This article presents a new approach to intelligent and efficient fault diagnosis for bearings, which are important mechanical components that are widely used in modern industry. They are are among the most common causes of induction motor failure. Traditional approaches that exploit vibration signals, have several shortcomings, such as the number of intrusive vibration sensors, complex calculations, and limited learning capability. Motor current signature analysis uses a non-intrusive sensor to easily collect stator current signals from a power source. To rapidly process rotating machine failures and automatically provide an accurate diagnosis in the face of increasing condition data, conventional deep neural network models present limitations and inaccurate fault diagnosis results. To overcome this problem, a new intelligent deep learning architecture was proposed. To enhance the predictive capability of our deep neural network model (DNN), we set out to associate specific coefficients to each level and variation, and the results obtained were compared with other models such as support vector machines, k-nearest neighbors, decision tree, long short-term memory (LSTM), and convolutional neural network (CNN). Experimental tests for the early detection of bearing faults under different loads verify the effectiveness of the proposed approach, providing a fast and, reliable diagnosis capable of achieving a high diagnostic accuracy that is superior to existing methods.