Based on the mechanism of thermally modulated reflected light, visible light images combined with machine learning methods can be used to estimate the surface temperature of metal equipment at ambient temperature under sunlight conditions. However, the surface conditions of on-site equipment and camera imaging parameters vary greatly across different scenarios, leading to poor generalization of models trained solely on laboratory image databases. To address this, it is necessary to update the original laboratory database by incorporating on-site images and retrain the model accordingly; on the other hand, since most of the on-site equipment is working normally, there are few images capturing fault-induced high temperatures. Even if the method of updating and retraining on-site images is used, the data imbalance in the image database can still cause significant measurement errors in these high-temperature images. This study studies image database update schemes to address both multi-scenario and data imbalance problems and demonstrates that retraining with as little as 5% scenario-specific images or 1% high-temperature images significantly improves temperature prediction accuracy, which was validated through on-site experiments at a substation. By comparing four machine learning algorithms (random forest regression, gradient boosted regression trees, decision trees, and k-nearest neighbors), this study reveals that RFR yields the best performance. These findings enhance the practical applicability of visible light image-based temperature measurement models in engineering contexts.
Li et al. (Fri,) studied this question.