<p>In the rolling bearing fault diagnosis process, there are more redundant spectra, which cannot be accurately converted, leading to insufficient feature extraction and a low correct fault diagnosis rate. Hence, a fault diagnosis approach for rolling bearings utilizing an attention-based residual deep network is proposed. Derived from Hertz contact theory, the rolling bearing structure is modeled by a simplified dynamic relationship. The rolling bearing fault vibration characteristics are dynamically analyzed, and the rolling bearing fault characteristic frequency is calculated. The non-redundant spectra in the abnormal vibration signals are converted into the corresponding Gram angle field and Gram Angle Sum Field (GASF), and the time-frequency feature matrix reflecting the vibration characteristics of rolling bearings is constructed. Innovatively, the attention mechanism is presented from the channel and space to improve the residual unit of the ResNet network, and the attention residual depth network is constructed to attain the rolling bearing fault diagnosis. The experimental findings indicate that the mean diagnostic accuracy for rolling bearing fault generation using the design method is 98.3%, the mean recall rate is up to 98.75%, and the cross-entropy loss convergence value of 0.121 is significantly lower than the random baseline, which belongs to a smaller loss. The above results indicated that the design approach has a more accurate diagnosis for various rolling bearing fault types.</p>