detect and diagnose incipient faults in the ball bearings components. The model is learned on nonlinear vibration signals that were recorded from a laboratory experimental setup performed on Machine Fault Simulator (MFS), focusing on defects in the inner race (IR), outer race (OR) and composite faults (CF). By leveraging deep features extracted through neural networks and employing transfer learning with Gated Recurrent Unit (GRU) network, the study enhances classification accuracy and effectively resolves the problem of vanishing gradient problem. The GRU has the advantage of efficiently extracting temporal features which have a long-term dependency in the time-series data, while the CNN captures both local and global spatial features. This combined approach not only enhances fault detection accuracy but also reduces the computational burden on the classifier. Overall, the result of this paper introduces a promising approach for both efficient and accurate fault detection and severity estimation in bearings of induction motors, potentially reducing the need for extensive manpower and sensor usage.</p>