Traditional power equipment defect-detection relies on manual verification, which places a high demand on the verifier’s experience, as well as a high workload and low efficiency, which can lead to false detection and missed detection. The Mask of the regions with CNN features (Mask RCNN) deep learning model is used to provide a defect-detection approach based on the Mask RCNN of Attention, Rotation, Genetic algorithm (ARG-Mask RCNN), which employs infrared imaging as the data source to assess the features of damaged insulators. For the backbone network of Mask RCNN, the structure of Residual Network 101 (ResNet101) is improved and the attention mechanism is added, which makes the model more alert to small targets and can quickly identify the location of small targets, improve the loss function, integrate the rotation mechanism into the loss function formula, and generate an anchor frame where a rotation angle is used to accurately locate the fault location. The initial hyperparameters of the network are improved, and the Genetic Algorithm Combined with Gradient Descent (GA-GD) algorithm is used to optimize the model hyperparameters, so that the model training results are as close to the global best as possible. The experimental results show that the average accuracy of the insulator fault-detection method proposed in this paper is as high as 98%, and the number of frames per second (FPS) is 5.75, which provides a guarantee of the safe, stable, and reliable operation of our country’s power system.

Insulators play an important role in supporting and securing live conductors. In the task of insulator fault detection, existing models often miss small targets due to the interference of complex backgrounds. Furthermore, due to the very large aspect ratio of the insulators themselves, most models contain extraneous background in their results. To solve the above problems, this paper makes full use of the characteristics of the insulator itself and optimizes YOLOv5 from four aspects: feature extraction, feature fusion, candidate frame generation, and loss function. We divide the image feature extraction process into two procedures. The first procedure is mainly responsible for extracting insulator images from complex backgrounds, and the second stage is based on the extracted insulator images for fault analysis. At the same time, considering the trouble that the objective is too small due to the shooting distance, we design an attention mechanism that is more suitable for practical needs to enhance the ability to capture small objects. Furthermore, we design a rotated candidate box to improve the quality of the predicted box and reduce irrelevant background. The FAIN_detection dataset is compared with various advanced detection methods. The proposed method is more effective than the previous model by verifying the self-made data. Notably, the model achieved an average accuracy of 98% at 29 frames per second (FPS), laying the foundation for automatic insulator fault detection.

Hyperparameters involved in neural networks (NNs) have a significant impact on the accuracy of model predictions. However, the values of the hyperparameters need to be manually preset, and finding the best hyperparameters has always puzzled researchers. In order to improve the accuracy and speed of target recognition by a neural network, an improved genetic algorithm is proposed to optimize the hyperparameters of the network by taking the loss function as the research object. Firstly, the role of all loss functions in object detection is analyzed, and a mathematical model is established according to the relationship between loss functions and hyperparameters. Secondly, an improved genetic algorithm is proposed, and the feasibility of the improved algorithm is verified by using complex fractal function and fractional calculus. Finally, the improved genetic algorithm is used to optimize the hyperparameters of the neural network, and the prediction accuracy of the model before and after the improvement is comprehensively analyzed. By comparing with state-of-the-art object detectors, our proposed method achieves the highest prediction accuracy in object detection. Based on an average accuracy rate of 95%, the detection speed is 20 frames per second, which shows the rationality and feasibility of the optimized model.

Transmission lines (TLs) are the backbone of the power system, enabling efficient and reliable transmission of power over long distances. Fault diagnosis is the key technology of maintenance. Currently, there are several computer vision-based fault detection methods for power equipment. However, valuable fault samples from complex environments are difficult to collect, and many device faults exist in the form of small targets. Inspired by sketch-based retrieval, we first propose a method for TL fault diagnosis (TLFD) based on hand-drawn sketches. First, to address the problem of incomplete feature extraction, a densely connected multiscale feature network is proposed. Second, to address the problem of insufficient use of prior knowledge, a domain-aware dual attention network is proposed. Finally, a similarity loss function is constructed to supervise the similarity between all images, which preserves the prior knowledge in the pretrained model. To study the instance-level relationship between hand-drawn sketches and real images, we created the “TLFD” data. Experiments demonstrate that we provide a new idea for power system fault diagnosis, and the model significantly outperforms many state-of-the-art models designed for fault diagnosis.