Visual inspection is one of the most common and reliable methods used by human experts to perform diagnostics in the industry. However, it requires costly, specific expertise that could benefit from being automated. Such specific tasks are precisely the type of narrow knowledge that machine learning algorithms are best at learning. Yet, data-driven diagnostics from images are lagging behind compared to that from sensor data. This is due to sparse image data, as human experts only need a few photographs to accurately diagnose machine degradation. This paper presents a methodology to incorporate expert knowledge into the development of a data-driven diagnostic model for hydrogenerators based on visual inspection of the presence of partial discharge degradation products. The proposed methodology is validated using real industrial images. It emphasizes on the integration of human knowledge to address multiple challenges such as data sparsity, knowledge conformity, and human interpretability.