Sheet metal manufacturing is a critical process in many industries, and detecting faults in sheet metal is essential to ensure the quality of the final product. Common visual faults in sheet metal include crazing, inclusion, patches, pitted surface, rolled in, and scratches that can affect the strength, durability, and overall quality of the material. Traditional defect diagnosis techniques were often time-consuming, expensive, required increased manpower, and limited to single defect detection. Hence, advanced methods for diagnosing faults are required that can provide immediate outcomes. The image acquisition step involves the identification of a suitable dataset with high-quality images with various types of surface defects. The experimental study was divided into two key phases: (i) the deep learning phase and (ii) the machine learning phase. In the deep learning phase, we innovatively utilize convolutional neural networks for feature extraction, employing the final fully connected layer of six pre-trained networks, namely AlexNet, Visual Geometry Group (VGG)16, VGG19, Residual Network (ResNet)50, DenseNet201, and GoogLeNet, to extract features from sheet metal images. During the machine learning phase, the J48 decision tree algorithm was used to perform feature selection from the extracted features. Post selection of features, three families of classifiers such as Bayes, lazy, and rree were applied to determine the best feature extractor-classifier pair. The combination of DenseNet201 features with the functional tree classifier achieved an overall classification accuracy of 99.72% compared to all other pre-trained network features and classifiers evaluated. This innovative integration of deep learning for feature extraction and machine learning for feature selection and classification sets a new benchmark for fault diagnosis in sheet metal manufacturing which shows a highly efficient and accurate solution compared to existing techniques.