An efficient foreign objects detection network for power substation
Introduction
The power substation is an important hub connecting a power plant and users. The safety and stability of its operating environment are of great importance. Once the equipment in the substation is in trouble, the safety of the power system and the stability of the power supply will be significantly affected [1]. However, substation accidents occur frequently due to foreign objects intrusion [2]. The intrusion of foreign objects in the substation will cause an irreversible impact on power equipment, even cause paralysis of the whole substation. Therefore, it is of considerable significance to accurately find foreign objects in time, and then take corresponding measures to remove them.
At present, the inspection of foreign objects in power substations is usually done by manual inspection. This method mainly depends on the subjective sensory qualitative judgment and analysis of inspectors, and it is very much affected by the working experiences of the inspectors. Some abnormal conditions will inevitably be ignored due to the weakness of subjectivity and fatigue of human eyes. At the same time, a substation is a high-risk place, so it is challenging to inspect in poor weather conditions.
Moving target detection is to separate the foreground and background containing the moving target from the stationary or slowly moving background environment, extract and locate the moving target, and prepare for subsequent target tracking, behavior understanding and recognition, etc. The result of moving target detection directly affects the accuracy of the subsequent tasks. ViBe(Visual Background Extractor) [3] is a pixel-level foreground detection algorithm with high real-time performance, low memory occupancy, and high foreground detection accuracy. But when the background is complex, there may be serious “ghosting” and “flashing points” problems in the processing results using Vibe.
Deep learning, esp. convolutional neural networks had achieved great progress in computer vision [4], [5] and began to widely used in the patrol inspection system [6]. Researchers are using R-CNN (Region-CNN) [7] and its variations for intrusion detection, such as power transmission lines, airfield pavement, and usually modified networks are used to get better performance. For example, Shi et al. [8] improved the accuracy of the network by expanding data with GAN (Generative Adversarial Networks) [9]. Liu et al. [10] proposed a new network based on Fast R-CNN [11], which significantly improved detection accuracy and speed. However, the current R-CNN series of target detection algorithms are all based on RPN (Region Proposal Network) [12] to generate candidate target boxes. Compared with the original R-CNN that generates target candidate boxes by sliding window, the selective search method filters out a large number of useless anchors. However, after NMS (Non-Maximum Suppression) [13] processing, RPN brings a series of complex calculations and lots of redundant candidate boxes, which is time-consuming. Therefore, we propose the FODN4PS method to reduce time consumption, and redundant boxes are eliminated, which helps to improve detection accuracy and performance for foreign objects detection.
We conducted a detailed evaluation of the proposed network, including the analysis of different data enhancement methods, the comparison with common classification networks, Faster R-CNN and Mask R-CNN. The contributions for this paper are as follows.
- •
We propose FODN4PS, which effectively avoid the problems of large amount of computation and redundant candidate boxes caused by RPN. FODN4PS changes the way of generating candidate target boxes and ensure that each target has only one candidate box for subsequent operations, which further reduce time consumption.
- •
We design a simple but efficient target detection approach composing of extraction of candidate target box and classification implemented by FODN4PS. Compared with the complex two-stage target detection algorithm, the candidate target box after the process of FODN4PS is more accurate, and after the training of classification network in advance, FODN4PS obtains higher accuracy.
- •
We use image rotating randomly and GAN to expand the number of data sets for improving training accuracy. Through comparative evaluations, we found that the accuracy is improved using these two methods at the same time.
- •
We make comprehensive comparative evaluations to prove that FODN4PS is more efficient than Faster RCNN and Mask RCNN.
This paper is organized as follows. Section 2 reviews related work. Section 3 introduces FODN4PS structure design, including MORE Net, classification network, and loss function. Section 4 evaluates FODN4PS. Conclusion and future work are given in section 5.
Section snippets
Related work
In our work, we use a set of methods to improve the performance and accuracy of foreign object detection, including firstly, data enhancement to improve the possible number of data samples to cater for different scenes in a power substation; and also, our work is related the classical computer vision problem of object detection and background modeling, therefore we will review the related work on data enhancement, object detection and background modeling.
Design of the FODN4PS approach
Currently the two-stage target detection algorithms are all based on RPN, which generates a lot of redundant anchors in the whole image. These algorithms obtain hundreds of regions of interest after the NMS processing, which causes a series of computation and increases overhead of recognition. Compared with the complex two-stage target detection algorithm, we optimize this process by extracting accurate target candidate boxes which do not involve a series of complex calculations. Therefore,
Hardware setting
The hardware we use for testing includes: the graphic card is NVIDIA GeForce GTX TITAN X with 12G memory, Intel i7 CPU. The software packages include python is 3.6 with CUDA version 10.1.
