Additive manufacturing (AM) and its various types are revolutionising the manufacturing industry and actively replacing conventional subtractive manufacturing. Among the various advantages of AM, reduction in weight, customisation, geometric independence, and a centralised supply chain are the most prominent. The widespread adoption of AM technologies is mainly hindered due to economic factors. The absence of in-situ monitoring and controlling of the AM printing process makes it dependent extensively on post-build quality inspection methods. Post-processing the 3D objects increases the time and cost of AM products. Sometimes, the cost of post-inspections is more than the actual printing cost. Moreover, the safety-critical application domains such as aerospace and medical sectors demand zero-defect 3D objects. Laser Powder Bed Fusion (LPBF) is a type of AM where complex geometrical 3D objects are built layer after layer by melting the pre-specified regions on a thin layer of the metal powder layer. LPBF is highly recommended for safety-critical parts production due to its high-quality and high-resolution capabilities. However, due to the complex nature of 3D printing, various defects can occur and compromise the desired properties of the final 3D builds. In order to reduce the extensive post-process quality inspections, the production cost, and the number of failed builds, in-situ monitoring of the printing process is essential. The metal printers are equipped with various sensors and produce vast amounts of data after each printing job. This sensor data can be used to monitor the printing process. Machine learning (ML) models are known for their outstanding pattern recognition, anomaly detection, and big data handling. Among the broad spectrum of defects of LPBF, micro-defects such as porosity, balling, surface deformation etc., are the most difficult to detect due to their sub-millimetre sizes. This work developed deep ML models to identify porosity and surface deformation defects from the powder bed images. Convolutional Neural Networks (CNN) are known for their automatic feature extraction and above-par accuracy. The study constructed a customised CNN from scratch to identify porosity defects on powder bed images with 97\% accuracy. Similarly, surface deformation defects were detected with 99\% accuracy from in-process images. The excellent performance of the proposed custom porosity CNN on two different defects emphasised the need to further explore the knowledge transfer capabilities of the model. The study further developed the custom porosity CNN model from the porosity dataset and used transfer learning to classify the surface deformation dataset by only retraining a part of the custom porosity CNN model. The model achieved an excellent accuracy of 93\%. The remarkable performance of the model on a relevant but unseen dataset validates the correctness of the methodology. Moreover, it shows the adaptability of the model in an unseen environment. The proposed models were assessed extensively against various evaluation metrics. In addition, Gradient-weighted Class Activation Mapping (Grad-CAM) and feature map visualisation were employed to debug and explain the decision-making process of the models. Finally, the experimental scope was extended from in-situ layer-wise defect detection to a multi-layer ex-situ X-ray Computed Tomography (XCT) dataset. Some defects, including under-study porosity defects, span multiple layers. Therefore, for a complete, conclusive analysis of porosity defects, pre-trained You Only Look Once (YOLO) model was used to classify and localise porosity defects on high-quality ex-situ XCT images. The experiments acquired a mAP@0.5 of 92.5%.