Multi Classes Defects Detection for Substation Maintenance Based on Ensemble Model

Training a deep learning model to detect the substation maintenance defects from monitoring images is a popular method in improving the intelligent supervision level of substation. However, these defects are usually accompanied with great sematic difference and wide range of training difficulty, which leads to a high difficulty in training a single model to defect multi classes of defects comprehensively. This paper proposes a multi class defects detection method based on ensemble model. First, according to the semantic information and training difficulty, the whole defects classes will be grouped by constructing evaluation indicators. Then, segmenting the original dataset according to the grouping results, and different group of defects will be detected by different sub-models. Finally, to improve the detection speed of these separated models, they will be integrated by reusing part of their front-end layers which can reduce redundant calculation of the independent detection proceeding of each model. The experiment results suggest that the proposed model can reduce the model parameters by 64.3% and improve the detection accuracy by 2.4% compared with YOLOX-s.