In the realm of heavy machine operations, the necessity for efficient fault diagnosis methodologies is evident, given the substantial risks posed by breakdowns in terms of costs, manpower resources, and time. This study leverages VGG16, guided Grad-CAM, and LRP techniques for fault diagnosis in three-phase induction motors and transformers. Departing from traditional methods, this research strategically incorporates non-destructive tech- niques such as thermal imaging. The methodology entails training a model on a thermal image dataset using the pre- trained VGG16 architecture. Subsequently, guided image datasets are generated through the Guided Grad-CAM process, complemented by corresponding LRP data, form- ing the foundation for training various CNN architectures with a softmax classifier for fault categorization. Signifi- cantly, this integrated approach enables the identification of potential fault areas based on thermal gradients. Pre- liminary results demonstrate promising outcomes, under- scoring the efficacy of this methodology in fault diagnosis. Moreover, extending beyond machinery applications, this study advocates for the integration of neural networks and thermal imaging in industries, offering prospects for predictive maintenance and enhanced operational efficiency in industrial diagnostics.