As industrial systems advance towards automation, the demand for accurate fault diagnosis methods grows. Deep learning (DL) faces challenges such as parameter volume and initialization instability. To overcome these hurdles, we proposed a DL framework integrating CNNs and TL. Signals are converted into 2-D images by Continuous wavelet transformation (CWT) improving representation. SVMs replace fully connected layers, enhancing efficiency. The method outperforms classical DL architectures. Validation includes convergence curves, testing reports and confusion matrices. It achieves highest accuracy in cross-domain diagnosis and fault detection. Across mechanical datasets, it sets a new standard for performance.