Data enhancement
The size of training set affects the performance of the trained model. However the data of foreign objects in the substation is hard to collect. To solve the problem of insufficient training data, we adopts GAN [55] method to generate new images and rotates images randomly to enhance the data. Firstly, We
Conclusions and future work
Foreign objects intrusion detection is essential for the safety of power grid operation. However. The existing work usually changes the R-CNN on the basis of retaining RPN components, which generates a lot of redundant anchors that incurs bad performance as computation are wasted on these anchors. We optimize this process by extracting accurate target candidate boxes that do not involve a series of complex calculations. In order to effectively detect foreign objects, this paper proposes a new
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgment
The research is supported by the National Natural Science Foundation of China (62072469), National Key R&D Program (2018yfe0116700), Shandong Natural Science Foundation (ZR2019MF049, parallel data-driven fault prediction under online and offline combined cloud computing environment), basic research fund of Central University (2015020031), West Coast artificial intelligence technology innovation center (2019-1-5, 2019-1-6), and the Opening Project of Shanghai Trusted Industrial Control Platform (
Liang Xu has obtained master degree at China University of Petroleum, majoring in software engineering. Now he is a Ph.D candidate in Beijing University of Science and Technology. His main research interests include deep learning, software engineering and big data processing
References (55)
- et al.
Stimulus-driven and concept-driven analysis for image caption generation
Neurocomputing
(2020) - et al.
Rcnn-based foreign object detection for securing power transmission lines
Procedia Comp. Sci.
(2019) - et al.
A coarse-to-fine algorithm for fast median filtering of image data with a huge number of levels
Signal Process.
(1994) - et al.
Expert diagnose system and real-time monitoring of an electric power installation
Electroteh. Electron. Automat.
(2014) - et al.
Risk of fire and explosion in electrical substations due to the formation of flammable mixtures
Sci. Rep.
(2020) - et al.
Vibe: a powerful random technique to estimate the background in video sequences
- et al.
Automated colorization of a grayscale image with seed points propagation
IEEE Trans. Multimed.
(2020) - et al.
R-cnn for small object detection
- et al.
Rich feature hierarchies for accurate object detection and semantic segmentation
- et al.
Data expansion for foreign object detection in power grid
Generative adversarial nets
Fast r-cnn
Proc. IEEE Int. Conf. Comp. Vision
Hypernet: Towards accurate region proposal generation and joint object detection
Efficient non-maximum suppression
Knowledge-aided convolutional neural network for small organ segmentation
IEEE J. Biomed. Health Inform.
Synthetic Data for Text Localisation in Natural Images
Deep Learning for Generic Object Detection: A Survey
Yolov4: Optimal Speed and Accuracy of Object Detection, arXiv Preprint
Similarity analysis of actual fake fingerprints and generated fake fingerprints by dcgan
Int. J. Fuzzy Logic Intellig. Syst.
Dcgan-based data augmentation for tomato leaf disease identification
IEEE Access
Gesture recognition based on cnn and dcgan for calculation and text output
IEEE Access
DCGAN-based data augmentation for tomato leaf disease identification
IEEE Access
Object recognition at night scene based on dcgan and faster r-cnn
IEEE Access
Image analysis using mathematical morphology, IEEE transactions on pattern analysis and machine intelligence
Pca-sift: A more distinctive representation for local image descriptors
Distinctive image features from scale-invariant keypoints
Int. J. Comput. Vis.
Machine learning for high-speed corner detection
Cited by (22)
IntelliSense technology in the new power systems
2023, Renewable and Sustainable Energy ReviewsEditorial to special issue on cross-media learning for visual question answering
2022, Image and Vision ComputingTSCDNet+: A highly robust substation anomaly detection method
2021, OptikCitation Excerpt :However, this method is only limited to related equipment, and can not detect non-electrical defects. Literature [3] proposes a foreign body detection network in substation, which can accurately obtain the position and category of the foreign body through simple and efficient object detection. Literature [4] detected fire, smoke, and abnormal intrusion through monitoring video and multi-stream CNN network structure.
Liang Xu has obtained master degree at China University of Petroleum, majoring in software engineering. Now he is a Ph.D candidate in Beijing University of Science and Technology. His main research interests include deep learning, software engineering and big data processing
Yongkang Song is master student at China University of Petroleum, his main research area is computer vision.
Weishan Zhang is an professor at China University of Petroleum. His main research areas are deep learning, computer vision, and big data processing.
Yunyun An is a researcher at Huangdao District Power Supply Company of Shandong Electric Power Company Qingdao, State Grid, Qingdao. Her main research interest is big data processing. She has obtained master degree at China University of Petroleum.
Ye Wang is a researcher at PLA 9144, his research interest is big data processing.
Huansheng Ning is a professor in the School of Computer and Communication Engineering, University of Science and Technology Beijing, China. His current research interests include Internet of Things, Cybermatics.