Direct energy deposition (DED) is an emerging technology for remanufacturing as it enables fusion and deposition of metallic materials into complex geometries with high quality. The melting pool plays a critical role in quality control during the DED process. Ensuring stable melting pool geometry, temperature, and consistency is essential for producing defect-free components. Thermal imaging combined with unsupervised machine learning (ML) offers significant potential for in-situ defect prediction and quality control in the DED process. Moreover, in-situ thermal imaging generates incremental datasets, allowing for the continuous improvement of ML model predictions without the need for additional labelling as the dataset grows. In this work, we investigate the impact of self-organizing map (SOM)-based incremental learning parameters on in-situ thermal monitoring of the DED process using infrared (IR) imaging. Parameters including map size, neighborhood radius, learning rate, number of components, and the decay rate for neighborhood radius and learning rate were evaluated under low and high settings. Their effects on adjustment time for processing new IR images and final model accuracy, measured by quantization error (QE), were analysed. The findings provide a valuable starting point for researchers aiming to optimize SOM-based incremental learning for real-time defect detection using IR imaging of the DED melt pool.