Latent knowledge reasoning incorporated for multi-fitting decoupling detection on electric transmission line

Detection by fitting is a basic task in the fault diagnosis of electric transmission lines. To effectively solve the problem of missed and false detection in multi-fitting detection caused by occlusion between fittings and complex settings, we propose a decoupling detection method based on the latent knowledge reasoning model (LKRM) inspired by the fixed combination structure of fittings. This method consists of a scenario knowledge reasoning (SCKR) module, spatial knowledge reasoning (SPKR) and decoupling detection module (DDM). To obtain knowledge about the combination and structure of fittings, the scenario perception method of SCKR is employed to determine the scenario information in each image, and the reasoning network is utilized for further effective fusion of the scenario information. To obtain relevant knowledge about the spatial location of the fittings, spatial-visual features and geometric features are extracted by the spatial perception network in SPKR, and spatial knowledge is inferred through a graph convolution network (GCN). Fitting detection will be trained through the non-coupling method used by the decoupled module, and the final detection decision will then be made. The experimental results show that the LKRM achieves better detection results than the other advanced object detection models. The accuracy rate is improved by 10.8% compared with the original baseline model. The qualitative experiment shows that the LKRM can effectively solve the problem of object occlusion in a complex setting. •A latent knowledge reasoning model for fitting detection was proposed.•Scenario and spatial latent knowledge of fitting’s structure were extracted.•The accuracy rate is improved by 10.8% compared to original baseline model.•Decoupling method was used to balance two sub-networks.•The problem of object occlusion in a complex setting can be solved effectively.