In recent years

several efforts have been made to develop power transmission line abnormal target detection models based on edge devices. Typically

updates to these models rely on participation of the cloud

which means that network resource shortages can lead to update failures

followed by unsatisfactory recognition and detection performance in practical use. To address this problem

this article proposes an edge visual incremental perception framework

based on deep semi-supervised learning

for monitoring power transmission lines. After generation of the initial model using a small amount of labeled data

models trained using this framework can update themselves based on unlabeled data. A teacher-student joint training strategy

a data augmentation strategy

and a model updating strategy are also designed and adopted to improve the performance of the models trained with this framework. The proposed framework is then examined with various transmission line datasets with 1%

2%

5%

and 10% labeled data. General performance enhancement is thus confirmed against traditional supervised learning strategies. With the 10% labeled data training set

the recognition accuracy of the model is improved to exceed 80%

meeting the practical needs of power system operation

and thus clearly validating the effectiveness of the framework.