Industrial gearboxes are critical components in various mechanical systems, but their reliability can be severely affected by faults, leading to costly downtimes and safety hazards. Traditional fault detection methods, such as manual inspections and vibration analysis, often fall short in identifying faults in their early stages and are sensitive to changes in operating conditions. This paper presents a novel approach utilizing Convolutional Neural Networks (CNNs), specifically a modified VGG16 architecture, for the detection of multiple and compound faults in industrial gearboxes. The proposed method leverages Short-Time Fourier Transform (STFT) spectrograms generated from vibration signals, allowing the model to self-learn fault features. Data augmentation through a striding technique enhances the training dataset, leading to improved fault classification accuracy. Experimental results demonstrate that the modified VGG16 model achieves a classification accuracy of 99.52%, significantly outperforming traditional methods. The VGG 16 Model is compared with other CNN models and it was found that VGG16 performs better in all scenarios. This study provides a robust, scalable solution for real-world gearbox fault diagnosis under varying operating conditions.